165 lines
7.1 KiB
Python
165 lines
7.1 KiB
Python
from itertools import chain
|
|
|
|
import ignite.distributed as idist
|
|
import torch
|
|
import torch.nn as nn
|
|
from omegaconf import OmegaConf
|
|
|
|
from engine.base.i2i import EngineKernel, run_kernel
|
|
from engine.util.build import build_model
|
|
from loss.I2I.edge_loss import EdgeLoss
|
|
from loss.I2I.perceptual_loss import PerceptualLoss
|
|
from loss.gan import GANLoss
|
|
from model.weight_init import generation_init_weights
|
|
|
|
|
|
class TAFGEngineKernel(EngineKernel):
|
|
def __init__(self, config):
|
|
super().__init__(config)
|
|
perceptual_loss_cfg = OmegaConf.to_container(config.loss.perceptual)
|
|
perceptual_loss_cfg.pop("weight")
|
|
self.perceptual_loss = PerceptualLoss(**perceptual_loss_cfg).to(idist.device())
|
|
|
|
gan_loss_cfg = OmegaConf.to_container(config.loss.gan)
|
|
gan_loss_cfg.pop("weight")
|
|
self.gan_loss = GANLoss(**gan_loss_cfg).to(idist.device())
|
|
|
|
self.recon_loss = nn.L1Loss() if config.loss.recon.level == 1 else nn.MSELoss()
|
|
self.style_recon_loss = nn.L1Loss() if config.loss.style_recon.level == 1 else nn.MSELoss()
|
|
self.content_recon_loss = nn.L1Loss() if config.loss.content_recon.level == 1 else nn.MSELoss()
|
|
self.cycle_loss = nn.L1Loss() if config.loss.cycle.level == 1 else nn.MSELoss()
|
|
|
|
self.edge_loss = EdgeLoss("HED", hed_pretrained_model_path=config.loss.edge.hed_pretrained_model_path).to(
|
|
idist.device())
|
|
|
|
def _process_batch(self, batch, inference=False):
|
|
# batch["b"] = batch["b"] if inference else batch["b"][0].expand(batch["a"].size())
|
|
return batch
|
|
|
|
def build_models(self) -> (dict, dict):
|
|
generators = dict(
|
|
main=build_model(self.config.model.generator)
|
|
)
|
|
discriminators = dict(
|
|
a=build_model(self.config.model.discriminator),
|
|
b=build_model(self.config.model.discriminator)
|
|
)
|
|
self.logger.debug(discriminators["a"])
|
|
self.logger.debug(generators["main"])
|
|
|
|
for m in chain(generators.values(), discriminators.values()):
|
|
generation_init_weights(m)
|
|
|
|
return generators, discriminators
|
|
|
|
def setup_after_g(self):
|
|
for discriminator in self.discriminators.values():
|
|
discriminator.requires_grad_(True)
|
|
|
|
def setup_before_g(self):
|
|
for discriminator in self.discriminators.values():
|
|
discriminator.requires_grad_(False)
|
|
|
|
def forward(self, batch, inference=False) -> dict:
|
|
generator = self.generators["main"]
|
|
batch = self._process_batch(batch, inference)
|
|
|
|
styles = dict()
|
|
contents = dict()
|
|
images = dict()
|
|
with torch.set_grad_enabled(not inference):
|
|
for ph in "ab":
|
|
contents[ph], styles[ph] = generator.encode(batch[ph]["edge"], batch[ph]["img"], ph, ph)
|
|
for ph in ("a2b", "b2a"):
|
|
images[f"fake_{ph[-1]}"] = generator.decode(contents[ph[0]], styles[ph[-1]], ph[-1])
|
|
contents["recon_a"], styles["recon_b"] = generator.encode(
|
|
self.edge_loss.edge_extractor(images["fake_b"]), images["fake_b"], "b", "b")
|
|
images["a2a"] = generator.decode(contents["a"], styles["a"], "a")
|
|
images["b2b"] = generator.decode(contents["b"], styles["recon_b"], "b")
|
|
images["cycle_a"] = generator.decode(contents["recon_a"], styles["a"], "a")
|
|
