Will the test code be updated to properly work with own 512 and 224 pretrained models? #261

New Issue

Open

opened 2022-05-08 09:17:21 +02:00 by netrunner-exe · 9 comments

netrunner-exe commented 2022-05-08 09:17:21 +02:00 (Migrated from github.com)

Copy Link

Copy Source

No description provided.

boreas-l commented 2022-05-09 10:39:10 +02:00 (Migrated from github.com)

Copy Link

Copy Source

I find:
1、in train.py and test_wholeimage_swapsingle.py, the initial network was not the same...
in train.py:

from models.projected_model import fsModel  
# model initial  
model = fsModel()  
# fsModel() source from **projected_model.py**, and in that, the Generator_Adain_Upsample source from fs_networks_fix

and in test_wholeimage_swapsingle.py:

from models.models import create_model  
# model initial
model = create_model(opt)  
# fsModel() source from fs_model.py, and in that, the Generator_Adain_Upsample source from fs_networks_512  

in general, the Generator_Adain_Upsample is diffenent between train.py and test_wholeimage_swapsingle.py.

2、so, after train, in your own pretrained_weight, you need initial network like train.py instead of test_wholeimage_swapsingle.py.

# test.py  
from models.projected_model import fsModel  

model = fsModel()  
# remenber add Gdeep=True in TestOptions 
model.initialize(opt)  
model.netG.eval()  

# and data preprocess like train.py
def data_preprocess(img):  
    img_tensor = transforms.ToTensor()(img)  
    img_tensor = img_tensor.view(-1, 3, img.size[0], img.size[1])  
    mean = torch.tensor([0.485, 0.456, 0.406]).cuda().view(1, 3, 1, 1)  
    std = torch.tensor([0.229, 0.224, 0.225]).cuda().view(1, 3, 1, 1)  
    img_tensor = img_tensor.cuda(non_blocking=True)  
    img_tensor = img_tensor.sub_(mean)._div(std)  
    return img_tensor  

I find: 1、in train.py and test_wholeimage_swapsingle.py, the initial network was not the same... in train.py: ``` from models.projected_model import fsModel # model initial model = fsModel() # fsModel() source from **projected_model.py**, and in that, the Generator_Adain_Upsample source from fs_networks_fix ``` and in test_wholeimage_swapsingle.py: ``` from models.models import create_model # model initial model = create_model(opt) # fsModel() source from fs_model.py, and in that, the Generator_Adain_Upsample source from fs_networks_512 ``` in general, the Generator_Adain_Upsample is diffenent between train.py and test_wholeimage_swapsingle.py. 2、so, after train, in your own pretrained_weight, you need initial network like train.py instead of test_wholeimage_swapsingle.py. ``` # test.py from models.projected_model import fsModel model = fsModel() # remenber add Gdeep=True in TestOptions model.initialize(opt) model.netG.eval() # and data preprocess like train.py def data_preprocess(img): img_tensor = transforms.ToTensor()(img) img_tensor = img_tensor.view(-1, 3, img.size[0], img.size[1]) mean = torch.tensor([0.485, 0.456, 0.406]).cuda().view(1, 3, 1, 1) std = torch.tensor([0.229, 0.224, 0.225]).cuda().view(1, 3, 1, 1) img_tensor = img_tensor.cuda(non_blocking=True) img_tensor = img_tensor.sub_(mean)._div(std) return img_tensor ```

👍 5

YaoQ commented 2022-05-09 10:55:25 +02:00 (Migrated from github.com)

Copy Link

Copy Source

The preprocess and the postprocess part for own trained model are different the inference demo code, and you have to update the demo code according to train.py

The preprocess and the postprocess part for own trained model are different the inference demo code, and you have to update the demo code according to train.py

netrunner-exe commented 2022-05-09 13:54:44 +02:00 (Migrated from github.com)

Copy Link

Copy Source

I find: 1、in train.py and test_wholeimage_swapsingle.py, the initial network was not the same... in train.py:

from models.projected_model import fsModel  
# model initial  
model = fsModel()  
# fsModel() source from **projected_model.py**, and in that, the Generator_Adain_Upsample source from fs_networks_fix

and in test_wholeimage_swapsingle.py:

from models.models import create_model  
# model initial
model = create_model(opt)  
# fsModel() source from fs_model.py, and in that, the Generator_Adain_Upsample source from fs_networks_512  

in general, the Generator_Adain_Upsample is diffenent between train.py and test_wholeimage_swapsingle.py.

2、so, after train, in your own pretrained_weight, you need initial network like train.py instead of test_wholeimage_swapsingle.py.

