diff --git a/cyclegan/.gitignore b/cyclegan/.gitignore new file mode 100644 index 0000000..4a26633 --- /dev/null +++ b/cyclegan/.gitignore @@ -0,0 +1,40 @@ +datasets/ +checkpoints/ +results/ +build/ +dist/ +*.png +torch.egg-info/ +*/**/__pycache__ +torch/version.py +torch/csrc/generic/TensorMethods.cpp +torch/lib/*.so* +torch/lib/*.dylib* +torch/lib/*.h +torch/lib/build +torch/lib/tmp_install +torch/lib/include +torch/lib/torch_shm_manager +torch/csrc/cudnn/cuDNN.cpp +torch/csrc/nn/THNN.cwrap +torch/csrc/nn/THNN.cpp +torch/csrc/nn/THCUNN.cwrap +torch/csrc/nn/THCUNN.cpp +torch/csrc/nn/THNN_generic.cwrap +torch/csrc/nn/THNN_generic.cpp +torch/csrc/nn/THNN_generic.h +docs/src/**/* +test/data/legacy_modules.t7 +test/data/gpu_tensors.pt +test/htmlcov +test/.coverage +*/*.pyc +*/**/*.pyc +*/**/**/*.pyc +*/**/**/**/*.pyc +*/**/**/**/**/*.pyc +*/*.so* +*/**/*.so* +*/**/*.dylib* +test/data/legacy_serialized.pt +*~ diff --git a/cyclegan/LICENSE b/cyclegan/LICENSE new file mode 100644 index 0000000..d75f0ee --- /dev/null +++ b/cyclegan/LICENSE @@ -0,0 +1,58 @@ +Copyright (c) 2017, Jun-Yan Zhu and Taesung Park +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +* Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +* Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + +--------------------------- LICENSE FOR pix2pix -------------------------------- +BSD License + +For pix2pix software +Copyright (c) 2016, Phillip Isola and Jun-Yan Zhu +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +* Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +* Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +----------------------------- LICENSE FOR DCGAN -------------------------------- +BSD License + +For dcgan.torch software + +Copyright (c) 2015, Facebook, Inc. All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + +Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + +Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + +Neither the name Facebook nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. diff --git a/cyclegan/README.md b/cyclegan/README.md new file mode 100644 index 0000000..9f8f579 --- /dev/null +++ b/cyclegan/README.md @@ -0,0 +1,220 @@ + + +


+ +# CycleGAN +### [PyTorch](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix) | [project page](https://junyanz.github.io/CycleGAN/) | [paper](https://arxiv.org/pdf/1703.10593.pdf) + +Torch implementation for learning an image-to-image translation (i.e. [pix2pix](https://github.com/phillipi/pix2pix)) **without** input-output pairs, for example: + + + + + +[Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks](https://junyanz.github.io/CycleGAN/) + [Jun-Yan Zhu](https://people.eecs.berkeley.edu/~junyanz/)\*, [Taesung Park](https://taesung.me/)\*, [Phillip Isola](http://web.mit.edu/phillipi/), [Alexei A. Efros](https://people.eecs.berkeley.edu/~efros/) + Berkeley AI Research Lab, UC Berkeley + In ICCV 2017. (* equal contributions) + +This package includes CycleGAN, [pix2pix](https://github.com/phillipi/pix2pix), as well as other methods like [BiGAN](https://arxiv.org/abs/1605.09782)/[ALI](https://ishmaelbelghazi.github.io/ALI/) and Apple's paper [S+U learning](https://arxiv.org/pdf/1612.07828.pdf). +The code was written by [Jun-Yan Zhu](https://github.com/junyanz) and [Taesung Park](https://github.com/taesung). +**Update**: Please check out [PyTorch](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix) implementation for CycleGAN and pix2pix. +The PyTorch version is under active development and can produce results comparable or better than this Torch version. + +## Other implementations: +

[Tensorflow] (by Harry Yang), +[Tensorflow] (by Archit Rathore), +[Tensorflow] (by Van Huy), +[Tensorflow] (by Xiaowei Hu), + [Tensorflow-simple] (by Zhenliang He), + [TensorLayer] (by luoxier), +[Chainer] (by Yanghua Jin), +[Minimal PyTorch] (by yunjey), +[Mxnet] (by Ldpe2G), +[lasagne/Keras] (by tjwei), +[Keras] (by Simon Karlsson)

+ + +## Applications +### Monet Paintings to Photos + + +### Collection Style Transfer + + +### Object Transfiguration + + +### Season Transfer + + +### Photo Enhancement: Narrow depth of field + + + + +## Prerequisites +- Linux or OSX +- NVIDIA GPU + CUDA CuDNN (CPU mode and CUDA without CuDNN may work with minimal modification, but untested) +- For MAC users, you need the Linux/GNU commands `gfind` and `gwc`, which can be installed with `brew install findutils coreutils`. + +## Getting Started +### Installation +- Install torch and dependencies from https://github.com/torch/distro +- Install torch packages `nngraph`, `class`, `display` +```bash +luarocks install nngraph +luarocks install class +luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec +``` +- Clone this repo: +```bash +git clone https://github.com/junyanz/CycleGAN +cd CycleGAN +``` + +### Apply a Pre-trained Model +- Download the test photos (taken by [Alexei Efros](https://www.flickr.com/photos/aaefros)): +``` +bash ./datasets/download_dataset.sh ae_photos +``` +- Download the pre-trained model `style_cezanne` (For CPU model, use `style_cezanne_cpu`): +``` +bash ./pretrained_models/download_model.sh style_cezanne +``` +- Now, let's generate Paul Cézanne style images: +``` +DATA_ROOT=./datasets/ae_photos name=style_cezanne_pretrained model=one_direction_test phase=test loadSize=256 fineSize=256 resize_or_crop="scale_width" th test.lua +``` +The test results will be saved to `./results/style_cezanne_pretrained/latest_test/index.html`. +Please refer to [Model Zoo](#model-zoo) for more pre-trained models. +`./examples/test_vangogh_style_on_ae_photos.sh` is an example script that downloads the pretrained Van Gogh style network and runs it on Efros's photos. + +### Train +- Download a dataset (e.g. zebra and horse images from ImageNet): +```bash +bash ./datasets/download_dataset.sh horse2zebra +``` +- Train a model: +```bash +DATA_ROOT=./datasets/horse2zebra name=horse2zebra_model th train.lua +``` +- (CPU only) The same training command without using a GPU or CUDNN. Setting the environment variables ```gpu=0 cudnn=0``` forces CPU only +```bash +DATA_ROOT=./datasets/horse2zebra name=horse2zebra_model gpu=0 cudnn=0 th train.lua +``` +- (Optionally) start the display server to view results as the model trains. (See [Display UI](#display-ui) for more details): +```bash +th -ldisplay.start 8000 0.0.0.0 +``` + +### Test +- Finally, test the model: +```bash +DATA_ROOT=./datasets/horse2zebra name=horse2zebra_model phase=test th test.lua +``` +The test results will be saved to an HTML file here: `./results/horse2zebra_model/latest_test/index.html`. + + +## Model Zoo +Download the pre-trained models with the following script. The model will be saved to `./checkpoints/model_name/latest_net_G.t7`. +```bash +bash ./pretrained_models/download_model.sh model_name +``` +- `orange2apple` (orange -> apple) and `apple2orange`: trained on ImageNet categories `apple` and `orange`. +- `horse2zebra` (horse -> zebra) and `zebra2horse` (zebra -> horse): trained on ImageNet categories `horse` and `zebra`. +- `style_monet` (landscape photo -> Monet painting style), `style_vangogh` (landscape photo -> Van Gogh painting style), `style_ukiyoe` (landscape photo -> Ukiyo-e painting style), `style_cezanne` (landscape photo -> Cezanne painting style): trained on paintings and Flickr landscape photos. +- `monet2photo` (Monet paintings -> real landscape): trained on paintings and Flickr landscape photographs. +- `cityscapes_photo2label` (street scene -> label) and `cityscapes_label2photo` (label -> street scene): trained on the Cityscapes dataset. +- `map2sat` (map -> aerial photo) and `sat2map` (aerial photo -> map): trained on Google maps. +- `iphone2dslr_flower` (iPhone photos of flowers -> DSLR photos of flowers): trained on Flickr photos. + +CPU models can be downloaded using: +```bash +bash pretrained_models/download_model.sh _cpu +``` +, where `` can be `horse2zebra`, `style_monet`, etc. You just need to append `_cpu` to the target model. + +## Training and Test Details +To train a model, +```bash +DATA_ROOT=/path/to/data/ name=expt_name th train.lua +``` +Models are saved to `./checkpoints/expt_name` (can be changed by passing `checkpoint_dir=your_dir` in train.lua). +See `opt_train` in `options.lua` for additional training options. + +To test the model, +```bash +DATA_ROOT=/path/to/data/ name=expt_name phase=test th test.lua +``` +This will run the model named `expt_name` in both directions on all images in `/path/to/data/testA` and `/path/to/data/testB`. +A webpage with result images will be saved to `./results/expt_name` (can be changed by passing `results_dir=your_dir` in test.lua). +See `opt_test` in `options.lua` for additional test options. Please use `model=one_direction_test` if you only would like to generate outputs of the trained network in only one direction, and specify `which_direction=AtoB` or `which_direction=BtoA` to set the direction. + +There are other options that can be used. For example, you can specify `resize_or_crop=crop` option to avoid resizing the image to squares. This is indeed how we trained GTA2Cityscapes model in the projet [webpage](https://junyanz.github.io/CycleGAN/) and [Cycada](https://arxiv.org/pdf/1711.03213.pdf) model. We prepared the images at 1024px resolution, and used `resize_or_crop=crop fineSize=360` to work with the cropped images of size 360x360. We also used `lambda_identity=1.0`. + +## Datasets +Download the datasets using the following script. Many of the datasets were collected by other researchers. Please cite their papers if you use the data. +```bash +bash ./datasets/download_dataset.sh dataset_name +``` +- `facades`: 400 images from the [CMP Facades dataset](http://cmp.felk.cvut.cz/~tylecr1/facade/). [[Citation](datasets/bibtex/facades.tex)] +- `cityscapes`: 2975 images from the [Cityscapes training set](https://www.cityscapes-dataset.com/). [[Citation](datasets/bibtex/cityscapes.tex)]. Note: Due to license issue, we do not host the dataset on our repo. Please download the dataset directly from the Cityscapes webpage. Please refer to `./datasets/prepare_cityscapes_dataset.py` for more detail. +- `maps`: 1096 training images scraped from Google Maps. +- `horse2zebra`: 939 horse images and 1177 zebra images downloaded from [ImageNet](http://www.image-net.org/) using the keywords `wild horse` and `zebra` +- `apple2orange`: 996 apple images and 1020 orange images downloaded from [ImageNet](http://www.image-net.org/) using the keywords `apple` and `navel orange`. +- `summer2winter_yosemite`: 1273 summer Yosemite images and 854 winter Yosemite images were downloaded using Flickr API. See more details in our paper. +- `monet2photo`, `vangogh2photo`, `ukiyoe2photo`, `cezanne2photo`: The art images were downloaded from [Wikiart](https://www.wikiart.org/). The real photos are downloaded from Flickr using the combination of the tags *landscape* and *landscapephotography*. The training set size of each class is Monet:1074, Cezanne:584, Van Gogh:401, Ukiyo-e:1433, Photographs:6853. +- `iphone2dslr_flower`: both classes of images were downloaded from Flickr. The training set size of each class is iPhone:1813, DSLR:3316. See more details in our paper. + + +## Display UI +Optionally, for displaying images during training and test, use the [display package](https://github.com/szym/display). + +- Install it with: `luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec` +- Then start the server with: `th -ldisplay.start` +- Open this URL in your browser: [http://localhost:8000](http://localhost:8000) + +By default, the server listens on localhost. Pass `0.0.0.0` to allow external connections on any interface: +```bash +th -ldisplay.start 8000 0.0.0.0 +``` +Then open `http://(hostname):(port)/` in your browser to load the remote desktop. + +## Setup Training and Test data +To train CycleGAN model on your own datasets, you need to create a data folder with two subdirectories `trainA` and `trainB` that contain images from domain A and B. You can test your model on your training set by setting ``phase='train'`` in `test.lua`. You can also create subdirectories `testA` and `testB` if you have test data. + +You should **not** expect our method to work on just any random combination of input and output datasets (e.g. `cats<->keyboards`). From our experiments, we find it works better if two datasets share similar visual content. For example, `landscape painting<->landscape photographs` works much better than `portrait painting <-> landscape photographs`. `zebras<->horses` achieves compelling results while `cats<->dogs` completely fails. See the following section for more discussion. + + +## Failure cases + + +Our model does not work well when the test image is rather different from the images on which the model is trained, as is the case in the figure to the left (we trained on horses and zebras without riders, but test here one a horse with a rider). See additional typical failure cases [here](https://junyanz.github.io/CycleGAN/images/failures.jpg). On translation tasks that involve color and texture changes, like many of those reported above, the method often succeeds. We have also explored tasks that require geometric changes, with little success. For example, on the task of `dog<->cat` transfiguration, the learned translation degenerates into making minimal changes to the input. We also observe a lingering gap between the results achievable with paired training data and those achieved by our unpaired method. In some cases, this gap may be very hard -- or even impossible,-- to close: for example, our method sometimes permutes the labels for tree and building in the output of the cityscapes photos->labels task. + + + +## Citation +If you use this code for your research, please cite our [paper](https://junyanz.github.io/CycleGAN/): + +``` +@inproceedings{CycleGAN2017, + title={Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networkss}, + author={Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A}, + booktitle={Computer Vision (ICCV), 2017 IEEE International Conference on}, + year={2017} +} + +``` + + +## Related Projects: +[pix2pix](https://github.com/phillipi/pix2pix): Image-to-image translation using conditional adversarial nets +[iGAN](https://github.com/junyanz/iGAN): Interactive Image Generation via Generative Adversarial Networks + +## Cat Paper Collection +If you love cats, and love reading cool graphics, vision, and ML papers, please check out the Cat Paper [Collection](https://github.com/junyanz/CatPapers). + + +## Acknowledgments +Code borrows from [pix2pix](https://github.com/phillipi/pix2pix) and [DCGAN](https://github.com/soumith/dcgan.torch). The data loader is modified from [DCGAN](https://github.com/soumith/dcgan.torch) and [Context-Encoder](https://github.com/pathak22/context-encoder). The generative network is adopted from [neural-style](https://github.com/jcjohnson/neural-style) with [Instance Normalization](https://github.com/DmitryUlyanov/texture_nets/blob/master/InstanceNormalization.lua). diff --git a/cyclegan/data/aligned_data_loader.lua b/cyclegan/data/aligned_data_loader.lua new file mode 100644 index 0000000..e2637e0 --- /dev/null +++ b/cyclegan/data/aligned_data_loader.lua @@ -0,0 +1,44 @@ +-------------------------------------------------------------------------------- +-- Subclass of BaseDataLoader that provides data from two datasets. +-- The samples from the datasets are aligned +-- The datasets are of the same size +-------------------------------------------------------------------------------- +require 'data.base_data_loader' + +local class = require 'class' +data_util = paths.dofile('data_util.lua') + +AlignedDataLoader = class('AlignedDataLoader', 'BaseDataLoader') + +function AlignedDataLoader:__init(conf) + BaseDataLoader.__init(self, conf) + conf = conf or {} +end + +function AlignedDataLoader:name() + return 'AlignedDataLoader' +end + +function AlignedDataLoader:Initialize(opt) + opt.align_data = 1 + self.idx_A = {1, opt.input_nc} + self.idx_B = {opt.input_nc+1, opt.input_nc+opt.output_nc} + local nc = 3--opt.input_nc + opt.output_nc + self.data = data_util.load_dataset('', opt, nc) +end + +-- actually fetches the data +-- |return|: a table of two tables, each corresponding to +-- the batch for dataset A and dataset B +function AlignedDataLoader:LoadBatchForAllDatasets() + local batch_data, path = self.data:getBatch() + local batchA = batch_data[{ {}, self.idx_A, {}, {} }] + local batchB = batch_data[{ {}, self.idx_B, {}, {} }] + + return batchA, batchB, path, path +end + +-- returns the size of each dataset +function AlignedDataLoader:size(dataset) + return self.data:size() +end diff --git a/cyclegan/data/base_data_loader.lua b/cyclegan/data/base_data_loader.lua new file mode 100644 index 0000000..65fa830 --- /dev/null +++ b/cyclegan/data/base_data_loader.lua @@ -0,0 +1,53 @@ +-------------------------------------------------------------------------------- +-- Base Class for Providing Data +-------------------------------------------------------------------------------- + +local class = require 'class' +require 'torch' + +BaseDataLoader = class('BaseDataLoader') + +function BaseDataLoader:__init(conf) + conf = conf or {} + self.data_tm = torch.Timer() +end + +function BaseDataLoader:name() + return 'BaseDataLoader' +end + +function BaseDataLoader:Initialize(opt) +end + +-- actually fetches the data +-- |return|: a table of two tables, each corresponding to +-- the batch for dataset A and dataset B +function BaseDataLoader:LoadBatchForAllDatasets() + return {},{},{},{} +end + +-- returns the next batch +-- a wrapper of getBatch(), which is meant to be overriden by subclasses +-- |return|: a table of two tables, each corresponding to +-- the batch for dataset A and dataset B +function BaseDataLoader:GetNextBatch() + self.data_tm:reset() + self.data_tm:resume() + local dataA, dataB, pathA, pathB = self:LoadBatchForAllDatasets() + self.data_tm:stop() + return dataA, dataB, pathA, pathB +end + +function BaseDataLoader:time_elapsed_to_fetch_data() + return self.data_tm:time().real +end + +-- returns the size of each dataset +function BaseDataLoader:size(dataset) + return 0 +end + + + + + diff --git a/cyclegan/data/data.lua b/cyclegan/data/data.lua new file mode 100644 index 0000000..d43586e --- /dev/null +++ b/cyclegan/data/data.lua @@ -0,0 +1,103 @@ +--[[ + This data loader is a modified version of the one from dcgan.torch + (see https://github.com/soumith/dcgan.torch/blob/master/data/data.lua). + + Copyright (c) 2016, Deepak Pathak [See LICENSE file for details] +]]-- + +local Threads = require 'threads' +Threads.serialization('threads.sharedserialize') + +local data = {} + +local result = {} +local unpack = unpack and unpack or table.unpack + +function data.new(n, opt_) + opt_ = opt_ or {} + local self = {} + for k,v in pairs(data) do + self[k] = v + end + + local donkey_file = 'donkey_folder.lua' +-- print('n..' .. n) + if n > 0 then + local options = opt_ + self.threads = Threads(n, + function() require 'torch' end, + function(idx) + opt = options + tid = idx + local seed = (opt.manualSeed and opt.manualSeed or 0) + idx + torch.manualSeed(seed) + torch.setnumthreads(1) + print(string.format('Starting donkey with id: %d seed: %d', tid, seed)) + assert(options, 'options not found') + assert(opt, 'opt not given') + print(opt) + paths.dofile(donkey_file) + end + + ) + else + if donkey_file then paths.dofile(donkey_file) end +-- print('empty threads') + self.threads = {} + function self.threads:addjob(f1, f2) f2(f1()) end + function self.threads:dojob() end + function self.threads:synchronize() end + end + + local nSamples = 0 + self.threads:addjob(function() return trainLoader:size() end, + function(c) nSamples = c end) + self.threads:synchronize() + self._size = nSamples + + for i = 1, n do + self.threads:addjob(self._getFromThreads, + self._pushResult) + end +-- print(self.threads) + return self +end + +function data._getFromThreads() + assert(opt.batchSize, 'opt.batchSize not found') + return trainLoader:sample(opt.batchSize) +end + +function data._pushResult(...) + local res = {...