mirror of
https://github.com/facefusion/facefusion-labs.git
synced 2026-04-19 15:56:37 +02:00
Henry Ruhs
2e6394565a
* add sync batchnorm * replace random.choice with hash * fifty percent reduction * fix discriminator input * restore dataset.py * Remove duplicates * add discriminator_ratio to config * fix onnx export bug: replace round() with int() * Fix embedding naming * Introduce ModelWithConfigCheckpoint callback (#86) * Fix dist ini * Style: Refactor typing and improve code clarity in training.py (#88) * Add type casting for trainer params * Add type casting for trainer params * Add type casting for trainer params * Remove inplace activations for torch.compile compatibility (#89) * Fix README * improvise with norm layers & weighted average * add skip layer * use gelu instead of leaky_relu * cleanup * cleanup * Update dependencies * Different defaults and enable validation * Different defaults and enable validation * Revert to higher batch size * Just use copy over copy2 --------- Co-authored-by: harisreedhar <h4harisreedhar.s.s@gmail.com> Co-authored-by: NeuroDonu <112660822+NeuroDonu@users.noreply.github.com> Co-authored-by: Harisreedhar <46858047+harisreedhar@users.noreply.github.com>
HyperSwap
Hyper accurate face swapping for everyone.
Preview
Installation
pip install -r requirements.txt
Setup
This config.ini utilizes the VGGFace2 dataset to train the HyperSwap model.
[training.dataset]
file_pattern = .datasets/vggface2/**/*.jpg
convert_template = vggfacehq_512_to_arcface_128
multiplier = 1
transform_size = 256
usage_mode = both
batch_mode = same
batch_ratio = 0.2
[training.loader]
batch_size = 8
num_workers = 8
split_ratio = 0.9995
[training.model]
generator_embedder_path = .models/blendface.pt
loss_embedder_path = .models/arcface.pt
gazer_path = .models/gazer.pt
face_masker_path = .models/face_masker.pt
[training.model.generator]
source_channels = 512
output_size = 256
num_blocks = 2
[training.model.discriminator]
input_channels = 3
num_filters = 64
num_layers = 5
num_discriminators = 3
kernel_size = 4
[training.model.masker]
input_channels = 67
output_channels = 1
num_filters = 16
[training.losses]
adversarial_weight = 1.0
cycle_weight = 1.0
feature_weight = 10.0
reconstruction_weight = 10.0
identity_weight = 20.0
gaze_weight = 0.05
mask_weight = 5.0
[training.trainer]
accumulate_size = 4
discriminator_ratio = 0.4
gradient_clip = 20.0
max_epochs = 50
strategy = auto
precision = 16-mixed
sync_batchnorm = false
preview_frequency = 100
[training.modifier]
mask_factor = 0.01
noise_factor = 0.05
[training.optimizer.generator]
learning_rate = 0.0004
momentum = 0.5
scheduler_factor = 0.7
scheduler_patience = 2000
[training.optimizer.discriminator]
learning_rate = 0.0002
momentum = 0.5
scheduler_factor = 0.7
scheduler_patience = 2000
[training.logger]
logger_path = .logs
logger_name = hyperswap
[training.output]
directory_path = .outputs
file_pattern = hyperswap_{epoch}_{step}
resume_path = .outputs/last.ckpt
[exporting]
directory_path = .exports
source_path = .outputs/last.ckpt
target_path = .exports/hyperswap_256.onnx
target_size = 256
ir_version = 10
opset_version = 15
precision = full
[inferencing]
generator_path = .outputs/last.ckpt
embedder_path = .models/arcface.pt
source_path = .assets/source.jpg
target_path = .assets/target.jpg
output_path = .outputs/output.jpg
Training
Train the model.
python train.py
Launch the TensorBoard to monitor the training.
tensorboard --logdir .logs
Exporting
Export the model to ONNX.
python export.py
Inferencing
Inference the model.
python infer.py