mirror of
https://github.com/facefusion/facefusion-labs.git
synced 2026-04-19 15:56:37 +02:00
harisreedhar
250 lines
11 KiB
Python
250 lines
11 KiB
Python
import os
|
|
import warnings
|
|
from configparser import ConfigParser
|
|
from typing import Tuple
|
|
|
|
import torch
|
|
import torchvision
|
|
from lightning import LightningModule, Trainer
|
|
from lightning.pytorch.callbacks import ModelCheckpoint
|
|
from lightning.pytorch.loggers import TensorBoardLogger
|
|
from torch import Tensor, nn
|
|
from torch.utils.data import Dataset, random_split
|
|
from torchdata.stateful_dataloader import StatefulDataLoader
|
|
|
|
from .dataset import DynamicDataset
|
|
from .helper import calc_embedding, overlay_mask
|
|
from .models.discriminator import Discriminator
|
|
from .models.generator import Generator
|
|
from .models.loss import AdversarialLoss, AttributeLoss, DiscriminatorLoss, GazeLoss, IdentityLoss, MaskLoss, MotionLoss, ReconstructionLoss
|
|
from .networks.masknet import MaskNet
|
|
from .types import Batch, Embedding, OptimizerSet
|
|
|
|
warnings.filterwarnings('ignore', category = UserWarning, module = 'torch')
|
|
|
|
CONFIG_PARSER = ConfigParser()
|
|
CONFIG_PARSER.read('config.ini')
|
|
|
|
|
|
class FaceSwapperTrainer(LightningModule):
|
|
def __init__(self, config_parser : ConfigParser) -> None:
|
|
super().__init__()
|
|
self.config_embedder_path = config_parser.get('training.model', 'embedder_path')
|
|
self.config_gazer_path = config_parser.get('training.model', 'gazer_path')
|
|
self.config_motion_extractor_path = config_parser.get('training.model', 'motion_extractor_path')
|
|
self.config_face_parser_path = config_parser.get('training.model', 'face_parser_path')
|
|
self.config_accumulate_size = config_parser.getfloat('training.trainer', 'accumulate_size')
|
|
self.config_learning_rate = config_parser.getfloat('training.trainer', 'learning_rate')
|
|
self.config_preview_frequency = config_parser.getint('training.trainer', 'preview_frequency')
|
|
self.embedder = torch.jit.load(self.config_embedder_path, map_location = 'cpu').eval()
|
|
self.gazer = torch.jit.load(self.config_gazer_path, map_location = 'cpu').eval()
|
|
self.motion_extractor = torch.jit.load(self.config_motion_extractor_path, map_location = 'cpu').eval()
|
|
self.face_parser = torch.jit.load(self.config_face_parser_path, map_location ='cpu').eval()
|
|
self.generator = Generator(config_parser)
|
|
self.discriminator = Discriminator(config_parser)
|
|
self.masker = MaskNet(config_parser)
|
|
self.discriminator_loss = DiscriminatorLoss()
|
|
self.adversarial_loss = AdversarialLoss(config_parser)
|
|
self.attribute_loss = AttributeLoss(config_parser)
|
|
self.reconstruction_loss = ReconstructionLoss(config_parser, self.embedder)
|
|
self.identity_loss = IdentityLoss(config_parser, self.embedder)
|
|
self.motion_loss = MotionLoss(config_parser, self.motion_extractor)
|
|
self.gaze_loss = GazeLoss(config_parser, self.gazer)
|
|
self.mask_loss = MaskLoss(config_parser, self.face_parser)
|
|
self.automatic_optimization = False
|
|
|
|
def forward(self, source_embedding : Embedding, target_tensor : Tensor) -> Tuple[Tensor, Tensor]:
|
|
output_tensor, mask_tensor = self.generator(source_embedding, target_tensor)
|
|
return output_tensor, mask_tensor
|
|
|
|
def configure_optimizers(self) -> Tuple[OptimizerSet, OptimizerSet, OptimizerSet]:
|
|
generator_optimizer = torch.optim.AdamW(self.generator.parameters(), lr = self.config_learning_rate, betas = (0.0, 0.999), weight_decay = 1e-4)
|
|
discriminator_optimizer = torch.optim.AdamW(self.discriminator.parameters(), lr = self.config_learning_rate, betas = (0.0, 0.999), weight_decay = 1e-4)
|
|
masker_optimizer = torch.optim.AdamW(self.masker.parameters(), lr = self.config_learning_rate, betas = (0.0, 0.999), weight_decay = 1e-4)
|
|
generator_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(generator_optimizer, T_0 = 300, T_mult = 2)
|
|
discriminator_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(discriminator_optimizer, T_0 = 300, T_mult = 2)
|
|
masker_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(masker_optimizer, T_0 = 300, T_mult = 2)