} + if res == nil then + self.threads:synchronize() + end + result[1] = res +end + + + +function data:getBatch() + -- queue another job + self.threads:addjob(self._getFromThreads, self._pushResult) + self.threads:dojob() + local res = result[1] + + img_data = res[1] + img_paths = res[3] + + result[1] = nil + if torch.type(img_data) == 'table' then + img_data = unpack(img_data) + end + + + return img_data, img_paths +end + +function data:size() + return self._size +end + +return data diff --git a/cyclegan/data/data_util.lua b/cyclegan/data/data_util.lua new file mode 100644 index 0000000..319560a --- /dev/null +++ b/cyclegan/data/data_util.lua @@ -0,0 +1,24 @@ +local data_util = {} + +require 'torch' +-- options = require '../options.lua' +-- load dataset from the file system +-- |name|: name of the dataset. It's currently either 'A' or 'B' +function data_util.load_dataset(name, opt, nc) + local tensortype = torch.getdefaulttensortype() + torch.setdefaulttensortype('torch.FloatTensor') + + local new_opt = options.clone(opt) + new_opt.manualSeed = torch.random(1, 10000) -- fix seed + new_opt.nc = nc + torch.manualSeed(new_opt.manualSeed) + local data_loader = paths.dofile('../data/data.lua') + new_opt.phase = new_opt.phase .. name + local data = data_loader.new(new_opt.nThreads, new_opt) + print("Dataset Size " .. name .. ": ", data:size()) + + torch.setdefaulttensortype(tensortype) + return data +end + +return data_util diff --git a/cyclegan/data/dataset.lua b/cyclegan/data/dataset.lua new file mode 100644 index 0000000..45328d3 --- /dev/null +++ b/cyclegan/data/dataset.lua @@ -0,0 +1,398 @@ +--[[ + Copyright (c) 2015-present, Facebook, Inc. + All rights reserved. + + This source code is licensed under the BSD-style license found in the + LICENSE file in the root directory of this source tree. An additional grant + of patent rights can be found in the PATENTS file in the same directory. +]]-- + +require 'torch' +torch.setdefaulttensortype('torch.FloatTensor') +local ffi = require 'ffi' +local class = require('pl.class') +local dir = require 'pl.dir' +local tablex = require 'pl.tablex' +local argcheck = require 'argcheck' +require 'sys' +require 'xlua' +require 'image' + +local dataset = torch.class('dataLoader') + +local initcheck = argcheck{ + pack=true, + help=[[ + A dataset class for images in a flat folder structure (folder-name is class-name). + Optimized for extremely large datasets (upwards of 14 million images). + Tested only on Linux (as it uses command-line linux utilities to scale up) +]], + {check=function(paths) + local out = true; + for k,v in ipairs(paths) do + if type(v) ~= 'string' then + print('paths can only be of string input'); + out = false + end + end + return out + end, + name="paths", + type="table", + help="Multiple paths of directories with images"}, + + {name="sampleSize", + type="table", + help="a consistent sample size to resize the images"}, + + {name="split", + type="number", + help="Percentage of split to go to Training" + }, + {name="serial_batches", + type="number", + help="if randomly sample training images"}, + + {name="samplingMode", + type="string", + help="Sampling mode: random | balanced ", + default = "balanced"}, + + {name="verbose", + type="boolean", + help="Verbose mode during initialization", + default = false}, + + {name="loadSize", + type="table", + help="a size to load the images to, initially", + opt = true}, + + {name="forceClasses", + type="table", + help="If you want this loader to map certain classes to certain indices, " + .. "pass a classes table that has {classname : classindex} pairs." + .. " For example: {3 : 'dog', 5 : 'cat'}" + .. "This function is very useful when you want two loaders to have the same " + .. "class indices (trainLoader/testLoader for example)", + opt = true}, + + {name="sampleHookTrain", + type="function", + help="applied to sample during training(ex: for lighting jitter). " + .. "It takes the image path as input", + opt = true}, + + {name="sampleHookTest", + type="function", + help="applied to sample during testing", + opt = true}, +} + +function dataset:__init(...) + + -- argcheck + local args = initcheck(...) + print(args) + for k,v in pairs(args) do self[k] = v end + + if not self.loadSize then self.loadSize = self.sampleSize; end + + if not self.sampleHookTrain then self.sampleHookTrain = self.defaultSampleHook end + if not self.sampleHookTest then self.sampleHookTest = self.defaultSampleHook end + self.image_count = 1 +-- print('image_count_init', self.image_count) + -- find class names + self.classes = {} + local classPaths = {} + if self.forceClasses then + for k,v in pairs(self.forceClasses) do + self.classes[k] = v + classPaths[k] = {} + end + end + local function tableFind(t, o) for k,v in pairs(t) do if v == o then return k end end end + -- loop over each paths folder, get list of unique class names, + -- also store the directory paths per class + -- for each class, + for k,path in ipairs(self.paths) do +-- print('path', path) + local dirs = {} -- hack + dirs[1] = path +-- local dirs = dir.getdirectories(path); + for k,dirpath in ipairs(dirs) do + local class = paths.basename(dirpath) + local idx = tableFind(self.classes, class) +-- print(class) +-- print(idx) + if not idx then + table.insert(self.classes, class) + idx = #self.classes + classPaths[idx] = {} + end + if not tableFind(classPaths[idx], dirpath) then + table.insert(classPaths[idx], dirpath); + end + end + end + + self.classIndices = {} + for k,v in ipairs(self.classes) do + self.classIndices[v] = k + end + + -- define command-line tools, try your best to maintain OSX compatibility + local wc = 'wc' + local cut = 'cut' + local find = 'find -H' -- if folder name is symlink, do find inside it after dereferencing + + if ffi.os == 'OSX' then + wc = 'gwc' + cut = 'gcut' + find = 'gfind' + end + ---------------------------------------------------------------------- + -- Options for the GNU find command + local extensionList = {'jpg', 'png','JPG','PNG','JPEG', 'ppm', 'PPM', 'bmp', 'BMP'} + local findOptions = ' -iname "*.' .. extensionList[1] .. '"' + for i=2,#extensionList do + findOptions = findOptions .. ' -o -iname "*.' .. extensionList[i] .. '"' + end + + -- find the image path names + self.imagePath = torch.CharTensor() -- path to each image in dataset + self.imageClass = torch.LongTensor() -- class index of each image (class index in self.classes) + self.classList = {} -- index of imageList to each image of a particular class + self.classListSample = self.classList -- the main list used when sampling data + + print('running "find" on each class directory, and concatenate all' + .. ' those filenames into a single file containing all image paths for a given class') + -- so, generates one file per class + local classFindFiles = {} + for i=1,#self.classes do + classFindFiles[i] = os.tmpname() + end + local combinedFindList = os.tmpname(); + + local tmpfile = os.tmpname() + local tmphandle = assert(io.open(tmpfile, 'w')) + -- iterate over classes + for i, class in ipairs(self.classes) do + -- iterate over classPaths + for j,path in ipairs(classPaths[i]) do + local command = find .. ' "' .. path .. '" ' .. findOptions + .. ' >>"' .. classFindFiles[i] .. '" \n' + tmphandle:write(command) + end + end + io.close(tmphandle) + os.execute('bash ' .. tmpfile) + os.execute('rm -f ' .. tmpfile) + + print('now combine all the files to a single large file') + local tmpfile = os.tmpname() + local tmphandle = assert(io.open(tmpfile, 'w')) + -- concat all finds to a single large file in the order of self.classes + for i=1,#self.classes do + local command = 'cat "' .. classFindFiles[i] .. '" >>' .. combinedFindList .. ' \n' + tmphandle:write(command) + end + io.close(tmphandle) + os.execute('bash ' .. tmpfile) + os.execute('rm -f ' .. tmpfile) + + --========================================================================== + print('load the large concatenated list of sample paths to self.imagePath') + local cmd = wc .. " -L '" + .. combinedFindList .. "' |" + .. cut .. " -f1 -d' '" + print('cmd..' .. cmd) + local maxPathLength = tonumber(sys.fexecute(wc .. " -L '" + .. combinedFindList .. "' |" + .. cut .. " -f1 -d' '")) + 1 + local length = tonumber(sys.fexecute(wc .. " -l '" + .. combinedFindList .. "' |" + .. cut .. " -f1 -d' '")) + assert(length > 0, "Could not find any image file in the given input paths") + assert(maxPathLength > 0, "paths of files are length 0?") + self.imagePath:resize(length, maxPathLength):fill(0) + local s_data = self.imagePath:data() + local count = 0 + for line in io.lines(combinedFindList) do + ffi.copy(s_data, line) + s_data = s_data + maxPathLength + if self.verbose and count % 10000 == 0 then + xlua.progress(count, length) + end; + count = count + 1 + end + + self.numSamples = self.imagePath:size(1) + if self.verbose then print(self.numSamples .. ' samples found.') end + --========================================================================== + print('Updating classList and imageClass appropriately') + self.imageClass:resize(self.numSamples) + local runningIndex = 0 + for i=1,#self.classes do + if self.verbose then xlua.progress(i, #(self.classes)) end + local length = tonumber(sys.fexecute(wc .. " -l '" + .. classFindFiles[i] .. "' |" + .. cut .. " -f1 -d' '")) + if length == 0 then + error('Class has zero samples') + else + self.classList[i] = torch.linspace(runningIndex + 1, runningIndex + length, length):long() + self.imageClass[{{runningIndex + 1, runningIndex + length}}]:fill(i) + end + runningIndex = runningIndex + length + end + + --========================================================================== + -- clean up temporary files + print('Cleaning up temporary files') + local tmpfilelistall = '' + for i=1,#(classFindFiles) do + tmpfilelistall = tmpfilelistall .. ' "' .. classFindFiles[i] .. '"' + if i % 1000 == 0 then + os.execute('rm -f ' .. tmpfilelistall) + tmpfilelistall = '' + end + end + os.execute('rm -f ' .. tmpfilelistall) + os.execute('rm -f "' .. combinedFindList .. '"') + --========================================================================== + + if self.split == 100 then + self.testIndicesSize = 0 + else + print('Splitting training and test sets to a ratio of ' + .. self.split .. '/' .. (100-self.split)) + self.classListTrain = {} + self.classListTest = {} + self.classListSample = self.classListTrain + local totalTestSamples = 0 + -- split the classList into classListTrain and classListTest + for i=1,#self.classes do + local list = self.classList[i] + local count = self.classList[i]:size(1) + local splitidx = math.floor((count * self.split / 100) + 0.5) -- +round + local perm = torch.randperm(count) + self.classListTrain[i] = torch.LongTensor(splitidx) + for j=1,splitidx do + self.classListTrain[i][j] = list[perm[j]] + end + if splitidx == count then -- all samples were allocated to train set + self.classListTest[i] = torch.LongTensor() + else + self.classListTest[i] = torch.LongTensor(count-splitidx) + totalTestSamples = totalTestSamples + self.classListTest[i]:size(1) + local idx = 1 + for j=splitidx+1,count do + self.classListTest[i][idx] = list[perm[j]] + idx = idx + 1 + end + end + end + -- Now combine classListTest into a single tensor + self.testIndices = torch.LongTensor(totalTestSamples) + self.testIndicesSize = totalTestSamples + local tdata = self.testIndices:data() + local tidx = 0 + for i=1,#self.classes do + local list = self.classListTest[i] + if list:dim() ~= 0 then + local ldata = list:data() + for j=0,list:size(1)-1 do + tdata[tidx] = ldata[j] + tidx = tidx + 1 + end + end + end + end +end + +-- size(), size(class) +function dataset:size(class, list) + list = list or self.classList + if not class then + return self.numSamples + elseif type(class) == 'string' then + return list[self.classIndices[class]]:size(1) + elseif type(class) == 'number' then + return list[class]:size(1) + end +end + +-- getByClass +function dataset:getByClass(class) + local index = 0 + if self.serial_batches == 1 then + index = math.fmod(self.image_count-1, self.classListSample[class]:nElement())+1 + self.image_count = self.image_count +1 + else + index = math.ceil(torch.uniform() * self.classListSample[class]:nElement()) + end + + local imgpath = ffi.string(torch.data(self.imagePath[self.classListSample[class][index]])) + return self:sampleHookTrain(imgpath), imgpath +end + +-- converts a table of samples (and corresponding labels) to a clean tensor +local function tableToOutput(self, dataTable, scalarTable) + local data, scalarLabels, labels + if opt.resize_or_crop == 'crop' or opt.resize_or_crop == 'scale_width' or opt.resize_or_crop == 'scale_height' then + assert(#scalarTable == 1) + data = torch.Tensor(1, + dataTable[1]:size(1), dataTable[1]:size(2), dataTable[1]:size(3)) + data[1]:copy(dataTable[1]) + scalarLabels = torch.LongTensor(#scalarTable):fill(-1111) + else + local quantity = #scalarTable + data = torch.Tensor(quantity, + self.sampleSize[1], self.sampleSize[2], self.sampleSize[3]) + scalarLabels = torch.LongTensor(quantity):fill(-1111) + for i=1,#dataTable do + data[i]:copy(dataTable[i]) + scalarLabels[i] = scalarTable[i] + end + end + return data, scalarLabels +end + +-- sampler, samples from the training set. +function dataset:sample(quantity) + assert(quantity) + local dataTable = {} + local scalarTable = {} + local samplePaths = {} + for i=1,quantity do + local class = torch.random(1, #self.classes) + local out, imgpath = self:getByClass(class) + table.insert(dataTable, out) + table.insert(scalarTable, class) + samplePaths[i] = imgpath + end + + local data, scalarLabels = tableToOutput(self, dataTable, scalarTable) + return data, scalarLabels, samplePaths-- filePaths +end + +function dataset:get(i1, i2) + local indices = torch.range(i1, i2); + local quantity = i2 - i1 + 1; + assert(quantity > 0) + -- now that indices has been initialized, get the samples + local dataTable = {} + local scalarTable = {} + for i=1,quantity do + -- load the sample + local imgpath = ffi.string(torch.data(self.imagePath[indices[i]])) + local out = self:sampleHookTest(imgpath) + table.insert(dataTable, out) + table.insert(scalarTable, self.imageClass[indices[i]]) + end + local data, scalarLabels = tableToOutput(self, dataTable, scalarTable) + return data, scalarLabels +end + +return dataset diff --git a/cyclegan/data/donkey_folder.lua b/cyclegan/data/donkey_folder.lua new file mode 100644 index 0000000..803b9ea --- /dev/null +++ b/cyclegan/data/donkey_folder.lua @@ -0,0 +1,200 @@ + +--[[ + This data loader is a modified version of the one from dcgan.torch + (see https://github.com/soumith/dcgan.torch/blob/master/data/donkey_folder.lua). + Copyright (c) 2016, Deepak Pathak [See LICENSE file for details] + Copyright (c) 2015-present, Facebook, Inc. + All rights reserved. + This source code is licensed under the BSD-style license found in the + LICENSE file in the root directory of this source tree. An additional grant + of patent rights can be found in the PATENTS file in the same directory. +]]-- + +require 'image' +paths.dofile('dataset.lua') +-- This file contains the data-loading logic and details. +-- It is run by each data-loader thread. +------------------------------------------ +-------- COMMON CACHES and PATHS +-- Check for existence of opt.data +if opt.DATA_ROOT then + opt.data = paths.concat(opt.DATA_ROOT, opt.phase) +else + print(os.getenv('DATA_ROOT')) + opt.data = paths.concat(os.getenv('DATA_ROOT'), opt.phase) +end + +if not paths.dirp(opt.data) then + error('Did not find directory: ' .. opt.data) +end + +-- a cache file of the training metadata (if doesnt exist, will be created) +local cache_prefix = opt.data:gsub('/', '_') +os.execute(('mkdir -p %s'):format(opt.cache_dir)) +local trainCache = paths.concat(opt.cache_dir, cache_prefix .. '_trainCache.t7') + +-------------------------------------------------------------------------------------------- +local input_nc = opt.nc -- input channels +local loadSize = {input_nc, opt.loadSize} +local sampleSize = {input_nc, opt.fineSize} + +local function loadImage(path) + local input = image.load(path, 3, 'float') + local h = input:size(2) + local w = input:size(3) + + local imA = image.crop(input, 0, 0, w/2, h) + imA = image.scale(imA, loadSize[2], loadSize[2]) + local imB = image.crop(input, w/2, 0, w, h) + imB = image.scale(imB, loadSize[2], loadSize[2]) + + local perm = torch.LongTensor{3, 2, 