# test.py  
from models.projected_model import fsModel  

model = fsModel()  
# remenber add Gdeep=True in TestOptions 
model.initialize(opt)  
model.netG.eval()  

# and data preprocess like train.py
def data_preprocess(img):  
    img_tensor = transforms.ToTensor()(img)  
    img_tensor = img_tensor.view(-1, 3, img.size[0], img.size[1])  
    mean = torch.tensor([0.485, 0.456, 0.406]).cuda().view(1, 3, 1, 1)  
    std = torch.tensor([0.229, 0.224, 0.225]).cuda().view(1, 3, 1, 1)  
    img_tensor = img_tensor.cuda(non_blocking=True)  
    img_tensor = img_tensor.sub_(mean)._div(std)  
    return img_tensor  

Thank you very much! I tried to make changes but got errors. But still, the fact that at least someone managed to run the test code pleases me. Otherwise, what is the point of training if it is impossible to fully run the test code with own model? Just because it's been asked for so long? Let's wait and hope that the developer will fix the test code and give us the opportunity to run all test code with old models and with our own pretrained models as it should be.

> I find: 1、in train.py and test_wholeimage_swapsingle.py, the initial network was not the same... in train.py: > > ``` > from models.projected_model import fsModel > # model initial > model = fsModel() > # fsModel() source from **projected_model.py**, and in that, the Generator_Adain_Upsample source from fs_networks_fix > ``` > > and in test_wholeimage_swapsingle.py: > > ``` > from models.models import create_model > # model initial > model = create_model(opt) > # fsModel() source from fs_model.py, and in that, the Generator_Adain_Upsample source from fs_networks_512 > ``` > > in general, the Generator_Adain_Upsample is diffenent between train.py and test_wholeimage_swapsingle.py. > > 2、so, after train, in your own pretrained_weight, you need initial network like train.py instead of test_wholeimage_swapsingle.py. > > ``` > # test.py > from models.projected_model import fsModel > > model = fsModel() > # remenber add Gdeep=True in TestOptions > model.initialize(opt) > model.netG.eval() > > # and data preprocess like train.py > def data_preprocess(img): > img_tensor = transforms.ToTensor()(img) > img_tensor = img_tensor.view(-1, 3, img.size[0], img.size[1]) > mean = torch.tensor([0.485, 0.456, 0.406]).cuda().view(1, 3, 1, 1) > std = torch.tensor([0.229, 0.224, 0.225]).cuda().view(1, 3, 1, 1) > img_tensor = img_tensor.cuda(non_blocking=True) > img_tensor = img_tensor.sub_(mean)._div(std) > return img_tensor > ``` Thank you very much! I tried to make changes but got errors. But still, the fact that at least someone managed to run the test code pleases me. Otherwise, what is the point of training if it is impossible to fully run the test code with own model? Just because it's been asked for so long? Let's wait and hope that the developer will fix the test code and give us the opportunity to run all test code with old models and with our own pretrained models as it should be.

netrunner-exe commented 2022-05-10 11:16:05 +02:00 (Migrated from github.com)

Copy Link

Copy Source

I find: 1、in train.py and test_wholeimage_swapsingle.py, the initial network was not the same... in train.py:

from models.projected_model import fsModel  
# model initial  
model = fsModel()  
# fsModel() source from **projected_model.py**, and in that, the Generator_Adain_Upsample source from fs_networks_fix

and in test_wholeimage_swapsingle.py:

from models.models import create_model  
# model initial
model = create_model(opt)  
# fsModel() source from fs_model.py, and in that, the Generator_Adain_Upsample source from fs_networks_512  

in general, the Generator_Adain_Upsample is diffenent between train.py and test_wholeimage_swapsingle.py.

2、so, after train, in your own pretrained_weight, you need initial network like train.py instead of test_wholeimage_swapsingle.py.

# test.py  
from models.projected_model import fsModel  

model = fsModel()  
# remenber add Gdeep=True in TestOptions 
model.initialize(opt)  
model.netG.eval()  

# and data preprocess like train.py
def data_preprocess(img):  
    img_tensor = transforms.ToTensor()(img)  
    img_tensor = img_tensor.view(-1, 3, img.size[0], img.size[1])  
    mean = torch.tensor([0.485, 0.456, 0.406]).cuda().view(1, 3, 1, 1)  
    std = torch.tensor([0.229, 0.224, 0.225]).cuda().view(1, 3, 1, 1)  
    img_tensor = img_tensor.cuda(non_blocking=True)  
    img_tensor = img_tensor.sub_(mean)._div(std)  
    return img_tensor  

Can you provide full working with own pretrained 512 model example of test_wholeimage_swapsingle.py or manual on what and where to change or add?