|
|
|
|
generator_config =\
|
|
{
|
|
'optimizer': generator_optimizer,
|
|
'lr_scheduler':
|
|
{
|
|
'scheduler': generator_scheduler,
|
|
'interval': 'step'
|
|
}
|
|
}
|
|
discriminator_config =\
|
|
{
|
|
'optimizer': discriminator_optimizer,
|
|
'lr_scheduler':
|
|
{
|
|
'scheduler': discriminator_scheduler,
|
|
'interval': 'step'
|
|
}
|
|
}
|
|
masker_config =\
|
|
{
|
|
'optimizer': masker_optimizer,
|
|
'lr_scheduler':
|
|
{
|
|
'scheduler': masker_scheduler,
|
|
'interval': 'step'
|
|
}
|
|
}
|
|
return generator_config, discriminator_config, masker_config
|
|
|
|
def training_step(self, batch : Batch, batch_index : int) -> Tensor:
|
|
source_tensor, target_tensor = batch
|
|
do_update = (batch_index + 1) % self.config_accumulate_size == 0
|
|
generator_optimizer, discriminator_optimizer, masker_optimizer = self.optimizers() #type:ignore[attr-defined]
|
|
|
|
self.toggle_optimizer(generator_optimizer)
|
|
source_embedding = calc_embedding(self.embedder, source_tensor, (0, 0, 0, 0))
|
|
target_attributes = self.generator.get_attributes(target_tensor)
|
|
generator_output_tensor = self.generator(source_embedding, target_tensor)
|
|
generator_output_attributes = self.generator.get_attributes(generator_output_tensor)
|
|
discriminator_output_tensors = self.discriminator(generator_output_tensor)
|
|
adversarial_loss, weighted_adversarial_loss = self.adversarial_loss(discriminator_output_tensors)
|
|
attribute_loss, weighted_attribute_loss = self.attribute_loss(target_attributes, generator_output_attributes)
|
|
reconstruction_loss, weighted_reconstruction_loss = self.reconstruction_loss(source_tensor, target_tensor, generator_output_tensor)
|
|
identity_loss, weighted_identity_loss = self.identity_loss(generator_output_tensor, source_tensor)
|
|
pose_loss, weighted_pose_loss, expression_loss, weighted_expression_loss = self.motion_loss(target_tensor, generator_output_tensor)
|
|
gaze_loss, weighted_gaze_loss = self.gaze_loss(target_tensor, generator_output_tensor)
|
|
generator_loss = weighted_adversarial_loss + weighted_attribute_loss + weighted_reconstruction_loss + weighted_identity_loss + weighted_pose_loss + weighted_gaze_loss + weighted_expression_loss
|
|
|
|
self.manual_backward(generator_loss)
|
|
|
|
if do_update:
|
|
generator_optimizer.step()
|
|
generator_optimizer.zero_grad()
|
|
|
|
self.untoggle_optimizer(generator_optimizer)
|
|
|
|
self.toggle_optimizer(masker_optimizer)
|
|
mask_tensor = self.masker(target_tensor, target_attributes[-1].detach())
|
|
mask_loss = self.mask_loss(target_tensor, mask_tensor)
|
|
|
|
self.manual_backward(mask_loss)
|
|
|
|
if do_update:
|
|
masker_optimizer.step()
|
|
masker_optimizer.zero_grad()
|
|
|
|
self.untoggle_optimizer(masker_optimizer)
|
|
|
|
self.toggle_optimizer(discriminator_optimizer)
|
|
discriminator_source_tensors = self.discriminator(source_tensor)
|
|
discriminator_output_tensors = self.discriminator(generator_output_tensor.detach())
|
|
discriminator_loss = self.discriminator_loss(discriminator_source_tensors, discriminator_output_tensors)
|
|
|
|
self.manual_backward(discriminator_loss)
|
|
|
|
if do_update:
|
|
discriminator_optimizer.step()
|
|
discriminator_optimizer.zero_grad()
|
|
|
|
self.untoggle_optimizer(discriminator_optimizer)
|
|
|
|
if self.global_step % self.config_preview_frequency == 0:
|
|
self.generate_preview(source_tensor, target_tensor, generator_output_tensor, mask_tensor)
|
|
|
|
self.log('generator_loss', generator_loss, prog_bar = True)
|
|
self.log('discriminator_loss', discriminator_loss, prog_bar = True)
|
|
self.log('adversarial_loss', adversarial_loss)
|
|
self.log('attribute_loss', attribute_loss)
|
|
self.log('reconstruction_loss', reconstruction_loss)
|
|
self.log('identity_loss', identity_loss)
|
|
self.log('pose_loss', pose_loss)
|
|
self.log('gaze_loss', gaze_loss)
|
|
self.log('mask_loss', mask_loss)
|
|
return generator_loss
|
|
|
|
def validation_step(self, batch : Batch, batch_index : int) -> Tensor:
|
|
source_tensor, target_tensor = batch
|
|
source_embedding = calc_embedding(self.embedder, source_tensor, (0, 0, 0, 0))