1} + imA = imA:index(1, perm) + imA = imA:mul(2):add(-1) + imB = imB:index(1, perm) + imB = imB:mul(2):add(-1) + + assert(imA:max()<=1,"A: badly scaled inputs") + assert(imA:min()>=-1,"A: badly scaled inputs") + assert(imB:max()<=1,"B: badly scaled inputs") + assert(imB:min()>=-1,"B: badly scaled inputs") + + + local oW = sampleSize[2] + local oH = sampleSize[2] + local iH = imA:size(2) + local iW = imA:size(3) + + if iH~=oH then + h1 = math.ceil(torch.uniform(1e-2, iH-oH)) + end + + if iW~=oW then + w1 = math.ceil(torch.uniform(1e-2, iW-oW)) + end + if iH ~= oH or iW ~= oW then + imA = image.crop(imA, w1, h1, w1 + oW, h1 + oH) + imB = image.crop(imB, w1, h1, w1 + oW, h1 + oH) + end + + if opt.flip == 1 and torch.uniform() > 0.5 then + imA = image.hflip(imA) + imB = image.hflip(imB) + end + + local concatenated = torch.cat(imA,imB,1) + + return concatenated +end + + +local function loadSingleImage(path) + local im = image.load(path, input_nc, 'float') + if opt.resize_or_crop == 'resize_and_crop' then + im = image.scale(im, loadSize[2], loadSize[2]) + end + if input_nc == 3 then + local perm = torch.LongTensor{3, 2, 1} + im = im:index(1, perm)--:mul(256.0): brg, rgb + im = im:mul(2):add(-1) + end + assert(im:max()<=1,"A: badly scaled inputs") + assert(im:min()>=-1,"A: badly scaled inputs") + + local oW = sampleSize[2] + local oH = sampleSize[2] + local iH = im:size(2) + local iW = im:size(3) + if (opt.resize_or_crop == 'resize_and_crop' ) then + local h1, w1 = 0, 0 + if iH~=oH then + h1 = math.ceil(torch.uniform(1e-2, iH-oH)) + end + if iW~=oW then + w1 = math.ceil(torch.uniform(1e-2, iW-oW)) + end + if iH ~= oH or iW ~= oW then + im = image.crop(im, w1, h1, w1 + oW, h1 + oH) + end + elseif (opt.resize_or_crop == 'combined') then + local sH = math.min(math.ceil(oH * torch.uniform(1+1e-2, 2.0-1e-2)), iH-1e-2) + local sW = math.min(math.ceil(oW * torch.uniform(1+1e-2, 2.0-1e-2)), iW-1e-2) + local h1 = math.ceil(torch.uniform(1e-2, iH-sH)) + local w1 = math.ceil(torch.uniform(1e-2, iW-sW)) + im = image.crop(im, w1, h1, w1 + sW, h1 + sH) + im = image.scale(im, oW, oH) + elseif (opt.resize_or_crop == 'crop') then + local w = math.min(math.min(oH, iH),iW) + w = math.floor(w/4)*4 + local x = math.floor(torch.uniform(0, iW - w)) + local y = math.floor(torch.uniform(0, iH - w)) + im = image.crop(im, x, y, x+w, y+w) + elseif (opt.resize_or_crop == 'scale_width') then + w = oW + h = torch.floor(iH * oW/iW) + im = image.scale(im, w, h) + elseif (opt.resize_or_crop == 'scale_height') then + h = oH + w = torch.floor(iW * oH / iH) + im = image.scale(im, w, h) + end + + if opt.flip == 1 and torch.uniform() > 0.5 then + im = image.hflip(im) + end + + return im + +end + +-- channel-wise mean and std. Calculate or load them from disk later in the script. +local mean,std +-------------------------------------------------------------------------------- +-- Hooks that are used for each image that is loaded + +-- function to load the image, jitter it appropriately (random crops etc.) +local trainHook_singleimage = function(self, path) + collectgarbage() + -- print('load single image') + local im = loadSingleImage(path) + return im +end + +-- function that loads images that have juxtaposition +-- of two images from two domains +local trainHook_doubleimage = function(self, path) + -- print('load double image') + collectgarbage() + + local im = loadImage(path) + return im +end + + +if opt.align_data > 0 then + sample_nc = input_nc*2 + trainHook = trainHook_doubleimage +else + sample_nc = input_nc + trainHook = trainHook_singleimage +end + +trainLoader = dataLoader{ + paths = {opt.data}, + loadSize = {input_nc, loadSize[2], loadSize[2]}, + sampleSize = {sample_nc, sampleSize[2], sampleSize[2]}, + split = 100, + serial_batches = opt.serial_batches, + verbose = true + } + +trainLoader.sampleHookTrain = trainHook +collectgarbage() + +-- do some sanity checks on trainLoader +do + local class = trainLoader.imageClass + local nClasses = #trainLoader.classes + assert(class:max() <= nClasses, "class logic has error") + assert(class:min() >= 1, "class logic has error") +end diff --git a/cyclegan/data/unaligned_data_loader.lua b/cyclegan/data/unaligned_data_loader.lua new file mode 100644 index 0000000..526ba73 --- /dev/null +++ b/cyclegan/data/unaligned_data_loader.lua @@ -0,0 +1,49 @@ +-------------------------------------------------------------------------------- +-- Subclass of BaseDataLoader that provides data from two datasets. +-- The samples from the datasets are not aligned. +-- The datasets can have different sizes +-------------------------------------------------------------------------------- +require 'data.base_data_loader' + +local class = require 'class' +data_util = paths.dofile('data_util.lua') + +UnalignedDataLoader = class('UnalignedDataLoader', 'BaseDataLoader') + +function UnalignedDataLoader:__init(conf) + BaseDataLoader.__init(self, conf) + conf = conf or {} +end + +function UnalignedDataLoader:name() + return 'UnalignedDataLoader' +end + +function UnalignedDataLoader:Initialize(opt) + opt.align_data = 0 + self.dataA = data_util.load_dataset('A', opt, opt.input_nc) + self.dataB = data_util.load_dataset('B', opt, opt.output_nc) +end + +-- actually fetches the data +-- |return|: a table of two tables, each corresponding to +-- the batch for dataset A and dataset B +function UnalignedDataLoader:LoadBatchForAllDatasets() + local batchA, pathA = self.dataA:getBatch() + local batchB, pathB = self.dataB:getBatch() + return batchA, batchB, pathA, pathB +end + +-- returns the size of each dataset +function UnalignedDataLoader:size(dataset) + if dataset == 'A' then + return self.dataA:size() + end + + if dataset == 'B' then + return self.dataB:size() + end + + return math.max(self.dataA:size(), self.dataB:size()) + -- return the size of the largest dataset by default +end diff --git a/cyclegan/examples/test_vangogh_style_on_ae_photos.sh b/cyclegan/examples/test_vangogh_style_on_ae_photos.sh new file mode 100644 index 0000000..60ce1bb --- /dev/null +++ b/cyclegan/examples/test_vangogh_style_on_ae_photos.sh @@ -0,0 +1,19 @@ +#!/bin/sh + +## This script download the dataset and pre-trained network, +## and generates style transferred images. + +# Download the dataset. The downloaded dataset is stored in ./datasets/${DATASET_NAME} +DATASET_NAME='ae_photos' +bash ./datasets/download_dataset.sh $DATASET_NAME + +# Download the pre-trained model. The downloaded model is stored in ./models/${MODEL_NAME}_pretrained/latest_net_G.t7 +MODEL_NAME='style_vangogh' +bash ./pretrained_models/download_model.sh $MODEL_NAME + +# Run style transfer using the downloaded dataset and model +DATA_ROOT=./datasets/$DATASET_NAME name=${MODEL_NAME}_pretrained model=one_direction_test phase=test how_many='all' loadSize=256 fineSize=256 resize_or_crop='scale_width' th test.lua + +if [ $? == 0 ]; then + echo "The result can be viewed at ./results/${MODEL_NAME}_pretrained/latest_test/index.html" +fi diff --git a/cyclegan/examples/train_maps.sh b/cyclegan/examples/train_maps.sh new file mode 100644 index 0000000..7c266ac --- /dev/null +++ b/cyclegan/examples/train_maps.sh @@ -0,0 +1,26 @@ +DB_NAME='maps' +GPU_ID=1 +DISPLAY_ID=1 +NET_G=resnet_6blocks +NET_D=basic +MODEL=cycle_gan +SAVE_EPOCH=5 +ALIGN_DATA=0 +LAMBDA=10 +NF=64 + + +EXPR_NAME=${DB_NAME}_${MODEL}_${LAMBDA} + +CHECKPOINT_DIR=./checkpoints/ +LOG_FILE=${CHECKPOINT_DIR}${EXPR_NAME}/log.txt +mkdir -p ${CHECKPOINT_DIR}${EXPR_NAME} + +DATA_ROOT=./datasets/$DB_NAME align_data=$ALIGN_DATA use_lsgan=1 \ +which_direction='AtoB' display_plot=$PLOT pool_size=50 niter=100 niter_decay=100 \ +which_model_netG=$NET_G which_model_netD=$NET_D model=$MODEL lr=0.0002 print_freq=200 lambda_A=$LAMBDA lambda_B=$LAMBDA \ +loadSize=143 fineSize=128 gpu=$GPU_ID display_winsize=128 \ +name=$EXPR_NAME flip=1 save_epoch_freq=$SAVE_EPOCH \ +continue_train=0 display_id=$DISPLAY_ID \ +checkpoints_dir=$CHECKPOINT_DIR\ + th train.lua | tee -a $LOG_FILE diff --git a/cyclegan/models/architectures.lua b/cyclegan/models/architectures.lua new file mode 100644 index 0000000..8b83cea --- /dev/null +++ b/cyclegan/models/architectures.lua @@ -0,0 +1,384 @@ +require 'nngraph' + + +---------------------------------------------------------------------------- +local function weights_init(m) + local name = torch.type(m) + if name:find('Convolution') then + m.weight:normal(0.0, 0.02) + m.bias:fill(0) + elseif name:find('Normalization') then + if m.weight then m.weight:normal(1.0, 0.02) end + if m.bias then m.bias:fill(0) end + end +end + + +normalization = nil + +function set_normalization(norm) +if norm == 'instance' then + require 'util.InstanceNormalization' + print('use InstanceNormalization') + normalization = nn.InstanceNormalization +elseif norm == 'batch' then + print('use SpatialBatchNormalization') + normalization = nn.SpatialBatchNormalization +end +end + +function defineG(input_nc, output_nc, ngf, which_model_netG, nz, arch) + local netG = nil + if which_model_netG == "encoder_decoder" then netG = defineG_encoder_decoder(input_nc, output_nc, ngf) + elseif which_model_netG == "unet128" then netG = defineG_unet128(input_nc, output_nc, ngf) + elseif which_model_netG == "unet256" then netG = defineG_unet256(input_nc, output_nc, ngf) + elseif which_model_netG == "resnet_6blocks" then netG = defineG_resnet_6blocks(input_nc, output_nc, ngf) + elseif which_model_netG == "resnet_9blocks" then netG = defineG_resnet_9blocks(input_nc, output_nc, ngf) + else error("unsupported netG model") + end + netG:apply(weights_init) + + return netG +end + +function defineD(input_nc, ndf, which_model_netD, n_layers_D, use_sigmoid) + local netD = nil + if which_model_netD == "basic" then netD = defineD_basic(input_nc, ndf, use_sigmoid) + elseif which_model_netD == "imageGAN" then netD = defineD_imageGAN(input_nc, ndf, use_sigmoid) + elseif which_model_netD == "n_layers" then netD = defineD_n_layers(input_nc, ndf, n_layers_D, use_sigmoid) + else error("unsupported netD model") + end + netD:apply(weights_init) + + return netD +end + +function defineG_encoder_decoder(input_nc, output_nc, ngf) + -- input is (nc) x 256 x 256 + local e1 = - nn.SpatialConvolution(input_nc, ngf, 4, 4, 2, 2, 1, 1) + -- input is (ngf) x 128 x 128 + local e2 = e1 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf, ngf * 2, 4, 4, 2, 2, 1, 1) - normalization(ngf * 2) + -- input is (ngf * 2) x 64 x 64 + local e3 = e2 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 2, ngf * 4, 4, 4, 2, 2, 1, 1) - normalization(ngf * 4) + -- input is (ngf * 4) x 32 x 32 + local e4 = e3 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 4, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 16 x 16 + local e5 = e4 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 8 x 8 + local e6 = e5 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 4 x 4 + local e7 = e6 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 2 x 2 + local e8 = e7 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) -- normalization(ngf * 8) + -- input is (ngf * 8) x 1 x 1 + + local d1 = e8 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 2 x 2 + local d2 = d1 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 4 x 4 + local d3 = d2 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 8 x 8 + local d4 = d3 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 16 x 16 + local d5 = d4 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8, ngf * 4, 4, 4, 2, 2, 1, 1) - normalization(ngf * 4) + -- input is (ngf * 4) x 32 x 32 + local d6 = d5 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 4, ngf * 2, 4, 4, 2, 2, 1, 1) - normalization(ngf * 2) + -- input is (ngf * 2) x 64 x 64 + local d7 = d6 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 2, ngf, 4, 4, 2, 2, 1, 1) - normalization(ngf) + -- input is (ngf) x128 x 128 + local d8 = d7 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf, output_nc, 4, 4, 2, 2, 1, 1) + -- input is (nc) x 256 x 256 + local o1 = d8 - nn.Tanh() + + local netG = nn.gModule({e1},{o1}) + return netG +end + + +function defineG_unet128(input_nc, output_nc, ngf) + local netG = nil + -- input is (nc) x 128 x 128 + local e1 = - nn.SpatialConvolution(input_nc, ngf, 4, 4, 2, 2, 1, 1) + -- input is (ngf) x 64 x 64 + local e2 = e1 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf, ngf * 2, 4, 4, 2, 2, 1, 1) - normalization(ngf * 2) + -- input is (ngf * 2) x 32 x 32 + local e3 = e2 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 2, ngf * 4, 4, 4, 2, 2, 1, 1) - normalization(ngf * 4) + -- input is (ngf * 4) x 16 x 16 + local e4 = e3 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 4, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 8 x 8 + local e5 = e4 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 4 x 4 + local e6 = e5 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 2 x 2 + local e7 = e6 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) -- normalization(ngf * 8) + -- input is (ngf * 8) x 1 x 1 + + local d1_ = e7 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 2 x 2 + local d1 = {d1_,e6} - nn.JoinTable(2) + local d2_ = d1 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 4 x 4 + local d2 = {d2_,e5} - nn.JoinTable(2) + local d3_ = d2 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 8 x 8 + local d3 = {d3_,e4} - nn.JoinTable(2) + local d4_ = d3 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 4, 4, 4, 2, 2, 1, 1) - normalization(ngf * 4) + -- input is (ngf * 8) x 16 x 16 + local d4 = {d4_,e3} - nn.JoinTable(2) + local d5_ = d4 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 4 * 2, ngf * 2, 4, 4, 2, 2, 1, 1) - normalization(ngf * 2) + -- input is (ngf * 4) x 32 x 32 + local d5 = {d5_,e2} - nn.JoinTable(2) + local d6_ = d5 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 2 * 2, ngf, 4, 4, 2, 2, 1, 1) - normalization(ngf) + -- input is (ngf * 2) x 64 x 64 + local d6 = {d6_,e1} - nn.JoinTable(2) + + local d7 = d6 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 2, output_nc, 4, 4, 2, 2, 1, 1) + -- input is (nc) x 128 x 128 + + local o1 = d7 - nn.Tanh() + local netG = nn.gModule({e1},{o1}) + return netG +end + + +function defineG_unet256(input_nc, output_nc, ngf) + local netG = nil + -- input is (nc) x 256 x 256 + local e1 = - nn.SpatialConvolution(input_nc, ngf, 4, 4, 2, 2, 1, 1) + -- input is (ngf) x 128 x 128 + local e2 = e1 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf, ngf * 2, 4, 4, 2, 2, 1, 1) - normalization(ngf * 2) + -- input is (ngf * 2) x 64 x 64 + local e3 = e2 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 2, ngf * 4, 4, 4, 2, 2, 1, 1) - normalization(ngf * 4) + -- input is (ngf * 4) x 32 x 32 + local e4 = e3 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 4, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 16 x 16 + local e5 = e4 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 8 x 8 + local e6 = e5 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 4 x 4 + local e7 = e6 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 2 x 2 + local e8 = e7 - nn.LeakyReLU(0.2, true) - nn.SpatialConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) -- - normalization(ngf * 8) + -- input is (ngf * 8) x 1 x 1 + + local d1_ = e8 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 2 x 2 + local d1 = {d1_,e7} - nn.JoinTable(2) + local d2_ = d1 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 4 x 4 + local d2 = {d2_,e6} - nn.JoinTable(2) + local d3_ = d2 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) - nn.Dropout(0.5) + -- input is (ngf * 8) x 8 x 8 + local d3 = {d3_,e5} - nn.JoinTable(2) + local d4_ = d3 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 8, 4, 4, 2, 2, 1, 1) - normalization(ngf * 8) + -- input is (ngf * 8) x 16 x 16 + local d4 = {d4_,e4} - nn.JoinTable(2) + local d5_ = d4 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 8 * 2, ngf * 4, 4, 4, 2, 2, 1, 1) - normalization(ngf * 4) + -- input is (ngf * 4) x 32 x 32 + local d5 = {d5_,e3} - nn.JoinTable(2) + local d6_ = d5 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 4 * 2, ngf * 2, 4, 4, 2, 2, 1, 1) - normalization(ngf * 2) + -- input is (ngf * 2) x 64 x 64 + local d6 = {d6_,e2} - nn.JoinTable(2) + local d7_ = d6 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 2 * 2, ngf, 4, 4, 2, 2, 1, 1) - normalization(ngf) + -- input is (ngf) x128 x 128 + local d7 = {d7_,e1} - nn.JoinTable(2) + local d8 = d7 - nn.ReLU(true) - nn.SpatialFullConvolution(ngf * 2, output_nc, 4, 4, 2, 2, 1, 1) + -- input is (nc) x 256 x 256 + + local o1 = d8 - nn.Tanh() + local netG = nn.gModule({e1},{o1}) + return netG +end + +-------------------------------------------------------------------------------- +-- Justin Johnson's model from https://github.com/jcjohnson/fast-neural-style/ +-------------------------------------------------------------------------------- + +local function build_conv_block(dim, padding_type) + local conv_block = nn.Sequential() + local p = 0 + if padding_type == 'reflect' then + conv_block:add(nn.SpatialReflectionPadding(1, 1, 1, 1)) + elseif padding_type == 'replicate' then + conv_block:add(nn.SpatialReplicationPadding(1, 1, 1, 1)) + elseif padding_type == 'zero' then + p = 1 + end + conv_block:add(nn.SpatialConvolution(dim, dim, 3, 3, 1, 1, p, p)) + conv_block:add(normalization(dim)) + conv_block:add(nn.ReLU(true)) + if padding_type == 'reflect' then + conv_block:add(nn.SpatialReflectionPadding(1, 1, 1, 1)) + elseif padding_type == 'replicate' then + conv_block:add(nn.SpatialReplicationPadding(1, 1, 1, 1)) + end + conv_block:add(nn.SpatialConvolution(dim, dim, 3, 3, 1, 1, p, p)) + conv_block:add(normalization(dim)) + return conv_block +end + + +local function build_res_block(dim, padding_type) + local conv_block = build_conv_block(dim, padding_type) + local res_block = nn.Sequential() + local concat = nn.ConcatTable() + concat:add(conv_block) + concat:add(nn.Identity()) + + res_block:add(concat):add(nn.CAddTable()) + return res_block +end + +function defineG_resnet_6blocks(input_nc, output_nc, ngf) + padding_type = 'reflect' + local ks = 3 + local netG = nil + local f = 7 + local p = (f - 1) / 2 + local data = -nn.Identity() + local e1 = data - nn.SpatialReflectionPadding(p, p, p, p) - nn.SpatialConvolution(input_nc, ngf, f, f, 1, 