> I find: 1、in train.py and test_wholeimage_swapsingle.py, the initial network was not the same... in train.py: > > ``` > from models.projected_model import fsModel > # model initial > model = fsModel() > # fsModel() source from **projected_model.py**, and in that, the Generator_Adain_Upsample source from fs_networks_fix > ``` > > and in test_wholeimage_swapsingle.py: > > ``` > from models.models import create_model > # model initial > model = create_model(opt) > # fsModel() source from fs_model.py, and in that, the Generator_Adain_Upsample source from fs_networks_512 > ``` > > in general, the Generator_Adain_Upsample is diffenent between train.py and test_wholeimage_swapsingle.py. > > 2、so, after train, in your own pretrained_weight, you need initial network like train.py instead of test_wholeimage_swapsingle.py. > > ``` > # test.py > from models.projected_model import fsModel > > model = fsModel() > # remenber add Gdeep=True in TestOptions > model.initialize(opt) > model.netG.eval() > > # and data preprocess like train.py > def data_preprocess(img): > img_tensor = transforms.ToTensor()(img) > img_tensor = img_tensor.view(-1, 3, img.size[0], img.size[1]) > mean = torch.tensor([0.485, 0.456, 0.406]).cuda().view(1, 3, 1, 1) > std = torch.tensor([0.229, 0.224, 0.225]).cuda().view(1, 3, 1, 1) > img_tensor = img_tensor.cuda(non_blocking=True) > img_tensor = img_tensor.sub_(mean)._div(std) > return img_tensor > ``` Can you provide full working with own pretrained 512 model example of `test_wholeimage_swapsingle.py` or manual on what and where to change or add?

boreas-l commented 2022-05-10 11:52:10 +02:00 (Migrated from github.com)

Copy Link

Copy Source

so glad, can you provide your email? or another way?

> so glad, can you provide your email? or another way?

netrunner-exe commented 2022-05-10 11:55:48 +02:00 (Migrated from github.com)

Copy Link

Copy Source

so glad, can you provide your email? or another way?

netrunner.exe@gmail.com @boreas-l

> > > > so glad, can you provide your email? or another way? netrunner.exe@gmail.com @boreas-l

netrunner-exe commented 2022-05-10 16:29:08 +02:00 (Migrated from github.com)

Copy Link

Copy Source

so glad, can you provide your email? or another way?

That would be great, looking forward to it!

> > > > so glad, can you provide your email? or another way? That would be great, looking forward to it!

BeaverInGreenland commented 2022-05-20 11:28:00 +02:00 (Migrated from github.com)

Copy Link

Copy Source

According to what is said here, how should we modify the arguments to make an inference on our own pretrained models ?

According to what is said [here](https://github.com/neuralchen/SimSwap/blob/main/docs/guidance/usage.md), how should we modify the arguments to make an inference on our own pretrained models ?

netrunner-exe commented 2022-05-20 13:00:39 +02:00 (Migrated from github.com)

Copy Link

Copy Source

According to what is said here, how should we modify the arguments to make an inference on our own pretrained models ?

Hah, in usage doesn't writed anything about how to use own pretrained models. Currently, the test code is incompatible with published training code and own pretrained models...Do not waste your time and resources - there is no point in training, since you will not be able to use your model in any way:) Therefore, we are waiting for the announced SimSwap ++, maybe we will have more luck with it :)

> According to what is said [here](https://github.com/neuralchen/SimSwap/blob/main/docs/guidance/usage.md), how should we modify the arguments to make an inference on our own pretrained models ? Hah, in usage doesn't writed anything about how to use own pretrained models. Currently, the test code is incompatible with published training code and own pretrained models...Do not waste your time and resources - there is no point in training, since you will not be able to use your model in any way:) Therefore, we are waiting for the announced SimSwap ++, maybe we will have more luck with it :)

Sign in to join this conversation.

No Branch/Tag Specified

Branches Tags

main

512_beta

1.0

Labels

Clear labels

GIF

GUI

Increase inference time

A solution to increase inference time

bug

Something isn't working

documentation

Improvements or additions to documentation

duplicate

This issue or pull request already exists

enhancement

New feature or request

good first issue

Good for newcomers

help wanted

Extra attention is needed

invalid

This doesn't seem right

question

Further information is requested

wontfix

This will not be worked on

No labels

Milestone

No items

No Milestone

Projects

Clear projects

No projects

Assignees

Clear assignees

CalvinBackup (Calvin Backup)

No Assignees

1 Participants

Notifications

Subscribe

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: CalvinBackup/neuralchen-SimSwap#261