|
|
output_tensor = self.generator(source_embedding, target_tensor)
|
|
output_embedding = calc_embedding(self.embedder, output_tensor, (0, 0, 0, 0))
|
|
validation_score = (nn.functional.cosine_similarity(source_embedding, output_embedding).mean() + 1) * 0.5
|
|
self.log('validation_score', validation_score, prog_bar = True)
|
|
return validation_score
|
|
|
|
def generate_preview(self, source_tensor : Tensor, target_tensor : Tensor, output_tensor : Tensor, mask_tensor : Tensor) -> None:
|
|
preview_limit = 8
|
|
preview_cells = []
|
|
mask_tensor = overlay_mask(target_tensor, mask_tensor)
|
|
|
|
for source_tensor, target_tensor, output_tensor, mask_tensor in zip(source_tensor[:preview_limit], target_tensor[:preview_limit], output_tensor[:preview_limit], mask_tensor[:preview_limit]):
|
|
preview_cell = torch.cat([ source_tensor, target_tensor, output_tensor, mask_tensor ], dim = 2)
|
|
preview_cells.append(preview_cell)
|
|
|
|
preview_cells = torch.cat(preview_cells, dim = 1).unsqueeze(0)
|
|
preview_grid = torchvision.utils.make_grid(preview_cells, normalize = True, scale_each = True)
|
|
self.logger.experiment.add_image('preview', preview_grid, self.global_step) # type:ignore[attr-defined]
|
|
|
|
|
|
def create_loaders(dataset : Dataset[Tensor]) -> Tuple[StatefulDataLoader[Tensor], StatefulDataLoader[Tensor]]:
|
|
config_batch_size = CONFIG_PARSER.getint('training.loader', 'batch_size')
|
|
config_num_workers = CONFIG_PARSER.getint('training.loader', 'num_workers')
|
|
|
|
training_dataset, validate_dataset = split_dataset(dataset)
|
|
training_loader = StatefulDataLoader(training_dataset, batch_size = config_batch_size, shuffle = True, num_workers = config_num_workers, drop_last = True, pin_memory = True, persistent_workers = True)
|
|
validation_loader = StatefulDataLoader(validate_dataset, batch_size = config_batch_size, shuffle = False, num_workers = config_num_workers, pin_memory = True, persistent_workers = True)
|
|
return training_loader, validation_loader
|
|
|
|
|
|
def split_dataset(dataset : Dataset[Tensor]) -> Tuple[Dataset[Tensor], Dataset[Tensor]]:
|
|
config_split_ratio = CONFIG_PARSER.getfloat('training.loader', 'split_ratio')
|
|
|
|
dataset_size = len(dataset) # type:ignore[arg-type]
|
|
training_size = int(dataset_size * config_split_ratio)
|
|
validation_size = int(dataset_size - training_size)
|
|
training_dataset, validate_dataset = random_split(dataset, [ training_size, validation_size ])
|
|
return training_dataset, validate_dataset
|
|
|
|
|
|
def create_trainer() -> Trainer:
|
|
config_max_epochs = CONFIG_PARSER.getint('training.trainer', 'max_epochs')
|
|
config_precision = CONFIG_PARSER.get('training.trainer', 'precision')
|
|
config_strategy = CONFIG_PARSER.get('training.trainer', 'strategy')
|
|
config_directory_path = CONFIG_PARSER.get('training.output', 'directory_path')
|
|
config_file_pattern = CONFIG_PARSER.get('training.output', 'file_pattern')
|
|
logger = TensorBoardLogger('.logs', name = 'face_swapper')
|
|
|
|
return Trainer(
|
|
logger = logger,
|
|
log_every_n_steps = 10,
|
|
max_epochs = config_max_epochs,
|
|
precision = config_precision,
|
|
strategy = config_strategy,
|
|
callbacks =
|
|
[
|
|
ModelCheckpoint(
|
|
monitor = 'generator_loss',
|
|
dirpath = config_directory_path,
|
|
filename = config_file_pattern,
|
|
every_n_train_steps = 1000,
|
|
save_top_k = 3,
|
|
save_last = True
|
|
)
|
|
],
|
|
val_check_interval = 1000
|
|
)
|
|
|
|
|
|
def train() -> None:
|
|
config_resume_path = CONFIG_PARSER.get('training.output', 'resume_path')
|
|
|
|
if torch.cuda.is_available():
|
|
torch.set_float32_matmul_precision('high')
|
|
|
|
dataset = DynamicDataset(CONFIG_PARSER)
|
|
training_loader, validation_loader = create_loaders(dataset)
|
|
face_swapper_trainer = FaceSwapperTrainer(CONFIG_PARSER)
|
|
trainer = create_trainer()
|
|
|
|
if os.path.isfile(config_resume_path):
|
|
trainer.fit(face_swapper_trainer, training_loader, validation_loader, ckpt_path = config_resume_path)
|
|
else:
|
|
trainer.fit(face_swapper_trainer, training_loader, validation_loader)
|