1) - normalization(ngf) - nn.ReLU(true) + local e2 = e1 - nn.SpatialConvolution(ngf, ngf*2, ks, ks, 2, 2, 1, 1) - normalization(ngf*2) - nn.ReLU(true) + local e3 = e2 - nn.SpatialConvolution(ngf*2, ngf*4, ks, ks, 2, 2, 1, 1) - normalization(ngf*4) - nn.ReLU(true) + local d1 = e3 - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) + - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) + local d2 = d1 - nn.SpatialFullConvolution(ngf*4, ngf*2, ks, ks, 2, 2, 1, 1,1,1) - normalization(ngf*2) - nn.ReLU(true) + local d3 = d2 - nn.SpatialFullConvolution(ngf*2, ngf, ks, ks, 2, 2, 1, 1,1,1) - normalization(ngf) - nn.ReLU(true) + local d4 = d3 - nn.SpatialReflectionPadding(p, p, p, p) - nn.SpatialConvolution(ngf, output_nc, f, f, 1, 1) - nn.Tanh() + netG = nn.gModule({data},{d4}) + return netG +end + +function defineG_resnet_9blocks(input_nc, output_nc, ngf) + padding_type = 'reflect' + local ks = 3 + local netG = nil + local f = 7 + local p = (f - 1) / 2 + local data = -nn.Identity() + local e1 = data - nn.SpatialReflectionPadding(p, p, p, p) - nn.SpatialConvolution(input_nc, ngf, f, f, 1, 1) - normalization(ngf) - nn.ReLU(true) + local e2 = e1 - nn.SpatialConvolution(ngf, ngf*2, ks, ks, 2, 2, 1, 1) - normalization(ngf*2) - nn.ReLU(true) + local e3 = e2 - nn.SpatialConvolution(ngf*2, ngf*4, ks, ks, 2, 2, 1, 1) - normalization(ngf*4) - nn.ReLU(true) + local d1 = e3 - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) + - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) + - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) - build_res_block(ngf*4, padding_type) + local d2 = d1 - nn.SpatialFullConvolution(ngf*4, ngf*2, ks, ks, 2, 2, 1, 1,1,1) - normalization(ngf*2) - nn.ReLU(true) + local d3 = d2 - nn.SpatialFullConvolution(ngf*2, ngf, ks, ks, 2, 2, 1, 1,1,1) - normalization(ngf) - nn.ReLU(true) + local d4 = d3 - nn.SpatialReflectionPadding(p, p, p, p) - nn.SpatialConvolution(ngf, output_nc, f, f, 1, 1) - nn.Tanh() + netG = nn.gModule({data},{d4}) + return netG +end + +function defineD_imageGAN(input_nc, ndf, use_sigmoid) + local netD = nn.Sequential() + + -- input is (nc) x 256 x 256 + netD:add(nn.SpatialConvolution(input_nc, ndf, 4, 4, 2, 2, 1, 1)) + netD:add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf) x 128 x 128 + netD:add(nn.SpatialConvolution(ndf, ndf * 2, 4, 4, 2, 2, 1, 1)) + netD:add(nn.SpatialBatchNormalization(ndf * 2)):add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf*2) x 64 x 64 + netD:add(nn.SpatialConvolution(ndf * 2, ndf*4, 4, 4, 2, 2, 1, 1)) + netD:add(nn.SpatialBatchNormalization(ndf * 4)):add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf*4) x 32 x 32 + netD:add(nn.SpatialConvolution(ndf * 4, ndf * 8, 4, 4, 2, 2, 1, 1)) + netD:add(nn.SpatialBatchNormalization(ndf * 8)):add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf*8) x 16 x 16 + netD:add(nn.SpatialConvolution(ndf * 8, ndf * 8, 4, 4, 2, 2, 1, 1)) + netD:add(nn.SpatialBatchNormalization(ndf * 8)):add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf*8) x 8 x 8 + netD:add(nn.SpatialConvolution(ndf * 8, ndf * 8, 4, 4, 2, 2, 1, 1)) + netD:add(nn.SpatialBatchNormalization(ndf * 8)):add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf*8) x 4 x 4 + netD:add(nn.SpatialConvolution(ndf * 8, 1, 4, 4, 2, 2, 1, 1)) + -- state size: 1 x 1 x 1 + if use_sigmoid then + netD:add(nn.Sigmoid()) + end + + return netD +end + + + +function defineD_basic(input_nc, ndf, use_sigmoid) + n_layers = 3 + return defineD_n_layers(input_nc, ndf, n_layers, use_sigmoid) +end + +-- rf=1 +function defineD_pixelGAN(input_nc, ndf, use_sigmoid) + + local netD = nn.Sequential() + + -- input is (nc) x 256 x 256 + netD:add(nn.SpatialConvolution(input_nc, ndf, 1, 1, 1, 1, 0, 0)) + netD:add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf) x 256 x 256 + netD:add(nn.SpatialConvolution(ndf, ndf * 2, 1, 1, 1, 1, 0, 0)) + netD:add(normalization(ndf * 2)):add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf*2) x 256 x 256 + netD:add(nn.SpatialConvolution(ndf * 2, 1, 1, 1, 1, 1, 0, 0)) + -- state size: 1 x 256 x 256 + if use_sigmoid then + netD:add(nn.Sigmoid()) + -- state size: 1 x 30 x 30 + end + + return netD +end + +-- if n=0, then use pixelGAN (rf=1) +-- else rf is 16 if n=1 +-- 34 if n=2 +-- 70 if n=3 +-- 142 if n=4 +-- 286 if n=5 +-- 574 if n=6 +function defineD_n_layers(input_nc, ndf, n_layers, use_sigmoid, kw, dropout_ratio) + + if dropout_ratio == nil then + dropout_ratio = 0.0 + end + + if kw == nil then + kw = 4 + end + padw = math.ceil((kw-1)/2) + + if n_layers==0 then + return defineD_pixelGAN(input_nc, ndf, use_sigmoid) + else + + local netD = nn.Sequential() + + -- input is (nc) x 256 x 256 + -- print('input_nc', input_nc) + netD:add(nn.SpatialConvolution(input_nc, ndf, kw, kw, 2, 2, padw, padw)) + netD:add(nn.LeakyReLU(0.2, true)) + + local nf_mult = 1 + local nf_mult_prev = 1 + for n = 1, n_layers-1 do + nf_mult_prev = nf_mult + nf_mult = math.min(2^n,8) + netD:add(nn.SpatialConvolution(ndf * nf_mult_prev, ndf * nf_mult, kw, kw, 2, 2, padw,padw)) + netD:add(normalization(ndf * nf_mult)):add(nn.Dropout(dropout_ratio)) + netD:add(nn.LeakyReLU(0.2, true)) + end + + -- state size: (ndf*M) x N x N + nf_mult_prev = nf_mult + nf_mult = math.min(2^n_layers,8) + netD:add(nn.SpatialConvolution(ndf * nf_mult_prev, ndf * nf_mult, kw, kw, 1, 1, padw, padw)) + netD:add(normalization(ndf * nf_mult)):add(nn.LeakyReLU(0.2, true)) + -- state size: (ndf*M*2) x (N-1) x (N-1) + netD:add(nn.SpatialConvolution(ndf * nf_mult, 1, kw, kw, 1, 1, padw,padw)) + -- state size: 1 x (N-2) x (N-2) + if use_sigmoid then + netD:add(nn.Sigmoid()) + end + -- state size: 1 x (N-2) x (N-2) + return netD + end +end diff --git a/cyclegan/models/base_model.lua b/cyclegan/models/base_model.lua new file mode 100644 index 0000000..fedd4bd --- /dev/null +++ b/cyclegan/models/base_model.lua @@ -0,0 +1,66 @@ +-------------------------------------------------------------------------------- +-- Base Class for Providing Models +-------------------------------------------------------------------------------- + +local class = require 'class' + +BaseModel = class('BaseModel') + +function BaseModel:__init(conf) + conf = conf or {} +end + +-- Returns the name of the model +function BaseModel:model_name() + return 'DoesNothingModel' +end + +-- Defines models and networks +function BaseModel:Initialize(opt) + models = {} + return models +end + +-- Runs the forward pass of the network +function BaseModel:Forward(input, opt) + output = {} + return output +end + +-- Runs the backprop gradient descent +-- Corresponds to a single batch of data +function BaseModel:OptimizeParameters(opt) +end + +-- This function can be used to reset momentum after each epoch +function BaseModel:RefreshParameters(opt) +end + +-- This function can be used to reset momentum after each epoch +function BaseModel:UpdateLearningRate(opt) +end +-- Save the current model to the file system +function BaseModel:Save(prefix, opt) +end + +-- returns a string that describes the current errors +function BaseModel:GetCurrentErrorDescription() + return "No Error exists in BaseModel" +end + +-- returns current errors +function BaseModel:GetCurrentErrors(opt) + return {} +end + +-- returns a table of image/label pairs that describe +-- the current results. +-- |return|: a table of table. List of image/label pairs +function BaseModel:GetCurrentVisuals(opt, size) + return {} +end + +-- returns a string that describes the display plot configuration +function BaseModel:DisplayPlot(opt) + return {} +end diff --git a/cyclegan/models/bigan_model.lua b/cyclegan/models/bigan_model.lua new file mode 100644 index 0000000..7bd6df8 --- /dev/null +++ b/cyclegan/models/bigan_model.lua @@ -0,0 +1,254 @@ +local class = require 'class' +require 'models.base_model' +require 'models.architectures' +require 'util.image_pool' +util = paths.dofile('../util/util.lua') +content = paths.dofile('../util/content_loss.lua') + +BiGANModel = class('BiGANModel', 'BaseModel') + +function BiGANModel:__init(conf) + BaseModel.__init(self, conf) + conf = conf or {} +end + +function BiGANModel:model_name() + return 'BiGANModel' +end + +function BiGANModel:InitializeStates(use_wgan) + optimState = {learningRate=opt.lr, beta1=opt.beta1,} + return optimState +end +-- Defines models and networks +function BiGANModel:Initialize(opt) + if opt.test == 0 then + self.realABPool = ImagePool(opt.pool_size) + self.fakeABPool = ImagePool(opt.pool_size) + end + -- define tensors + local d_input_nc = opt.input_nc + opt.output_nc + self.real_AB = torch.Tensor(opt.batchSize, d_input_nc, opt.fineSize, opt.fineSize) + self.fake_AB = torch.Tensor(opt.batchSize, d_input_nc, opt.fineSize, opt.fineSize) + -- load/define models + self.criterionGAN = nn.MSECriterion() + + local netG, netE, netD = nil, nil, nil + if opt.continue_train == 1 then + if opt.test == 1 then -- which_epoch option exists in test mode + netG = util.load_test_model('G', opt) + netE = util.load_test_model('E', opt) + netD = util.load_test_model('D', opt) + else + netG = util.load_model('G', opt) + netE = util.load_model('E', opt) + netD = util.load_model('D', opt) + end + else + -- netG_test = defineG(opt.input_nc, opt.output_nc, opt.ngf, "resnet_unet", opt.arch) + -- os.exit() + netD = defineD(d_input_nc, opt.ndf, opt.which_model_netD, opt.n_layers_D, false) -- no sigmoid layer + print('netD...', netD) + netG = defineG(opt.input_nc, opt.output_nc, opt.ngf, opt.which_model_netG, opt.arch) + print('netG...', netG) + netE = defineG(opt.output_nc, opt.input_nc, opt.ngf, opt.which_model_netG, opt.arch) + print('netE...', netE) + + end + + self.netD = netD + self.netG = netG + self.netE = netE + + -- define real/fake labels + netD_output_size = self.netD:forward(self.real_AB):size() + self.fake_label = torch.Tensor(netD_output_size):fill(0.0) + self.real_label = torch.Tensor(netD_output_size):fill(1.0) -- no soft smoothing + + self.optimStateD = self:InitializeStates() + self.optimStateG = self:InitializeStates() + self.optimStateE = self:InitializeStates() + self.A_idx = {{}, {1, opt.input_nc}, {}, {}} + self.B_idx = {{}, {opt.input_nc+1, opt.input_nc+opt.output_nc}, {}, {}} + self:RefreshParameters() + + print('---------- # Learnable Parameters --------------') + print(('G = %d'):format(self.parametersG:size(1))) + print(('E = %d'):format(self.parametersE:size(1))) + print(('D = %d'):format(self.parametersD:size(1))) + print('------------------------------------------------') + -- os.exit() +end + +-- Runs the forward pass of the network and +-- saves the result to member variables of the class +function BiGANModel:Forward(input, opt) + if opt.which_direction == 'BtoA' then + local temp = input.real_A + input.real_A = input.real_B + input.real_B = temp + end + self.real_AB[self.A_idx]:copy(input.real_A) + self.fake_AB[self.B_idx]:copy(input.real_B) + self.real_A = self.real_AB[self.A_idx] + self.real_B = self.fake_AB[self.B_idx] + self.fake_B = self.netG:forward(self.real_A):clone() + self.fake_A = self.netE:forward(self.real_B):clone() + self.real_AB[self.B_idx]:copy(self.fake_B) -- real_AB: real_A, fake_B -> real_label + self.fake_AB[self.A_idx]:copy(self.fake_A) -- fake_AB: fake_A, real_B -> fake_label + -- if opt.test == 0 then + -- self.real_AB = self.realABPool:Query(self.real_AB) -- batch history + -- self.fake_AB = self.fakeABPool:Query(self.fake_AB) -- batch history + -- end +end + +-- create closure to evaluate f(X) and df/dX of discriminator +function BiGANModel:fDx_basic(x, gradParams, netD, real_AB, fake_AB, opt) + util.BiasZero(netD) + gradParams:zero() + -- Real log(D_A(B)) + local output = netD:forward(real_AB):clone() + local errD_real = self.criterionGAN:forward(output, self.real_label) + local df_do = self.criterionGAN:backward(output, self.real_label) + netD:backward(real_AB, df_do) + -- Fake + log(1 - D_A(G(A))) + output = netD:forward(fake_AB):clone() + local errD_fake = self.criterionGAN:forward(output, self.fake_label) + local df_do2 = self.criterionGAN:backward(output, self.fake_label) + netD:backward(fake_AB, df_do2) + -- Compute loss + local errD = (errD_real + errD_fake) / 2.0 + return errD, gradParams +end + + +function BiGANModel:fDx(x, opt) + -- use image pool that stores the old fake images + real_AB = self.realABPool:Query(self.real_AB) + fake_AB = self.fakeABPool:Query(self.fake_AB) + self.errD, gradParams = self:fDx_basic(x, self.gradParametersD, self.netD, real_AB, fake_AB, opt) + return self.errD, gradParams +end + + + +function BiGANModel:fGx_basic(x, netG, netD, gradParametersG, opt) + util.BiasZero(netG) + util.BiasZero(netD) + gradParametersG:zero() + + -- First. G(A) should fake the discriminator + local output = netD:forward(self.real_AB):clone() + local errG = self.criterionGAN:forward(output, self.fake_label) + local dgan_loss_dd = self.criterionGAN:backward(output, self.fake_label) + local dgan_loss_do = netD:updateGradInput(self.real_AB, dgan_loss_dd) + netG:backward(self.real_A, dgan_loss_do[self.B_idx]) -- real_AB: real_A, fake_B -> real_label + return gradParametersG, errG +end + + +function BiGANModel:fGx(x, opt) + self.gradParametersG, self.errG = self:fGx_basic(x, self.netG, self.netD, + self.gradParametersG, opt) + return self.errG, self.gradParametersG +end + + +function BiGANModel:fEx_basic(x, netE, netD, gradParametersE, opt) + util.BiasZero(netE) + util.BiasZero(netD) + gradParametersE:zero() + + -- First. G(A) should fake the discriminator + local output = netD:forward(self.fake_AB):clone() + local errE= self.criterionGAN:forward(output, self.real_label) + local dgan_loss_dd = self.criterionGAN:backward(output, self.real_label) + local dgan_loss_do = netD:updateGradInput(self.fake_AB, dgan_loss_dd) + netE:backward(self.real_B, dgan_loss_do[self.A_idx])-- fake_AB: fake_A, real_B -> fake_label + return gradParametersE, errE +end + + +function BiGANModel:fEx(x, opt) + self.gradParametersE, self.errE = self:fEx_basic(x, self.netE, self.netD, + self.gradParametersE, opt) + return self.errE, self.gradParametersE +end + + +function BiGANModel:OptimizeParameters(opt) + local fG = function(x) return self:fGx(x, opt) end + local fE = function(x) return self:fEx(x, opt) end + local fD = function(x) return self:fDx(x, opt) end + optim.adam(fD, self.parametersD, self.optimStateD) + optim.adam(fG, self.parametersG, self.optimStateG) + optim.adam(fE, self.parametersE, self.optimStateE) +end + +function BiGANModel:RefreshParameters() + self.parametersD, self.gradParametersD = nil, nil -- nil them to avoid spiking memory + self.parametersG, self.gradParametersG = nil, nil + self.parametersE, self.gradParametersE = nil, nil + -- define parameters of optimization + self.parametersD, self.gradParametersD = self.netD:getParameters() + self.parametersG, self.gradParametersG = self.netG:getParameters() + self.parametersE, self.gradParametersE = self.netE:getParameters() +end + +function BiGANModel:Save(prefix, opt) + util.save_model(self.netG, prefix .. '_net_G.t7', 1) + util.save_model(self.netE, prefix .. '_net_E.t7', 1) + util.save_model(self.netD, prefix .. '_net_D.t7', 1) +end + +function BiGANModel:GetCurrentErrorDescription() + description = ('D: %.4f G: %.4f E: %.4f'):format( + self.errD and self.errD or -1, + self.errG and self.errG or -1, + self.errE and self.errE or -1) + return description +end + +function BiGANModel:GetCurrentErrors() + local errors = {errD=self.errD, errG=self.errG, errE=self.errE} + return errors +end + +-- returns a string that describes the display plot configuration +function BiGANModel:DisplayPlot(opt) + return 'errD,errG,errE' +end +function BiGANModel:UpdateLearningRate(opt) + local lrd = opt.lr / opt.niter_decay + local old_lr = self.optimStateD['learningRate'] + local lr = old_lr - lrd + self.optimStateD['learningRate'] = lr + self.optimStateG['learningRate'] = lr + self.optimStateE['learningRate'] = lr + print(('update learning rate: %f -> %f'):format(old_lr, lr)) +end + +local function MakeIm3(im) + -- print('before im_size', im:size()) + local im3 = nil + if im:size(2) == 1 then + im3 = torch.repeatTensor(im, 1,3,1,1) + else + im3 = im + end + -- print('after im_size', im:size()) + -- print('after im3_size', im3:size()) + return im3 +end +function BiGANModel:GetCurrentVisuals(opt, size) + if not size then + size = opt.display_winsize + end + + local visuals = {} + table.insert(visuals, {img=MakeIm3(self.real_A), label='real_A'}) + table.insert(visuals, {img=MakeIm3(self.fake_B), label='fake_B'}) + table.insert(visuals, {img=MakeIm3(self.real_B), label='real_B'}) + table.insert(visuals, {img=MakeIm3(self.fake_A), label='fake_A'}) + return visuals +end diff --git a/cyclegan/models/content_gan_model.lua b/cyclegan/models/content_gan_model.lua new file mode 100644 index 0000000..a3e7059 --- /dev/null +++ b/cyclegan/models/content_gan_model.lua @@ -0,0 +1,201 @@ +local class = require 'class' +require 'models.base_model' +require 'models.architectures' +require 'util.image_pool' +util = paths.dofile('../util/util.lua') +content = paths.dofile('../util/content_loss.lua') + +ContentGANModel = class('ContentGANModel', 'BaseModel') + +function ContentGANModel:__init(conf) + BaseModel.__init(self, conf) + conf = conf or {} +end + +function ContentGANModel:model_name() + return 'ContentGANModel' +end + +function ContentGANModel:InitializeStates() + local optimState = {learningRate=opt.lr, beta1=opt.beta1,} + return optimState +end +-- Defines models and networks +function ContentGANModel:Initialize(opt) + if opt.test == 0 then + self.fakePool = ImagePool(opt.pool_size) + end + -- define tensors + self.real_A = torch.Tensor(opt.batchSize, opt.input_nc, opt.fineSize, opt.fineSize) + self.fake_B = torch.Tensor(opt.batchSize, opt.output_nc, opt.fineSize, opt.fineSize) + self.real_B = self.fake_B:clone() --torch.Tensor(opt.batchSize, opt.output_nc, opt.fineSize, opt.fineSize) + + -- load/define models + self.criterionGAN = nn.MSECriterion() + self.criterionContent = nn.AbsCriterion() + self.contentFunc = content.defineContent(opt.content_loss, opt.layer_name) + self.netG, self.netD = nil, nil + if opt.continue_train == 1 then + if opt.which_epoch then -- which_epoch option exists in test mode + self.netG = util.load_test_model('G_A', opt) + self.netD = util.load_test_model('D_A', opt) + else + self.netG = util.load_model('G_A', opt) + self.netD = util.load_model('D_A', opt) + end + else + self.netG = defineG(opt.input_nc, opt.output_nc, opt.ngf, opt.which_model_netG) + print('netG...', self.netG) + self.netD = defineD(opt.output_nc, opt.ndf, opt.which_model_netD, opt.n_layers_D, false) + print('netD...', self.netD) + end + -- define real/fake labels + netD_output_size = self.netD:forward(self.real_A):size() + self.fake_label = torch.Tensor(netD_output_size):fill(0.0) + self.real_label = torch.Tensor(netD_output_size):fill(1.0) -- no soft smoothing + self.optimStateD = self:InitializeStates() + self.optimStateG = self:InitializeStates() + self:RefreshParameters() + print('---------- # Learnable Parameters --------------') + print(('G = %d'):format(self.parametersG:size(1))) + print(('D = %d'):format(self.parametersD:size(1))) + print('------------------------------------------------') + -- os.exit() +end + +-- Runs the forward pass of the network and +-- saves the result to member variables of the class +function ContentGANModel:Forward(input, opt) + if opt.which_direction == 'BtoA' then + local temp = input.real_A + input.real_A = input.real_B + input.real_B = temp + end + + self.real_A:copy(input.real_A) + self.real_B:copy(input.real_B) + self.fake_B = self.netG:forward(self.real_A):clone() + -- output = {self.fake_B} + output = {} + -- if opt.test == 1 then + + -- end + return output +end + +-- create closure to evaluate f(X) and df/dX of discriminator +function ContentGANModel:fDx_basic(x, gradParams, netD, netG, + real_target, fake_target, opt) + util.BiasZero(netD) + util.BiasZero(netG) + gradParams:zero() + + local errD_real, errD_rec, errD_fake, errD = 0, 0, 0, 0 + -- Real log(D_A(B)) + local output = netD:forward(real_target) + errD_real = self.criterionGAN:forward(output, self.real_label) + df_do = self.criterionGAN:backward(output, self.real_label) + netD:backward(real_target, df_do) + + -- Fake + log(1 - D_A(G_A(A))) + output = netD:forward(fake_target) + errD_fake = self.criterionGAN:forward(output, self.fake_label) + df_do = self.criterionGAN:backward(output, self.fake_label) + netD:backward(fake_target, df_do) + errD = (errD_real + errD_fake) / 2.0 + -- print('errD', errD + return errD, gradParams +end + + +function ContentGANModel:fDx(x, opt) + fake_B = self.fakePool:Query(self.fake_B) + self.errD, gradParams = self:fDx_basic(x, self.gradparametersD, self.netD, self.netG, + self.real_B, fake_B, opt) + return self.errD, gradParams +end + +function ContentGANModel:fGx_basic(x, netG_source, netD_source, real_source, real_target, fake_target, + gradParametersG_source, opt) + util.BiasZero(netD_source) + util.BiasZero(netG_source) + gradParametersG_source:zero() + -- GAN loss + -- local df_d_GAN = torch.zeros(fake_target:size()) + -- local errGAN = 0 + -- local errRec = 0 + --- Domain GAN loss: D_A(G_A(A)) + local output = netD_source.output -- [hack] forward was already executed in fDx, so save computation netD_source:forward(fake_B) --- + local errGAN = self.criterionGAN:forward(output, self.real_label) + local df_do = self.criterionGAN:backward(output, self.real_label) + local df_d_GAN = netD_source:updateGradInput(fake_target, df_do) ---:narrow(2,fake_AB:size(2)-output_nc+1, output_nc) + + -- content loss + -- print('content_loss', opt.content_loss) + -- function content.lossUpdate(criterionContent, real_source, fake_target, contentFunc, loss_type, weight) + local errContent, df_d_content = content.lossUpdate(self.criterionContent, real_source, fake_target, self.contentFunc, opt.content_loss, opt.lambda_A) + netG_source:forward(real_source) + netG_source:backward(real_source, df_d_GAN + df_d_content) + -- print('errD', errGAN) + return gradParametersG_source, errGAN, errContent +end + +function ContentGANModel:fGx(x, opt) + self.gradparametersG, self.errG, self.errCont = + self:fGx_basic(x, self.netG, self.netD, + self.real_A, self.real_B, self.fake_B, + self.gradparametersG, opt) + return self.errG, self.gradparametersG +end + +function ContentGANModel:OptimizeParameters(opt) + local fDx = function(x) return self:fDx(x, opt) end + local fGx = function(x) return self:fGx(x, opt) end + optim.adam(fDx, self.parametersD, self.optimStateD) + optim.adam(fGx, self.parametersG, self.optimStateG) +end + +function ContentGANModel:RefreshParameters() + self.parametersD, self.gradparametersD = nil, nil -- nil them to avoid spiking memory + self.parametersG, self.gradparametersG = nil, nil + -- define parameters of optimization + self.parametersG, self.gradparametersG = self.netG:getParameters() + self.parametersD, self.gradparametersD = self.netD:getParameters() +end + +function ContentGANModel:Save(prefix, opt) + util.save_model(self.netG, prefix .. '_net_G_A.t7', 1.0) + util.save_model(self.netD, prefix .. '_net_D_A.t7', 1.0) +end + +function ContentGANModel:GetCurrentErrorDescription() + description = ('G: %.4f D: %.4f Content: %.4f'):format(self.errG and self.errG or -1, + self.errD and self.errD or -1, + self.errCont and self.errCont or -1) + return description +end + + +function ContentGANModel:GetCurrentErrors() + local errors = {errG=self.errG and self.errG or -1, errD=self.errD and self.errD or -1, + errCont=self.errCont and self.errCont or -1} + return errors +end + +-- returns a string that describes the display plot configuration +function ContentGANModel:DisplayPlot(opt) + return 'errG,errD,errCont' +end + + +function ContentGANModel:GetCurrentVisuals(opt, size) + if not size then + size = opt.display_winsize + end + + local visuals = {} + table.insert(visuals, {img=self.real_A, label='real_A'}) + table.insert(visuals, {img=self.fake_B, label='fake_B'}) + table.insert(visuals, {img=self.real_B, label='real_B'}) + return visuals +end diff --git a/cyclegan/models/cycle_gan_model.lua b/cyclegan/models/cycle_gan_model.lua new file mode 100644 index 0000000..c1489c4 --- /dev/null +++ b/cyclegan/models/cycle_gan_model.lua @@ -0,0 +1,324 @@ +local class = require 'class' +require 'models.base_model' +require 'models.architectures' +require 'util.image_pool' + +util = paths.dofile('../util/util.lua') +CycleGANModel = class('CycleGANModel', 'BaseModel') + +function CycleGANModel:__init(conf) + BaseModel.__init(self, conf) + conf = conf or {} +end + +function CycleGANModel:model_name() + return 'CycleGANModel' +end + +function CycleGANModel:InitializeStates(use_wgan) + optimState = {learningRate=opt.lr, beta1=opt.beta1,} + return optimState +end +-- Defines models and networks +function CycleGANModel:Initialize(opt) + if opt.test == 0 then + self.fakeAPool = ImagePool(opt.pool_size) + self.fakeBPool = ImagePool(opt.pool_size) + end + -- define tensors + if opt.test == 0 then -- allocate tensors for training + self.real_A = torch.Tensor(opt.batchSize, opt.input_nc, opt.fineSize, opt.fineSize) + self.real_B = torch.Tensor(opt.batchSize, opt.output_nc, opt.fineSize, opt.fineSize) + self.fake_A = torch.Tensor(opt.batchSize, opt.input_nc, opt.fineSize, opt.fineSize) + self.fake_B = torch.Tensor(opt.batchSize, opt.output_nc, opt.fineSize, opt.fineSize) + self.rec_A = torch.Tensor(opt.batchSize, opt.input_nc, opt.fineSize, opt.fineSize) + self.rec_B = torch.Tensor(opt.batchSize, opt.output_nc, opt.fineSize, opt.fineSize) + end + -- load/define models + local use_lsgan = ((opt.use_lsgan ~= nil) and (opt.use_lsgan == 1)) + if not use_lsgan then + self.criterionGAN = nn.BCECriterion() + else + self.criterionGAN = nn.MSECriterion() + end + self.criterionRec = nn.AbsCriterion() + + local netG_A, netD_A, netG_B, netD_B = nil, nil, nil, nil + if opt.continue_train == 1 then + if opt.test == 1 then -- test mode + netG_A = util.load_test_model('G_A', opt) + netG_B = util.load_test_model('G_B', opt) + + --setup optnet to save a little bit of memory + if opt.use_optnet == 1 then + local sample_input = torch.randn(1, opt.input_nc, 2, 2) + local optnet = require 'optnet' + optnet.optimizeMemory(netG_A, sample_input, {inplace=true, reuseBuffers=true}) + optnet.optimizeMemory(netG_B, sample_input, {inplace=true, reuseBuffers=true}) + end + else + netG_A = util.load_model('G_A', opt) + netG_B = util.load_model('G_B', opt) + netD_A = util.load_model('D_A', opt) + netD_B = util.load_model('D_B', opt) + end + else + local use_sigmoid = (not use_lsgan) + -- netG_test = defineG(opt.input_nc, opt.output_nc, opt.ngf, "resnet_unet", opt.arch) + -- os.exit() + netG_A = defineG(opt.input_nc, opt.output_nc, opt.ngf, opt.which_model_netG, opt.arch) + print('netG_A...', netG_A) + netD_A = defineD(opt.output_nc, opt.ndf, opt.which_model_netD, opt.n_layers_D, use_sigmoid) -- no sigmoid layer + print('netD_A...', netD_A) + netG_B = defineG(opt.output_nc, opt.input_nc, opt.ngf, opt.which_model_netG, opt.arch) + print('netG_B...', netG_B) + netD_B = defineD(opt.input_nc, opt.ndf, opt.which_model_netD, opt.n_layers_D, use_sigmoid) -- no sigmoid layer + print('netD_B', netD_B) + end + + self.netD_A = netD_A + self.netG_A = netG_A + self.netG_B = netG_B + self.netD_B = netD_B + + -- define real/fake labels + if opt.test == 0 then + local D_A_size = self.netD_A:forward(self.real_B):size() -- hack: assume D_size_A = D_size_B + self.fake_label_A = torch.Tensor(D_A_size):fill(0.0) + self.real_label_A = torch.Tensor(D_A_size):fill(1.0) -- no soft smoothing + local D_B_size = self.netD_B:forward(self.real_A):size() -- hack: assume D_size_A = D_size_B + self.fake_label_B = torch.Tensor(D_B_size):fill(0.0) + self.real_label_B = torch.Tensor(D_B_size):fill(1.0) -- no soft smoothing + self.optimStateD_A = self:InitializeStates() + self.optimStateG_A = self:InitializeStates() + self.optimStateD_B = self:InitializeStates() + self.optimStateG_B = self:InitializeStates() + self:RefreshParameters() + print('---------- # Learnable Parameters --------------') + print(('G_A = %d'):format(self.parametersG_A:size(1))) + print(('D_A = %d'):format(self.parametersD_A:size(1))) + print(('G_B = %d'):format(self.parametersG_B:size(1))) + print(('D_B = %d'):format(self.parametersD_B:size(1))) + print('------------------------------------------------') + end +end + +-- Runs the forward pass of the network and +-- saves the result to member variables of the class +function CycleGANModel:Forward(input, opt) + if opt.which_direction == 'BtoA' then + local temp = input.real_A:clone() + input.real_A = input.real_B:clone() + input.real_B = temp + end + + if opt.test == 0 then + self.real_A:copy(input.real_A) + self.real_B:copy(input.real_B) + end + + if opt.test == 1 then -- forward for test + if opt.gpu > 0 then + self.real_A = input.real_A:cuda() + self.real_B = input.real_B:cuda() + else + self.real_A = input.real_A:clone() + self.real_B = input.real_B:clone() + end + self.fake_B = self.netG_A:forward(self.real_A):clone() + self.fake_A = self.netG_B:forward(self.real_B):clone() + self.rec_A = self.netG_B:forward(self.fake_B):clone() + self.rec_B = self.netG_A:forward(self.fake_A):clone() + end +end + +-- create closure to evaluate f(X) and df/dX of discriminator +function CycleGANModel:fDx_basic(x, gradParams, netD, netG, real, fake, real_label, fake_label, opt) + util.BiasZero(netD) + util.BiasZero(netG) + gradParams:zero() + -- Real log(D_A(B)) + local output = netD:forward(real) + local errD_real = self.criterionGAN:forward(output, real_label) + local df_do = self.criterionGAN:backward(output, real_label) + netD:backward(real, df_do) + -- Fake + log(1 - D_A(G_A(A))) + output = netD:forward(fake) + local errD_fake = self.criterionGAN:forward(output, fake_label) + local df_do2 = self.criterionGAN:backward(output, fake_label) + netD:backward(fake, df_do2) + -- Compute loss + local errD = (errD_real + errD_fake) / 2.0 + return errD, gradParams +end + + +function CycleGANModel:fDAx(x, opt) + -- use image pool that stores the old fake images + fake_B = self.fakeBPool:Query(self.fake_B) + self.errD_A, gradParams = self:fDx_basic(x, self.gradparametersD_A, self.netD_A, self.netG_A, + self.real_B, fake_B, self.real_label_A, self.fake_label_A, opt) + return self.errD_A, gradParams +end + + +function CycleGANModel:fDBx(x, opt) + -- use image pool that stores the old fake images + fake_A = self.fakeAPool:Query(self.fake_A) + self.errD_B, gradParams = self:fDx_basic(x, self.gradparametersD_B, self.netD_B, self.netG_B, + self.real_A, fake_A, self.real_label_B, self.fake_label_B, opt) + return self.errD_B, gradParams +end + + +function CycleGANModel:fGx_basic(x, gradParams, netG, netD, netE, real, real2, real_label, lambda1, lambda2, opt) + util.BiasZero(netD) + util.BiasZero(netG) + util.BiasZero(netE) -- inverse mapping + gradParams:zero() + + -- G should be identity if real2 is fed. + local errI = nil + local identity = nil + if opt.lambda_identity > 0 then + identity = netG:forward(real2):clone() + errI = self.criterionRec:forward(identity, real2) * lambda2 * opt.lambda_identity + local didentity_loss_do = self.criterionRec:backward(identity, real2):mul(lambda2):mul(opt.lambda_identity) + netG:backward(real2, didentity_loss_do) + end + + --- GAN loss: D_A(G_A(A)) + local fake = netG:forward(real):clone() + local output = netD:forward(fake) + local errG = self.criterionGAN:forward(output, real_label) + local df_do1 = self.criterionGAN:backward(output, real_label) + local df_d_GAN = netD:updateGradInput(fake, df_do1) -- + + -- forward cycle loss + local rec = netE:forward(fake):clone() + local errRec = self.criterionRec:forward(rec, real) * lambda1 + local df_do2 = self.criterionRec:backward(rec, real):mul(lambda1) + local df_do_rec = netE:updateGradInput(fake, df_do2) + + netG:backward(real, df_d_GAN + df_do_rec) + + -- backward cycle loss + local fake2 = netE:forward(real2)--:clone() + local rec2 = netG:forward(fake2)--:clone() + local errAdapt = self.criterionRec:forward(rec2, real2) * lambda2 + local df_do_coadapt = self.criterionRec:backward(rec2, real2):mul(lambda2) + netG:backward(fake2, df_do_coadapt) + + return gradParams, errG, errRec, errI, fake, rec, identity +end + +function CycleGANModel:fGAx(x, opt) + self.gradparametersG_A, self.errG_A, self.errRec_A, self.errI_A, self.fake_B, self.rec_A, self.identity_B = + self:fGx_basic(x, self.gradparametersG_A, self.netG_A, self.netD_A, self.netG_B, self.real_A, self.real_B, + self.real_label_A, opt.lambda_A, opt.lambda_B, opt) + return self.errG_A, self.gradparametersG_A +end + +function CycleGANModel:fGBx(x, opt) + self.gradparametersG_B, self.errG_B, self.errRec_B, self.errI_B, self.fake_A, self.rec_B, self.identity_A = + self:fGx_basic(x, self.gradparametersG_B, self.netG_B, self.netD_B, self.netG_A, self.real_B, self.real_A, + self.real_label_B, opt.lambda_B, opt.lambda_A, opt) + return self.errG_B, self.gradparametersG_B +end + + +function CycleGANModel:OptimizeParameters(opt) + local fDA = function(x) return self:fDAx(x, opt) end + local fGA = function(x) return self:fGAx(x, opt) end + local fDB = function(x) return self:fDBx(x, opt) end + local fGB = function(x) return self:fGBx(x, opt) end + + optim.adam(fGA, self.parametersG_A, self.optimStateG_A) + optim.adam(fDA, self.parametersD_A, self.optimStateD_A) + optim.adam(fGB, self.parametersG_B, self.optimStateG_B) + optim.adam(fDB, self.parametersD_B, self.optimStateD_B) +end + +function CycleGANModel:RefreshParameters() + self.parametersD_A, self.gradparametersD_A = nil, nil -- nil them to avoid spiking memory + self.parametersG_A, self.gradparametersG_A = nil, nil + self.parametersG_B, self.gradparametersG_B = nil, nil + self.parametersD_B, self.gradparametersD_B = nil, nil + -- define parameters of optimization + self.parametersG_A, self.gradparametersG_A = self.netG_A:getParameters() + self.parametersD_A, self.gradparametersD_A = self.netD_A:getParameters() + self.parametersG_B, self.gradparametersG_B = self.netG_B:getParameters() + self.parametersD_B, self.gradparametersD_B = self.netD_B:getParameters() +end + +function CycleGANModel:Save(prefix, opt) + util.save_model(self.netG_A, prefix .. '_net_G_A.t7', 1) + util.save_model(self.netD_A, prefix .. '_net_D_A.t7', 1) + util.save_model(self.netG_B, prefix .. '_net_G_B.t7', 1) + util.save_model(self.netD_B, prefix .. '_net_D_B.t7', 1) +end + +function CycleGANModel:GetCurrentErrorDescription() + description = ('[A] G: %.4f D: %.4f Rec: %.4f I: %.4f || [B] G: %.4f D: %.4f Rec: %.4f I:%.4f'):format( + self.errG_A and self.errG_A or -1, + self.errD_A and self.errD_A or -1, + self.errRec_A and self.errRec_A or -1, + self.errI_A and self.errI_A or -1, + self.errG_B and self.errG_B or -1, + self.errD_B and self.errD_B or -1, + self.errRec_B and self.errRec_B or -1, + self.errI_B and self.errI_B or -1) + return description +end + +function CycleGANModel:GetCurrentErrors() + local errors = {errG_A=self.errG_A, errD_A=self.errD_A, errRec_A=self.errRec_A, errI_A=self.errI_A, + errG_B=self.errG_B, errD_B=self.errD_B, errRec_B=self.errRec_B, errI_B=self.errI_B} + return errors +end + +-- returns a string that describes the display plot configuration +function CycleGANModel:DisplayPlot(opt) + if opt.lambda_identity > 0 then + return 'errG_A,errD_A,errRec_A,errI_A,errG_B,errD_B,errRec_B,errI_B' + else + return 'errG_A,errD_A,errRec_A,errG_B,errD_B,errRec_B' + end +end + +function CycleGANModel:UpdateLearningRate(opt) + local lrd = opt.lr / opt.niter_decay + local old_lr = self.optimStateD_A['learningRate'] + local lr = old_lr - lrd + self.optimStateD_A['learningRate'] = lr + self.optimStateD_B['learningRate'] = lr + self.optimStateG_A['learningRate'] = lr + self.optimStateG_B['learningRate'] = lr + print(('update learning rate: %f -> %f'):format(old_lr, lr)) +end + +local function MakeIm3(im) + if im:size(2) == 1 then + local im3 = torch.repeatTensor(im, 1,3,1,1) + return im3 + else + return im + end +end + +function CycleGANModel:GetCurrentVisuals(opt, size) + local visuals = {} + table.insert(visuals, {img=MakeIm3(self.real_A), label='real_A'}) + table.insert(visuals, {img=MakeIm3(self.fake_B), label='fake_B'}) + table.insert(visuals, {img=MakeIm3(self.rec_A), label='rec_A'}) + if opt.test == 0 and opt.lambda_identity > 0 then + table.insert(visuals, {img=MakeIm3(self.identity_A), label='identity_A'}) + end + table.insert(visuals, {img=MakeIm3(self.real_B), label='real_B'}) + table.insert(visuals, {img=MakeIm3(self.fake_A), label='fake_A'}) + table.insert(visuals, {img=MakeIm3(self.rec_B), label='rec_B'}) + if opt.test == 0 and opt.lambda_identity > 0 then + table.insert(visuals, {img=MakeIm3(self.identity_B), label='identity_B'}) + end + return visuals +end diff --git a/cyclegan/models/one_direction_test_model.lua b/cyclegan/models/one_direction_test_model.lua new file mode 100644 index 0000000..b06c186 --- /dev/null +++ b/cyclegan/models/one_direction_test_model.lua @@ -0,0 +1,79 @@ +local class = require 'class' +require 'models.base_model' +require 'models.architectures' +require 'util.image_pool' + +util = paths.dofile('../util/util.lua') +OneDirectionTestModel = class('OneDirectionTestModel', 'BaseModel') + +function OneDirectionTestModel:__init(conf) + BaseModel.__init(self, conf) + conf = conf or {} +end + +function OneDirectionTestModel:model_name() + return 'OneDirectionTestModel' +end + +-- Defines models and networks +function OneDirectionTestModel:Initialize(opt) + -- define tensors + self.real_A = torch.Tensor(opt.batchSize, opt.input_nc, opt.fineSize, opt.fineSize) + + -- load/define models + self.netG_A = util.load_test_model('G', opt) + + -- setup optnet to save a bit of memory + if opt.use_optnet == 1 then + local optnet = require 'optnet' + local sample_input = torch.randn(1, opt.input_nc, 2, 2) + optnet.optimizeMemory(self.netG_A, sample_input, {inplace=true, reuseBuffers=true}) + end + + self:RefreshParameters() + + print('---------- # Learnable Parameters --------------') + print(('G_A = %d'):format(self.parametersG_A:size(1))) + print('------------------------------------------------') +end + +-- Runs the forward pass of the network and +-- saves the result to member variables of the class +function OneDirectionTestModel:Forward(input, opt) + if opt.which_direction == 'BtoA' then + input.real_A = input.real_B:clone() + end + + self.real_A = input.real_A:clone() + if opt.gpu > 0 then + self.real_A = self.real_A:cuda() + end + + self.fake_B = self.netG_A:forward(self.real_A):clone() +end + +function OneDirectionTestModel:RefreshParameters() + self.parametersG_A, self.gradparametersG_A = nil, nil + self.parametersG_A, self.gradparametersG_A = self.netG_A:getParameters() +end + + +local function MakeIm3(im) + if im:size(2) == 1 then + local im3 = torch.repeatTensor(im, 1,3,1,1) + return im3 + else + return im + end +end + +function OneDirectionTestModel:GetCurrentVisuals(opt, size) + if not size then + size = opt.display_winsize + end + + local visuals = {} + table.insert(visuals, {img=MakeIm3(self.real_A), label='real_A'}) + table.insert(visuals, {img=MakeIm3(self.fake_B), label='fake_B'}) + return visuals +end diff --git a/cyclegan/models/pix2pix_model.lua b/cyclegan/models/pix2pix_model.lua new file mode 100644 index 0000000..0629937 --- /dev/null +++ b/cyclegan/models/pix2pix_model.lua @@ -0,0 +1,228 @@ +local class = require 'class' +require 'models.base_model' +require 'models.architectures' +require 'util.image_pool' +util = paths.dofile('../util/util.lua') +Pix2PixModel = class('Pix2PixModel', 'BaseModel') + +function Pix2PixModel:__init(conf) + conf = conf or {} +end + +-- Returns the name of the model +function Pix2PixModel:model_name() + return 'Pix2PixModel' +end + +function Pix2PixModel:InitializeStates() + return {learningRate=opt.lr, beta1=opt.beta1,} +end + +-- Defines models and networks +function Pix2PixModel:Initialize(opt) -- use lsgan + -- define tensors + local d_input_nc = opt.input_nc + opt.output_nc + self.real_AB = torch.Tensor(opt.batchSize, d_input_nc, opt.fineSize, opt.fineSize) + self.fake_AB = torch.Tensor(opt.batchSize, d_input_nc, opt.fineSize, opt.fineSize) + if opt.test == 0 then + self.fakeABPool = ImagePool(opt.pool_size) + end + -- load/define models + self.criterionGAN = nn.MSECriterion() + self.criterionL1 = nn.AbsCriterion() + + local netG, netD = nil, nil + if opt.continue_train == 1 then + if opt.test == 1 then -- only load model G for test + netG = util.load_test_model('G', opt) + else + netG = util.load_model('G', opt) + netD = util.load_model('D', opt) + end + else + netG = defineG(opt.input_nc, opt.output_nc, opt.ngf, opt.which_model_netG) + netD = defineD(d_input_nc, opt.ndf, opt.which_model_netD, opt.n_layers_D, false) -- with sigmoid + end + + self.netD = netD + self.netG = netG + + -- define real/fake labels + if opt.test == 0 then + netD_output_size = self.netD:forward(self.real_AB):size() + self.fake_label = torch.Tensor(netD_output_size):fill(0.0) + self.real_label = torch.Tensor(netD_output_size):fill(1.0) -- no soft smoothing + + self.optimStateD = self:InitializeStates() + self.optimStateG = self:InitializeStates() + + self:RefreshParameters() + + print('---------- # Learnable Parameters --------------') + print(('G = %d'):format(self.parametersG:size(1))) + print(('D = %d'):format(self.parametersD:size(1))) + print('------------------------------------------------') + end + + self.A_idx = {{}, {1, opt.input_nc}, {}, {}} + self.B_idx = {{}, {opt.input_nc+1, opt.input_nc+opt.output_nc}, {}, {}} +end + +-- Runs the forward pass of the network +function Pix2PixModel:Forward(input, opt) + if opt.which_direction == 'BtoA' then + local temp = input.real_A + input.real_A = input.real_B + input.real_B = temp + end + + if opt.test == 0 then + self.real_AB[self.A_idx]:copy(input.real_A) + self.real_AB[self.B_idx]:copy(input.real_B) + self.real_A = self.real_AB[self.A_idx] + self.real_B = self.real_AB[self.B_idx] + + self.fake_AB[self.A_idx]:copy(self.real_A) + self.fake_B = self.netG:forward(self.real_A):clone() + self.fake_AB[self.B_idx]:copy(self.fake_B) + else + if opt.gpu > 0 then + self.real_A = input.real_A:cuda() + self.real_B = input.real_B:cuda() + else + self.real_A = input.real_A:clone() + self.real_B = input.real_B:clone() + end + self.fake_B = self.netG:forward(self.real_A):clone() + end +end + +-- create closure to evaluate f(X) and df/dX of discriminator +function Pix2PixModel:fDx_basic(x, gradParams, netD, netG, real, fake, opt) + util.BiasZero(netD) + util.BiasZero(netG) + gradParams:zero() + + -- Real log(D(B)) + local output = netD:forward(real) + local errD_real = self.criterionGAN:forward(output, self.real_label) + local df_do = self.criterionGAN:backward(output, self.real_label) + netD:backward(real, df_do) + -- Fake + log(1 - D(G(A))) + output = netD:forward(fake) + local errD_fake = self.criterionGAN:forward(output, self.fake_label) + local df_do2 = self.criterionGAN:backward(output, self.fake_label) + netD:backward(fake, df_do2) + -- calculate loss + local errD = (errD_real + errD_fake) / 2.0 + return errD, gradParams +end + + +function Pix2PixModel:fDx(x, opt) + fake_AB = self.fakeABPool:Query(self.fake_AB) + self.errD, gradParams = self:fDx_basic(x, self.gradParametersD, self.netD, self.netG, + self.real_AB, fake_AB, opt) + return self.errD, gradParams +end + +function Pix2PixModel:fGx_basic(x, netG, netD, real, fake, gradParametersG, opt) + util.BiasZero(netG) + util.BiasZero(netD) + gradParametersG:zero() + + -- First. G(A) should fake the discriminator + local output = netD:forward(fake) + local errG = self.criterionGAN:forward(output, self.real_label) + local dgan_loss_dd = self.criterionGAN:backward(output, self.real_label) + local dgan_loss_do = netD:updateGradInput(fake, dgan_loss_dd) + + -- Second. G(A) should be close to the real + real_B = real[self.B_idx] + real_A = real[self.A_idx] + fake_B = fake[self.B_idx] + local errL1 = self.criterionL1:forward(fake_B, real_B) * opt.lambda_A + local dl1_loss_do = self.criterionL1:backward(fake_B, real_B) * opt.lambda_A + netG:backward(real_A, dgan_loss_do[self.B_idx] + dl1_loss_do) + + return gradParametersG, errG, errL1 +end + +function Pix2PixModel:fGx(x, opt) + self.gradParametersG, self.errG, self.errL1 = self:fGx_basic(x, self.netG, self.netD, + self.real_AB, self.fake_AB, self.gradParametersG, opt) + return self.errG, self.gradParametersG +end + +-- Runs the backprop gradient descent +-- Corresponds to a single batch of data +function Pix2PixModel:OptimizeParameters(opt) + local fD = function(x) return self:fDx(x, opt) end + local fG = function(x) return self:fGx(x, opt) end + optim.adam(fD, self.parametersD, self.optimStateD) + optim.adam(fG, self.parametersG, self.optimStateG) +end + +-- This function can be used to reset momentum after each epoch +function Pix2PixModel:RefreshParameters() + self.parametersD, self.gradParametersD = nil, nil -- nil them to avoid spiking memory + self.parametersG, self.gradParametersG = nil, nil + + -- define parameters of optimization + self.parametersG, self.gradParametersG = self.netG:getParameters() + self.parametersD, self.gradParametersD = self.netD:getParameters() +end + +-- This function updates the learning rate; lr for the first opt.niter iterations; graduatlly decreases the lr to 0 for the next opt.niter_decay iterations +function Pix2PixModel:UpdateLearningRate(opt) + local lrd = opt.lr / opt.niter_decay + local old_lr = self.optimStateD['learningRate'] + local lr = old_lr - lrd + self.optimStateD['learningRate'] = lr + self.optimStateG['learningRate'] = lr + print(('update learning rate: %f -> %f'):format(old_lr, lr)) +end + + +-- Save the current model to the file system +function Pix2PixModel:Save(prefix, opt) + util.save_model(self.netG, prefix .. '_net_G.t7', 1.0) + util.save_model(self.netD, prefix .. '_net_D.t7', 1.0) +end + +-- returns a string that describes the current errors +function Pix2PixModel:GetCurrentErrorDescription() + description = ('G: %.4f D: %.4f L1: %.4f'):format( + self.errG and self.errG or -1, self.errD and self.errD or -1, self.errL1 and self.errL1 or -1) + return description + +end + + +-- returns a string that describes the display plot configuration +function Pix2PixModel:DisplayPlot(opt) + return 'errG,errD,errL1' +end + + +-- returns current errors +function Pix2PixModel:GetCurrentErrors() + local errors = {errG=self.errG, errD=self.errD, errL1=self.errL1} + return errors +end + +-- returns a table of image/label pairs that describe +-- the current results. +-- |return|: a table of table. List of image/label pairs +function Pix2PixModel:GetCurrentVisuals(opt, size) + if not size then + size = opt.display_winsize + end + + local visuals = {} + table.insert(visuals, {img=self.real_A, label='real_A'}) + table.insert(visuals, {img=self.fake_B, label='fake_B'}) + table.insert(visuals, {img=self.real_B, label='real_B'}) + + return visuals +end diff --git a/cyclegan/options.lua b/cyclegan/options.lua new file mode 100644 index 0000000..2e4e6ca --- /dev/null +++ b/cyclegan/options.lua @@ -0,0 +1,143 @@ +-------------------------------------------------------------------------------- +-- Configure options +-------------------------------------------------------------------------------- + +local options = {} +-- options for train +local opt_train = { + DATA_ROOT = '', -- path to images (should have subfolders 'train', 'val', etc) + batchSize = 1, -- # images in batch + loadSize = 143, -- scale images to this size + fineSize = 128, -- then crop to this size + ngf = 64, -- # of gen filters in first conv layer + ndf = 64, -- # of discrim filters in first conv layer + input_nc = 3, -- # of input image channels + output_nc = 3, -- # of output image channels + niter = 100, -- # of iter at starting learning rate + niter_decay = 100, -- # of iter to linearly decay learning rate to zero + lr = 0.0002, -- initial learning rate for adam + beta1 = 0.5, -- momentum term of adam + ntrain = math.huge, -- # of examples per epoch. math.huge for full dataset + flip = 1, -- if flip the images for data argumentation + display_id = 10, -- display window id. + display_winsize = 128, -- display window size + display_freq = 25, -- display the current results every display_freq iterations + gpu = 1, -- gpu = 0 is CPU mode. gpu=X is GPU mode on GPU X + name = '', -- name of the experiment, should generally be passed on the command line + which_direction = 'AtoB', -- AtoB or BtoA + phase = 'train', -- train, val, test, etc + nThreads = 2, -- # threads for loading data + save_epoch_freq = 1, -- save a model every save_epoch_freq epochs (does not overwrite previously saved models) + save_latest_freq = 5000, -- save the latest model every latest_freq sgd iterations (overwrites the previous latest model) + print_freq = 50, -- print the debug information every print_freq iterations + save_display_freq = 2500, -- save the current display of results every save_display_freq_iterations + continue_train = 0, -- if continue training, load the latest model: 1: true, 0: false + serial_batches = 0, -- if 1, takes images in order to make batches, otherwise takes them randomly + checkpoints_dir = './checkpoints', -- models are saved here + cache_dir = './cache', -- cache files are saved here + cudnn = 1, -- set to 0 to not use cudnn + which_model_netD = 'basic', -- selects model to use for netD + which_model_netG = 'resnet_6blocks', -- selects model to use for netG + norm = 'instance', -- batch or instance normalization + n_layers_D = 3, -- only used if which_model_netD=='n_layers' + content_loss = 'pixel', -- content loss type: pixel, vgg + layer_name = 'pixel', -- layer used in content loss (e.g. relu4_2) + lambda_A = 10.0, -- weight for cycle loss (A -> B -> A) + lambda_B = 10.0, -- weight for cycle loss (B -> A -> B) + model = 'cycle_gan', -- which mode to run. 'cycle_gan', 'pix2pix', 'bigan', 'content_gan' + use_lsgan = 1, -- if 1, use least square GAN, if 0, use vanilla GAN + align_data = 0, -- if > 0, use the dataloader for where the images are aligned + pool_size = 50, -- the size of image buffer that stores previously generated images + resize_or_crop = 'resize_and_crop', -- resizing/cropping strategy: resize_and_crop | crop | scale_width | scale_height + lambda_identity = 0.5, -- use identity mapping. Setting opt.lambda_identity other than 0 has an effect of scaling the weight of the identity mapping loss. For example, if the weight of the identity loss should be 10 times smaller than the weight of the reconstruction loss, please set opt.lambda_identity = 0.1 + use_optnet = 0, -- use optnet to save GPU memory during test +} + +-- options for test +local opt_test = { + DATA_ROOT = '', -- path to images (should have subfolders 'train', 'val', etc) + loadSize = 128, -- scale images to this size + fineSize = 128, -- then crop to this size + flip = 0, -- horizontal mirroring data augmentation + display = 1, -- display samples while training. 0 = false + display_id = 200, -- display window id. + gpu = 1, -- gpu = 0 is CPU mode. gpu=X is GPU mode on GPU X + how_many = 'all', -- how many test images to run (set to all to run on every image found in the data/phase folder) + phase = 'test', -- train, val, test, etc + aspect_ratio = 1.0, -- aspect ratio of result images + norm = 'instance', -- batchnorm or isntance norm + name = '', -- name of experiment, selects which model to run, should generally should be passed on command line + input_nc = 3, -- # of input image channels + output_nc = 3, -- # of output image channels + serial_batches = 1, -- if 1, takes images in order to make batches, otherwise takes them randomly + cudnn = 1, -- set to 0 to not use cudnn (untested) + checkpoints_dir = './checkpoints', -- loads models from here + cache_dir = './cache', -- cache files are saved here + results_dir='./results/', -- saves results here + which_epoch = 'latest', -- which epoch to test? set to 'latest' to use latest cached model + model = 'cycle_gan', -- which mode to run. 'cycle_gan', 'pix2pix', 'bigan', 'content_gan'; to use pretrained model, select `one_direction_test` + align_data = 0, -- if > 0, use the dataloader for pix2pix + which_direction = 'AtoB', -- AtoB or BtoA + resize_or_crop = 'resize_and_crop', -- resizing/cropping strategy: resize_and_crop | crop | scale_width | scale_height +} + +-------------------------------------------------------------------------------- +-- util functions +-------------------------------------------------------------------------------- +function options.clone(opt) + local copy = {} + for orig_key, orig_value in pairs(opt) do + copy[orig_key] = orig_value + end + return copy +end + +function options.parse_options(mode) + if mode == 'train' then + opt = opt_train + opt.test = 0 + elseif mode == 'test' then + opt = opt_test + opt.test = 1 + else + print("Invalid option [" .. mode .. "]") + return nil + end + + -- one-line argument parser. parses enviroment variables to override the defaults + for k,v in pairs(opt) do opt[k] = tonumber(os.getenv(k)) or os.getenv(k) or opt[k] end + if mode == 'test' then + opt.nThreads = 1 + opt.continue_train = 1 + opt.batchSize = 1 -- test code only supports batchSize=1 + end + + -- print by keys + keyset = {} + for k,v in pairs(opt) do + table.insert(keyset, k) + end + table.sort(keyset) + print("------------------- Options -------------------") + for i,k in ipairs(keyset) do + print(('%+25s: %s'):format(k, opt[k])) + end + print("-----------------------------------------------") + + -- save opt to checkpoints + paths.mkdir(opt.checkpoints_dir) + paths.mkdir(paths.concat(opt.checkpoints_dir, opt.name)) + opt.visual_dir = paths.concat(opt.checkpoints_dir, opt.name, 'visuals') + paths.mkdir(opt.visual_dir) + -- save opt to the disk + fd = io.open(paths.concat(opt.checkpoints_dir, opt.name, 'opt_' .. mode .. '.txt'), 'w') + for i,k in ipairs(keyset) do + fd:write(("%+25s: %s\n"):format(k, opt[k])) + end + fd:close() + + return opt +end + + +return options diff --git a/cyclegan/pretrained_models/download_model.sh b/cyclegan/pretrained_models/download_model.sh new file mode 100644 index 0000000..7218e4d --- /dev/null +++ b/cyclegan/pretrained_models/download_model.sh @@ -0,0 +1,10 @@ +FILE=$1 + +echo "Note: available models are apple2orange, facades_photo2label, map2sat, orange2apple, style_cezanne, style_ukiyoe, summer2winter_yosemite, zebra2horse, facades_label2photo, horse2zebra,monet2photo, sat2map, style_monet,style_vangogh, winter2summer_yosemite, iphone2dslr_flower" + +echo "Specified [$FILE]" + +mkdir -p ./checkpoints/${FILE}_pretrained +URL=https://people.eecs.berkeley.edu/~taesung_park/CycleGAN/models/$FILE.t7 +MODEL_FILE=./checkpoints/${FILE}_pretrained/latest_net_G.t7 +wget -N $URL -O $MODEL_FILE diff --git a/cyclegan/pretrained_models/download_vgg.sh b/cyclegan/pretrained_models/download_vgg.sh new file mode 100644 index 0000000..2f89982 --- /dev/null +++ b/cyclegan/pretrained_models/download_vgg.sh @@ -0,0 +1,6 @@ +URL1=https://people.eecs.berkeley.edu/~taesung_park/projects/CycleGAN/models/places_vgg.caffemodel +MODEL_FILE1=./models/places_vgg.caffemodel +URL2=https://people.eecs.berkeley.edu/~taesung_park/projects/CycleGAN/models/places_vgg.prototxt +MODEL_FILE2=./models/places_vgg.prototxt +wget -N $URL1 -O $MODEL_FILE1 +wget -N $URL2 -O $MODEL_FILE2 diff --git a/cyclegan/pretrained_models/places_vgg.prototxt b/cyclegan/pretrained_models/places_vgg.prototxt new file mode 100644 index 0000000..f47002b --- /dev/null +++ b/cyclegan/pretrained_models/places_vgg.prototxt @@ -0,0 +1,593 @@ +name: "VGG-Places365" +input: "data" +input_dim: 1 +input_dim: 3 +input_dim: 224 +input_dim: 224 +layer { + name: "conv1_1" + type: "Convolution" + bottom: "data" + top: "conv1_1" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 64 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu1_1" + type: "ReLU" + bottom: "conv1_1" + top: "conv1_1" +} +layer { + name: "conv1_2" + type: "Convolution" + bottom: "conv1_1" + top: "conv1_2" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 64 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu1_2" + type: "ReLU" + bottom: "conv1_2" + top: "conv1_2" +} +layer { + name: "pool1" + type: "Pooling" + bottom: "conv1_2" + top: "pool1" + pooling_param { + pool: MAX + kernel_size: 2 + stride: 2 + } +} +layer { + name: "conv2_1" + type: "Convolution" + bottom: "pool1" + top: "conv2_1" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 128 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu2_1" + type: "ReLU" + bottom: "conv2_1" + top: "conv2_1" +} +layer { + name: "conv2_2" + type: "Convolution" + bottom: "conv2_1" + top: "conv2_2" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 128 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu2_2" + type: "ReLU" + bottom: "conv2_2" + top: "conv2_2" +} +layer { + name: "pool2" + type: "Pooling" + bottom: "conv2_2" + top: "pool2" + pooling_param { + pool: MAX + kernel_size: 2 + stride: 2 + } +} +layer { + name: "conv3_1" + type: "Convolution" + bottom: "pool2" + top: "conv3_1" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 256 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu3_1" + type: "ReLU" + bottom: "conv3_1" + top: "conv3_1" +} +layer { + name: "conv3_2" + type: "Convolution" + bottom: "conv3_1" + top: "conv3_2" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 256 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu3_2" + type: "ReLU" + bottom: "conv3_2" + top: "conv3_2" +} +layer { + name: "conv3_3" + type: "Convolution" + bottom: "conv3_2" + top: "conv3_3" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 256 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu3_3" + type: "ReLU" + bottom: "conv3_3" + top: "conv3_3" +} +layer { + name: "pool3" + type: "Pooling" + bottom: "conv3_3" + top: "pool3" + pooling_param { + pool: MAX + kernel_size: 2 + stride: 2 + } +} +layer { + name: "conv4_1" + type: "Convolution" + bottom: "pool3" + top: "conv4_1" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 512 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu4_1" + type: "ReLU" + bottom: "conv4_1" + top: "conv4_1" +} +layer { + name: "conv4_2" + type: "Convolution" + bottom: "conv4_1" + top: "conv4_2" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 512 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu4_2" + type: "ReLU" + bottom: "conv4_2" + top: "conv4_2" +} +layer { + name: "conv4_3" + type: "Convolution" + bottom: "conv4_2" + top: "conv4_3" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 512 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu4_3" + type: "ReLU" + bottom: "conv4_3" + top: "conv4_3" +} +layer { + name: "pool4" + type: "Pooling" + bottom: "conv4_3" + top: "pool4" + pooling_param { + pool: MAX + kernel_size: 2 + stride: 2 + } +} +layer { + name: "conv5_1" + type: "Convolution" + bottom: "pool4" + top: "conv5_1" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 512 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu5_1" + type: "ReLU" + bottom: "conv5_1" + top: "conv5_1" +} +layer { + name: "conv5_2" + type: "Convolution" + bottom: "conv5_1" + top: "conv5_2" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 512 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu5_2" + type: "ReLU" + bottom: "conv5_2" + top: "conv5_2" +} +layer { + name: "conv5_3" + type: "Convolution" + bottom: "conv5_2" + top: "conv5_3" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + convolution_param { + num_output: 512 + pad: 1 + kernel_size: 3 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu5_3" + type: "ReLU" + bottom: "conv5_3" + top: "conv5_3" +} +layer { + name: "pool5" + type: "Pooling" + bottom: "conv5_3" + top: "pool5" + pooling_param { + pool: MAX + kernel_size: 2 + stride: 2 + } +} +layer { + name: "fc6" + type: "InnerProduct" + bottom: "pool5" + top: "fc6" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + inner_product_param { + num_output: 4096 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu6" + type: "ReLU" + bottom: "fc6" + top: "fc6" +} +layer { + name: "drop6" + type: "Dropout" + bottom: "fc6" + top: "fc6" + dropout_param { + dropout_ratio: 0.5 + } +} +layer { + name: "fc7" + type: "InnerProduct" + bottom: "fc6" + top: "fc7" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + inner_product_param { + num_output: 4096 + weight_filler { + type: "gaussian" + std: 0.01 + } + bias_filler { + type: "constant" + value: 0.0 + } + } +} +layer { + name: "relu7" + type: "ReLU" + bottom: "fc7" + top: "fc7" +} +layer { + name: "drop7" + type: "Dropout" + bottom: "fc7" + top: "fc7" + dropout_param { + dropout_ratio: 0.5 + } +} +layer { + name: "fc8a" + type: "InnerProduct" + bottom: "fc7" + top: "fc8a" + param { + lr_mult: 1.0 + decay_mult: 1.0 + } + param { + lr_mult: 2.0 + decay_mult: 0.0 + } + inner_product_param { + num_output: 365 + } +} +layer { + name: "prob" + type: "Softmax" + bottom: "fc8a" + top: "prob" +} diff --git a/cyclegan/test.lua b/cyclegan/test.lua new file mode 100644 index 0000000..8872990 --- /dev/null +++ b/cyclegan/test.lua @@ -0,0 +1,142 @@ +-- usage: DATA_ROOT=/path/to/data/ name=expt1 which_direction=BtoA th test.lua +-- +-- code derived from https://github.com/soumith/dcgan.torch and https://github.com/phillipi/pix2pix +require 'image' +require 'nn' +require 'nngraph' +require 'models.architectures' + + +util = paths.dofile('util/util.lua') +options = require 'options' +opt = options.parse_options('test') + +-- initialize torch GPU/CPU mode +if opt.gpu > 0 then + require 'cutorch' + require 'cunn' + cutorch.setDevice(opt.gpu) + print ("GPU Mode") + torch.setdefaulttensortype('torch.CudaTensor') +else + torch.setdefaulttensortype('torch.FloatTensor') + print ("CPU Mode") +end + +-- setup visualization +visualizer = require 'util/visualizer' + +function TableConcat(t1,t2) + for i=1,#t2 do + t1[#t1+1] = t2[i] + end + return t1 +end + + +-- load data +local data_loader = nil +if opt.align_data > 0 then + require 'data.aligned_data_loader' + data_loader = AlignedDataLoader() +else + require 'data.unaligned_data_loader' + data_loader = UnalignedDataLoader() +end +print( "DataLoader " .. data_loader:name() .. " was created.") +data_loader:Initialize(opt) + +if opt.how_many == 'all' then + opt.how_many = data_loader:size() +end + +opt.how_many = math.min(opt.how_many, data_loader:size()) + +-- set batch/instance normalization +set_normalization(opt.norm) + +-- load model +opt.continue_train = 1 +-- define model +if opt.model == 'cycle_gan' then + require 'models.cycle_gan_model' + model = CycleGANModel() +elseif opt.model == 'one_direction_test' then + require 'models.one_direction_test_model' + model = OneDirectionTestModel() +elseif opt.model == 'pix2pix' then + require 'models.pix2pix_model' + model = Pix2PixModel() +elseif opt.model == 'bigan' then + require 'models.bigan_model' + model = BiGANModel() +elseif opt.model == 'content_gan' then + require 'models.content_gan_model' + model = ContentGANModel() +else + error('Please specify a correct model') +end +model:Initialize(opt) + +local pathsA = {} -- paths to images A tested on +local pathsB = {} -- paths to images B tested on +local web_dir = paths.concat(opt.results_dir, opt.name .. '/' .. opt.which_epoch .. '_' .. opt.phase) +paths.mkdir(web_dir) +local image_dir = paths.concat(web_dir, 'images') +paths.mkdir(image_dir) +s1 = opt.fineSize +s2 = opt.fineSize / opt.aspect_ratio + +visuals = {} + +for n = 1, math.floor(opt.how_many) do + print('processing batch ' .. n) + local cur_dataA, cur_dataB, cur_pathsA, cur_pathsB = data_loader:GetNextBatch() + + cur_pathsA = util.basename_batch(cur_pathsA) + cur_pathsB = util.basename_batch(cur_pathsB) + print('pathsA', cur_pathsA) + print('pathsB', cur_PathsB) + model:Forward({real_A=cur_dataA, real_B=cur_dataB}, opt) + + visuals = model:GetCurrentVisuals(opt, opt.fineSize) + + for i,visual in ipairs(visuals) do + if opt.resize_or_crop == 'scale_width' or opt.resize_or_crop == 'scale_height' then + s1 = nil + s2 = nil + end + visualizer.save_images(visual.img, paths.concat(image_dir, visual.label), {string.gsub(cur_pathsA[1],'.jpg','.png')}, s1, s2) + end + + + print('Saved images to: ', image_dir) + pathsA = TableConcat(pathsA, cur_pathsA) + pathsB = TableConcat(pathsB, cur_pathsB) +end + +labels = {} +for i,visual in ipairs(visuals) do + table.insert(labels, visual.label) +end + +-- make webpage +io.output(paths.concat(web_dir, 'index.html')) +io.write('') +io.write('') +for i = 1, #labels do + io.write('') +end +io.write('') + +for n = 1,math.floor(opt.how_many) do + io.write('') + io.write('') + for j = 1, #labels do + label = labels[j] + io.write('') + end + io.write('') +end + +io.write('
Image ' .. labels[i] .. '
' .. tostring(n) .. '
') diff --git a/cyclegan/train.lua b/cyclegan/train.lua new file mode 100644 index 0000000..6ef9974 --- /dev/null +++ b/cyclegan/train.lua @@ -0,0 +1,178 @@ +-- usage example: DATA_ROOT=/path/to/data/ which_direction=BtoA name=expt1 th train.lua +-- code derived from https://github.com/soumith/dcgan.torch and https://github.com/phillipi/pix2pix + +require 'torch' +require 'nn' +require 'optim' +util = paths.dofile('util/util.lua') +content = paths.dofile('util/content_loss.lua') +require 'image' +require 'models.architectures' + +-- load configuration file +options = require 'options' +opt = options.parse_options('train') + +-- setup visualization +visualizer = require 'util/visualizer' + +-- initialize torch GPU/CPU mode +if opt.gpu > 0 then + require 'cutorch' + require 'cunn' + cutorch.setDevice(opt.gpu) + print ("GPU Mode") + torch.setdefaulttensortype('torch.CudaTensor') +else + torch.setdefaulttensortype('torch.FloatTensor') + print ("CPU Mode") +end + +-- load data +local data_loader = nil +if opt.align_data > 0 then + require 'data.aligned_data_loader' + data_loader = AlignedDataLoader() +else + require 'data.unaligned_data_loader' + data_loader = UnalignedDataLoader() +end +print( "DataLoader " .. data_loader:name() .. " was created.") +data_loader:Initialize(opt) + +-- set batch/instance normalization +set_normalization(opt.norm) + +--- timer +local epoch_tm = torch.Timer() +local tm = torch.Timer() + +-- define model +local model = nil +local display_plot = nil +if opt.model == 'cycle_gan' then + assert(data_loader:name() == 'UnalignedDataLoader') + require 'models.cycle_gan_model' + model = CycleGANModel() +elseif opt.model == 'pix2pix' then + require 'models.pix2pix_model' + assert(data_loader:name() == 'AlignedDataLoader') + model = Pix2PixModel() +elseif opt.model == 'bigan' then + assert(data_loader:name() == 'UnalignedDataLoader') + require 'models.bigan_model' + model = BiGANModel() +elseif opt.model == 'content_gan' then + require 'models.content_gan_model' + assert(data_loader:name() == 'UnalignedDataLoader') + model = ContentGANModel() +else + error('Please specify a correct model') +end + +-- print the model name +print('Model ' .. model:model_name() .. ' was specified.') +model:Initialize(opt) + +-- set up the loss plot +require 'util/plot_util' +plotUtil = PlotUtil() +display_plot = model:DisplayPlot(opt) +plotUtil:Initialize(display_plot, opt.display_id, opt.name) + +-------------------------------------------------------------------------------- +-- Helper Functions +-------------------------------------------------------------------------------- +function visualize_current_results() + local visuals = model:GetCurrentVisuals(opt) + for i,visual in ipairs(visuals) do + visualizer.disp_image(visual.img, opt.display_winsize, + opt.display_id+i, opt.name .. ' ' .. visual.label) + end +end + +function save_current_results(epoch, counter) + local visuals = model:GetCurrentVisuals(opt) + for i,visual in ipairs(visuals) do + output_path = paths.concat(opt.visual_dir, 'train_epoch' .. epoch .. '_iter' .. counter .. '_' .. visual.label .. '.jpg') + visualizer.save_results(visual.img, output_path) + end +end + +function print_current_errors(epoch, counter_in_epoch) + print(('Epoch: [%d][%8d / %8d]\t Time: %.3f DataTime: %.3f ' + .. '%s'): + format(epoch, ((counter_in_epoch-1) / opt.batchSize), + math.floor(math.min(data_loader:size(), opt.ntrain) / opt.batchSize), + tm:time().real / opt.batchSize, + data_loader:time_elapsed_to_fetch_data() / opt.batchSize, + model:GetCurrentErrorDescription() + )) +end + +function plot_current_errors(epoch, counter_ratio, opt) + local errs = model:GetCurrentErrors(opt) + local plot_vals = { epoch + counter_ratio} + plotUtil:Display(plot_vals, errs) +end + +-------------------------------------------------------------------------------- +-- Main Training Loop +-------------------------------------------------------------------------------- +local counter = 0 +local num_batches = math.floor(math.min(data_loader:size(), opt.ntrain) / opt.batchSize) +print('#training iterations: ' .. opt.niter+opt.niter_decay ) + +for epoch = 1, opt.niter+opt.niter_decay do + epoch_tm:reset() + for counter_in_epoch = 1, math.min(data_loader:size(), opt.ntrain), opt.batchSize do + tm:reset() + -- load a batch and run G on that batch + local real_dataA, real_dataB, _, _ = data_loader:GetNextBatch() + + model:Forward({real_A=real_dataA, real_B=real_dataB}, opt) + -- run forward pass + opt.counter = counter + -- run backward pass + model:OptimizeParameters(opt) + -- display on the web server + if counter % opt.display_freq == 0 and opt.display_id > 0 then + visualize_current_results() + end + + -- logging + if counter % opt.print_freq == 0 then + print_current_errors(epoch, counter_in_epoch) + plot_current_errors(epoch, counter_in_epoch/num_batches, opt) + end + + -- save latest model + if counter % opt.save_latest_freq == 0 and counter > 0 then + print(('saving the latest model (epoch %d, iters %d)'):format(epoch, counter)) + model:Save('latest', opt) + end + + -- save latest results + if counter % opt.save_display_freq == 0 then + save_current_results(epoch, counter) + end + counter = counter + 1 + end + + -- save model at the end of epoch + if epoch % opt.save_epoch_freq == 0 then + print(('saving the model (epoch %d, iters %d)'):format(epoch, counter)) + model:Save('latest', opt) + model:Save(epoch, opt) + end + -- print the timing information after each epoch + print(('End of epoch %d / %d \t Time Taken: %.3f'): + format(epoch, opt.niter+opt.niter_decay, epoch_tm:time().real)) + + -- update learning rate + if epoch > opt.niter then + model:UpdateLearningRate(opt) + end + -- refresh parameters + model:RefreshParameters(opt) +end diff --git a/cyclegan/util/InstanceNormalization.lua b/cyclegan/util/InstanceNormalization.lua new file mode 100644 index 0000000..595ddfe --- /dev/null +++ b/cyclegan/util/InstanceNormalization.lua @@ -0,0 +1,106 @@ +require 'nn' + +--[[ + Implements instance normalization as described in the paper + + Instance Normalization: The Missing Ingredient for Fast Stylization + Dmitry Ulyanov, Andrea Vedaldi, Victor Lempitsky + https://arxiv.org/abs/1607.08022 + This implementation is based on + https://github.com/DmitryUlyanov/texture_nets +]] + +local InstanceNormalization, parent = torch.class('nn.InstanceNormalization', 'nn.Module') + +function InstanceNormalization:__init(nOutput, eps, momentum, affine) + parent.__init(self) + self.running_mean = torch.zeros(nOutput) + self.running_var = torch.ones(nOutput) + + self.eps = eps or 1e-5 + self.momentum = momentum or 0.0 + if affine ~= nil then + assert(type(affine) == 'boolean', 'affine has to be true/false') + self.affine = affine + else + self.affine = true + end + + self.nOutput = nOutput + self.prev_batch_size = -1 + + if self.affine then + self.weight = torch.Tensor(nOutput):uniform() + self.bias = torch.Tensor(nOutput):zero() + self.gradWeight = torch.Tensor(nOutput) + self.gradBias = torch.Tensor(nOutput) + end +end + +function InstanceNormalization:updateOutput(input) + self.output = self.output or input.new() + assert(input:size(2) == self.nOutput) + + local batch_size = input:size(1) + + if batch_size ~= self.prev_batch_size or (self.bn and self:type() ~= self.bn:type()) then + self.bn = nn.SpatialBatchNormalization(input:size(1)*input:size(2), self.eps, self.momentum, self.affine) + self.bn:type(self:type()) + self.bn.running_mean:copy(self.running_mean:repeatTensor(batch_size)) + self.bn.running_var:copy(self.running_var:repeatTensor(batch_size)) + + self.prev_batch_size = input:size(1) + end + + -- Get statistics + self.running_mean:copy(self.bn.running_mean:view(input:size(1),self.nOutput):mean(1)) + self.running_var:copy(self.bn.running_var:view(input:size(1),self.nOutput):mean(1)) + + -- Set params for BN + if self.affine then + self.bn.weight:copy(self.weight:repeatTensor(batch_size)) + self.bn.bias:copy(self.bias:repeatTensor(batch_size)) + end + + local input_1obj = input:contiguous():view(1,input:size(1)*input:size(2),input:size(3),input:size(4)) + self.output = self.bn:forward(input_1obj):viewAs(input) + + return self.output +end + +function InstanceNormalization:updateGradInput(input, gradOutput) + self.gradInput = self.gradInput or gradOutput.new() + + assert(self.bn) + + local input_1obj = input:contiguous():view(1,input:size(1)*input:size(2),input:size(3),input:size(4)) + local gradOutput_1obj = gradOutput:contiguous():view(1,input:size(1)*input:size(2),input:size(3),input:size(4)) + + if self.affine then + self.bn.gradWeight:zero() + self.bn.gradBias:zero() + end + + self.gradInput = self.bn:backward(input_1obj, gradOutput_1obj):viewAs(input) + + if self.affine then + self.gradWeight:add(self.bn.gradWeight:view(input:size(1),self.nOutput):sum(1)) + self.gradBias:add(self.bn.gradBias:view(input:size(1),self.nOutput):sum(1)) + end + return self.gradInput +end + +function InstanceNormalization:clearState() + self.output = self.output.new() + self.gradInput = self.gradInput.new() + + if self.bn then + self.bn:clearState() + end +end + +function InstanceNormalization:evaluate() +end + +function InstanceNormalization:training() +end diff --git a/cyclegan/util/VGG_preprocess.lua b/cyclegan/util/VGG_preprocess.lua new file mode 100644 index 0000000..1e74828 --- /dev/null +++ b/cyclegan/util/VGG_preprocess.lua @@ -0,0 +1,27 @@ +-- define nn module for VGG postprocessing +local VGG_postprocess, parent = torch.class('nn.VGG_postprocess', 'nn.Module') + +function VGG_postprocess:__init() + parent.__init(self) +end + +function VGG_postprocess:updateOutput(input) + self.output = input:add(1):mul(127.5) + -- print(self.output:max(), self.output:min()) + if self.output:max() > 255 or self.output:min() < 0 then + print(self.output:min(), self.output:max()) + end + -- assert(self.output:min()>=0,"badly scaled inputs") + -- assert(self.output:max()<=255,"badly scaled inputs") + + local mean_pixel = torch.FloatTensor({103.939, 116.779, 123.68}) + mean_pixel = mean_pixel:reshape(1,3,1,1) + mean_pixel = mean_pixel:repeatTensor(input:size(1), 1, input:size(3), input:size(4)):cuda() + self.output:add(-1, mean_pixel) + return self.output +end + +function VGG_postprocess:updateGradInput(input, gradOutput) + self.gradInput = gradOutput:div(127.5) + return self.gradInput +end diff --git a/cyclegan/util/content_loss.lua b/cyclegan/util/content_loss.lua new file mode 100644 index 0000000..d11d469 --- /dev/null +++ b/cyclegan/util/content_loss.lua @@ -0,0 +1,87 @@ +require 'torch' +require 'nn' +local content = {} + +function content.defineVGG(content_layer) + local contentFunc = nn.Sequential() + require 'loadcaffe' + require 'util/VGG_preprocess' + cnn = loadcaffe.load('../models/vgg.prototxt', '../models/vgg.caffemodel', 'cudnn') + contentFunc:add(nn.SpatialUpSamplingBilinear({oheight=224, owidth=224})) + contentFunc:add(nn.VGG_postprocess()) + for i = 1, #cnn do + local layer = cnn:get(i):clone() + local name = layer.name + local layer_type = torch.type(layer) + contentFunc:add(layer) + if name == content_layer then + print("Setting up content layer: ", layer.name) + break + end + end + cnn = nil + collectgarbage() + print(contentFunc) + return contentFunc +end + +function content.defineAlexNet(content_layer) + local contentFunc = nn.Sequential() + require 'loadcaffe' + require 'util/VGG_preprocess' + cnn = loadcaffe.load('../models/alexnet.prototxt', '../models/alexnet.caffemodel', 'cudnn') + contentFunc:add(nn.SpatialUpSamplingBilinear({oheight=224, owidth=224})) + contentFunc:add(nn.VGG_postprocess()) + for i = 1, #cnn do + local layer = cnn:get(i):clone() + local name = layer.name + local layer_type = torch.type(layer) + contentFunc:add(layer) + if name == content_layer then + print("Setting up content layer: ", layer.name) + break + end + end + cnn = nil + collectgarbage() + print(contentFunc) + return contentFunc +end + + + +function content.defineContent(content_loss, layer_name) + -- print('content_loss_define', content_loss) + if content_loss == 'pixel' or content_loss == 'none' then + return nil + elseif content_loss == 'vgg' then + return content.defineVGG(layer_name) + else + print("unsupported content loss") + return nil + end +end + + +function content.lossUpdate(criterionContent, real_source, fake_target, contentFunc, loss_type, weight) + if loss_type == 'none' then + local errCont = 0.0 + local df_d_content = torch.zeros(fake_target:size()) + return errCont, df_d_content + elseif loss_type == 'pixel' then + local errCont = criterionContent:forward(fake_target, real_source) * weight + local df_do_content = criterionContent:backward(fake_target, real_source)*weight + return errCont, df_do_content + elseif loss_type == 'vgg' then + local f_fake = contentFunc:forward(fake_target):clone() + local f_real = contentFunc:forward(real_source):clone() + local errCont = criterionContent:forward(f_fake, f_real) * weight + local df_do_tmp = criterionContent:backward(f_fake, f_real) * weight + local df_do_content = contentFunc:updateGradInput(fake_target, df_do_tmp)--:mul(weight) + return errCont, df_do_content + else error("unsupported content loss") + end +end + + +return content diff --git a/cyclegan/util/cudnn_convert_custom.lua b/cyclegan/util/cudnn_convert_custom.lua new file mode 100644 index 0000000..8b0c08d --- /dev/null +++ b/cyclegan/util/cudnn_convert_custom.lua @@ -0,0 +1,72 @@ +-- modified from https://github.com/NVIDIA/torch-cudnn/blob/master/convert.lua +-- removed error on nngraph + +-- modules that can be converted to nn seamlessly +local layer_list = { + 'BatchNormalization', + 'SpatialBatchNormalization', + 'SpatialConvolution', + 'SpatialCrossMapLRN', + 'SpatialFullConvolution', + 'SpatialMaxPooling', + 'SpatialAveragePooling', + 'ReLU', + 'Tanh', + 'Sigmoid', + 'SoftMax', + 'LogSoftMax', + 'VolumetricBatchNormalization', + 'VolumetricConvolution', + 'VolumetricFullConvolution', + 'VolumetricMaxPooling', + 'VolumetricAveragePooling', +} + +-- goes over a given net and converts all layers to dst backend +-- for example: net = cudnn_convert_custom(net, cudnn) +-- same as cudnn.convert with gModule check commented out +function cudnn_convert_custom(net, dst, exclusion_fn) + return net:replace(function(x) + --if torch.type(x) == 'nn.gModule' then + -- io.stderr:write('Warning: cudnn.convert does not work with nngraph yet. Ignoring nn.gModule') + -- return x + --end + local y = 0 + local src = dst == nn and cudnn or nn + local src_prefix = src == nn and 'nn.' or 'cudnn.' + local dst_prefix = dst == nn and 'nn.' or 'cudnn.' + + local function convert(v) + local y = {} + torch.setmetatable(y, dst_prefix..v) + if v == 'ReLU' then y = dst.ReLU() end -- because parameters + for k,u in pairs(x) do y[k] = u end + if src == cudnn and x.clearDesc then x.clearDesc(y) end + if src == cudnn and v == 'SpatialAveragePooling' then + y.divide = true + y.count_include_pad = v.mode == 'CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING' + end + if src == nn and string.find(v, 'Convolution') then + y.groups = 1 + end + return y + end + + if exclusion_fn and exclusion_fn(x) then + return x + end + local t = torch.typename(x) + if t == 'nn.SpatialConvolutionMM' then + y = convert('SpatialConvolution') + elseif t == 'inn.SpatialCrossResponseNormalization' then + y = convert('SpatialCrossMapLRN') + else + for i,v in ipairs(layer_list) do + if torch.typename(x) == src_prefix..v then + y = convert(v) + end + end + end + return y == 0 and x or y + end) +end diff --git a/cyclegan/util/image_pool.lua b/cyclegan/util/image_pool.lua new file mode 100644 index 0000000..f62cd85 --- /dev/null +++ b/cyclegan/util/image_pool.lua @@ -0,0 +1,59 @@ +local class = require 'class' +ImagePool= class('ImagePool') + +require 'torch' +require 'image' + +function ImagePool:__init(pool_size) + self.pool_size = pool_size + if pool_size > 0 then + self.num_imgs = 0 + self.images = {} + end +end + +function ImagePool:model_name() + return 'ImagePool' +end +-- +-- function ImagePool:Initialize(pool_size) +-- -- torch.manualSeed(0) +-- -- assert(pool_size > 0) +-- self.pool_size = pool_size +-- if pool_size > 0 then +-- self.num_imgs = 0 +-- self.images = {} +-- end +-- end + +function ImagePool:Query(image) + -- print('query image') + if self.pool_size == 0 then + -- print('get identical image') + return image + end + if self.num_imgs < self.pool_size then + -- self.images.insert(image:clone()) + self.num_imgs = self.num_imgs + 1 + self.images[self.num_imgs] = image + return image + else + local p = math.random() + -- print('p' ,p) + -- os.exit() + if p > 0.5 then + -- print('use old image') + -- random_id = torch.Tensor(1) + -- random_id:random(1, self.pool_size) + local random_id = math.random(self.pool_size) + -- print('random_id', random_id) + local tmp = self.images[random_id]:clone() + self.images[random_id] = image:clone() + return tmp + else + return image + end + + end + +end diff --git a/cyclegan/util/plot_util.lua b/cyclegan/util/plot_util.lua new file mode 100644 index 0000000..b20bcb6 --- /dev/null +++ b/cyclegan/util/plot_util.lua @@ -0,0 +1,44 @@ +local class = require 'class' +PlotUtil = class('PlotUtil') + + +require 'torch' +disp = require 'display' +util = require 'util/util' +require 'image' + +local unpack = unpack or table.unpack + +function PlotUtil:__init(conf) + conf = conf or {} +end + +function PlotUtil:model_name() + return 'PlotUtil' +end + +function PlotUtil:Initialize(display_plot, display_id, name) + self.display_plot = string.split(string.gsub(display_plot, "%s+", ""), ",") + + self.plot_config = { + title = name .. ' loss over time', + labels = {'epoch', unpack(self.display_plot)}, + ylabel = 'loss', + win = display_id, + } + + self.plot_data = {} + print('display_opt', self.display_plot) +end + + +function PlotUtil:Display(plot_vals, loss) + for k, v in ipairs(self.display_plot) do + if loss[v] ~= nil then + plot_vals[#plot_vals + 1] = loss[v] + end + end + + table.insert(self.plot_data, plot_vals) + disp.plot(self.plot_data, self.plot_config) +end diff --git a/cyclegan/util/util.lua b/cyclegan/util/util.lua new file mode 100644 index 0000000..2b00ede --- /dev/null +++ b/cyclegan/util/util.lua @@ -0,0 +1,266 @@ +-- +-- code derived from https://github.com/soumith/dcgan.torch +-- + +local util = {} + +require 'torch' + + +function util.BiasZero(net) + net:apply(function(m) if torch.type(m):find('Convolution') then m.bias:zero() end end) +end + + +function util.checkEqual(A, B, name) + local dif = (A:float()-B:float()):abs():mean() + print(name, dif) +end + +function util.containsValue(table, value) + for k, v in pairs(table) do + if v == value then return true end + end + return false +end + + +function util.CheckTensor(A, name) + print(name, A:min(), A:max(), A:mean()) +end + + +function util.normalize(img) + -- rescale image to 0 .. 1 + local min = img:min() + local max = img:max() + + img = torch.FloatTensor(img:size()):copy(img) + img:add(-min):mul(1/(max-min)) + return img +end + +function util.normalizeBatch(batch) + for i = 1, batch:size(1) do + batch[i] = util.normalize(batch[i]:squeeze()) + end + return batch +end + +function util.basename_batch(batch) + for i = 1, #batch do + batch[i] = paths.basename(batch[i]) + end + return batch +end + + + +-- default preprocessing +-- +-- Preprocesses an image before passing it to a net +-- Converts from RGB to BGR and rescales from [0,1] to [-1,1] +function util.preprocess(img) + -- RGB to BGR + if img:size(1) == 3 then + local perm = torch.LongTensor{3, 2, 1} + img = img:index(1, perm) + end + -- [0,1] to [-1,1] + img = img:mul(2):add(-1) + + -- check that input is in expected range + assert(img:max()<=1,"badly scaled inputs") + assert(img:min()>=-1,"badly scaled inputs") + + return img +end + +-- Undo the above preprocessing. +function util.deprocess(img) + -- BGR to RGB + if img:size(1) == 3 then + local perm = torch.LongTensor{3, 2, 1} + img = img:index(1, perm) + end + + -- [-1,1] to [0,1] + img = img:add(1):div(2) + + return img +end + +function util.preprocess_batch(batch) + for i = 1, batch:size(1) do + batch[i] = util.preprocess(batch[i]:squeeze()) + end + return batch +end + +function util.print_tensor(name, x) + print(name, x:size(), x:min(), x:mean(), x:max()) +end + +function util.deprocess_batch(batch) + for i = 1, batch:size(1) do + batch[i] = util.deprocess(batch[i]:squeeze()) + end + return batch +end + + +function util.scaleBatch(batch,s1,s2) + -- print('s1', s1) + -- print('s2', s2) + local scaled_batch = torch.Tensor(batch:size(1),batch:size(2),s1,s2) + for i = 1, batch:size(1) do + scaled_batch[i] = image.scale(batch[i],s1,s2):squeeze() + end + return scaled_batch +end + + + +function util.toTrivialBatch(input) + return input:reshape(1,input:size(1),input:size(2),input:size(3)) +end +function util.fromTrivialBatch(input) + return input[1] +end + +-- input is between -1 and 1 +function util.jitter(input) + local noise = torch.rand(input:size())/256.0 + input:add(1.0):mul(0.5*255.0/256.0):add(noise):add(-0.5):mul(2.0) + --local scaled = (input+1.0)*0.5 + --local jittered = scaled*255.0/256.0 + torch.rand(input:size())/256.0 + --local scaled_back = (jittered-0.5)*2.0 + --return scaled_back +end + +function util.scaleImage(input, loadSize) + + -- replicate bw images to 3 channels + if input:size(1)==1 then + input = torch.repeatTensor(input,3,1,1) + end + + input = image.scale(input, loadSize, loadSize) + + return input +end + +function util.getAspectRatio(path) + local input = image.load(path, 3, 'float') + local ar = input:size(3)/input:size(2) + return ar +end + +function util.loadImage(path, loadSize, nc) + local input = image.load(path, 3, 'float') + input= util.preprocess(util.scaleImage(input, loadSize)) + + if nc == 1 then + input = input[{{1}, {}, {}}] + end + + return input +end + +function file_exists(filename) + local f = io.open(filename,"r") + if f ~= nil then io.close(f) return true else return false end +end + +-- TO DO: loading code is rather hacky; clean it up and make sure it works on all types of nets / cpu/gpu configurations +function load_helper(filename, opt) + fileExists = file_exists(filename) + if not fileExists then + print('model not found! ' .. filename) + return nil + end + print(('loading previously trained model (%s)'):format(filename)) + if opt.norm == 'instance' then + print('use InstanceNormalization') + require 'util.InstanceNormalization' + end + + if opt.cudnn>0 then + require 'cudnn' + end + + local net = torch.load(filename) + if opt.gpu > 0 then + require 'cunn' + net:cuda() + + -- calling cuda on cudnn saved nngraphs doesn't change all variables to cuda, so do it below + if net.forwardnodes then + for i=1,#net.forwardnodes do + if net.forwardnodes[i].data.module then + net.forwardnodes[i].data.module:cuda() + end + end + end + + else + net:float() + end + net:apply(function(m) if m.weight then + m.gradWeight = m.weight:clone():zero(); + m.gradBias = m.bias:clone():zero(); end end) + return net +end + +function util.load_model(name, opt) + -- if opt['lambda_'.. name] > 0.0 then + -- print('not loading model '.. opt.checkpoints_dir .. opt.name .. + -- 'latest_net_' .. name .. '.t7' .. ' because opt.lambda is not greater than zero') + return load_helper(paths.concat(opt.checkpoints_dir, opt.name, + 'latest_net_' .. name .. '.t7'), opt) + -- end +end + +function util.load_test_model(name, opt) + return load_helper(paths.concat(opt.checkpoints_dir, opt.name, + opt.which_epoch .. '_net_' .. name .. '.t7'), opt) +end + + +-- load dataset from the file system +-- |name|: name of the dataset. It's currently either 'A' or 'B' +-- function util.load_dataset(name, nc, opt, nc) +-- local tensortype = torch.getdefaulttensortype() +-- torch.setdefaulttensortype('torch.FloatTensor') +-- +-- local new_opt = options.clone(opt) +-- new_opt.manualSeed = torch.random(1, 10000) -- fix seed +-- new_opt.nc = nc +-- torch.manualSeed(new_opt.manualSeed) +-- local data_loader = paths.dofile('../data/data.lua') +-- new_opt.phase = new_opt.phase .. name +-- local data = data_loader.new(new_opt.nThreads, new_opt) +-- print("Dataset Size " .. name .. ": ", data:size()) +-- +-- torch.setdefaulttensortype(tensortype) +-- return data +-- end + + + +function util.cudnn(net) + require 'cudnn' + require 'util/cudnn_convert_custom' + return cudnn_convert_custom(net, cudnn) +end + +function util.save_model(net, net_name, weight) + if weight > 0.0 then + torch.save(paths.concat(opt.checkpoints_dir, opt.name, net_name), net:clearState()) + end +end + + + + +return util diff --git a/cyclegan/util/visualizer.lua b/cyclegan/util/visualizer.lua new file mode 100644 index 0000000..2cf2ca3 --- /dev/null +++ b/cyclegan/util/visualizer.lua @@ -0,0 +1,63 @@ +------------------------------------------------------------- +-- Various utilities for visualization through the web server +------------------------------------------------------------- + +local visualizer = {} + +require 'torch' +disp = nil +print(opt) +if opt.display_id > 0 then -- [hack]: assume that opt already existed + disp = require 'display' +end +util = require 'util/util' +require 'image' + +-- function visualizer +function visualizer.disp_image(img_data, win_size, display_id, title) + images = util.deprocess_batch(util.scaleBatch(img_data:float(),win_size,win_size)) + disp.image(images, {win=display_id, title=title}) +end + +function visualizer.save_results(img_data, output_path) + local tensortype = torch.getdefaulttensortype() + torch.setdefaulttensortype('torch.FloatTensor') + local image_out = nil + local win_size = opt.display_winsize + images = torch.squeeze(util.deprocess_batch(util.scaleBatch(img_data:float(), win_size, win_size))) + + if images:dim() == 3 then + image_out = images + else + for i = 1,images:size(1) do + img = images[i] + if image_out == nil then + image_out = img + else + image_out = torch.cat(image_out, img) + end + end + end + image.save(output_path, image_out) + torch.setdefaulttensortype(tensortype) +end + +function visualizer.save_images(imgs, save_dir, impaths, s1, s2) + local tensortype = torch.getdefaulttensortype() + torch.setdefaulttensortype('torch.FloatTensor') + batchSize = imgs:size(1) + imgs_f = util.deprocess_batch(imgs):float() + paths.mkdir(save_dir) + for i = 1, batchSize do -- imgs_f[i]:size(2), imgs_f[i]:size(3)/opt.aspect_ratio + if s1 ~= nil and s2 ~= nil then + im_s = image.scale(imgs_f[i], s1, s2):float() + else + im_s = imgs_f[i]:float() + end + img_to_save = torch.FloatTensor(im_s:size()):copy(im_s) + image.save(paths.concat(save_dir, impaths[i]), img_to_save) + end + torch.setdefaulttensortype(tensortype) +end + +return visualizer