return dict(styles=styles, contents=contents, images=images)
|
|
|
|
def criterion_generators(self, batch, generated) -> dict:
|
|
batch = self._process_batch(batch)
|
|
loss = dict()
|
|
|
|
for ph in "ab":
|
|
loss[f"recon_image_{ph}"] = self.config.loss.recon.weight * self.recon_loss(
|
|
generated["images"][f"{ph}2{ph}"], batch[ph]["img"])
|
|
|
|
pred_fake = self.discriminators[ph](generated["images"][f"fake_{ph}"])
|
|
loss[f"gan_{ph}"] = 0
|
|
for sub_pred_fake in pred_fake:
|
|
# last output is actual prediction
|
|
loss[f"gan_{ph}"] += self.gan_loss(sub_pred_fake[-1], True) * self.config.loss.gan.weight
|
|
loss[f"recon_content_a"] = self.config.loss.content_recon.weight * self.content_recon_loss(
|
|
generated["contents"]["a"], generated["contents"]["recon_a"]
|
|
)
|
|
loss[f"recon_style_b"] = self.config.loss.style_recon.weight * self.style_recon_loss(
|
|
generated["styles"]["b"], generated["styles"]["recon_b"]
|
|
)
|
|
|
|
for ph in ("a2b", "b2a"):
|
|
if self.config.loss.perceptual.weight > 0:
|
|
loss[f"perceptual_{ph}"] = self.config.loss.perceptual.weight * self.perceptual_loss(
|
|
batch[ph[0]]["img"], generated["images"][f"fake_{ph[-1]}"]
|
|
)
|
|
if self.config.loss.edge.weight > 0:
|
|
loss[f"edge_a"] = self.config.loss.edge.weight * self.edge_loss(
|
|
generated["images"]["fake_b"], batch["a"]["edge"][:, 0:1, :, :]
|
|
)
|
|
loss[f"edge_b"] = self.config.loss.edge.weight * self.edge_loss(
|
|
generated["images"]["fake_a"], batch["b"]["edge"]
|
|
)
|
|
|
|
if self.config.loss.cycle.weight > 0:
|
|
loss[f"cycle_a"] = self.config.loss.cycle.weight * self.cycle_loss(
|
|
batch["a"]["img"], generated["images"]["cycle_a"]
|
|
)
|
|
return loss
|
|
|
|
def criterion_discriminators(self, batch, generated) -> dict:
|
|
loss = dict()
|
|
# batch = self._process_batch(batch)
|
|
for phase in self.discriminators.keys():
|
|
pred_real = self.discriminators[phase](batch[phase]["img"])
|
|
pred_fake = self.discriminators[phase](generated["images"][f"fake_{phase}"].detach())
|
|
loss[f"gan_{phase}"] = 0
|
|
for i in range(len(pred_fake)):
|
|
loss[f"gan_{phase}"] += (self.gan_loss(pred_fake[i][-1], False, is_discriminator=True)
|
|
+ self.gan_loss(pred_real[i][-1], True, is_discriminator=True)) / 2
|
|
return loss
|
|
|
|
def intermediate_images(self, batch, generated) -> dict:
|
|
"""
|
|
returned dict must be like: {"a": [img1, img2, ...], "b": [img3, img4, ...]}
|
|
:param batch:
|
|
:param generated: dict of images
|
|
:return: dict like: {"a": [img1, img2, ...], "b": [img3, img4, ...]}
|
|
"""
|
|
batch = self._process_batch(batch)
|
|
return dict(
|
|
a=[batch["a"]["edge"][:, 0:1, :, :].expand(-1, 3, -1, -1).detach(),
|
|
batch["a"]["img"].detach(),
|
|
generated["images"]["a2a"].detach(),
|
|
generated["images"]["fake_b"].detach(),
|
|
generated["images"]["cycle_a"].detach(),
|
|
],
|
|
b=[batch["b"]["edge"].expand(-1, 3, -1, -1).detach(),
|
|
batch["b"]["img"].detach(),
|
|
generated["images"]["b2b"].detach(),
|
|
generated["images"]["fake_a"].detach()]
|
|
)
|
|
|
|
def change_engine(self, config, trainer):
|
|
pass
|
|
# @trainer.on(Events.ITERATION_STARTED(once=int(config.max_iteration / 3)))
|
|
# def change_config(engine):
|
|
# with read_write(config):
|
|
# config.loss.perceptual.weight = 5
|
|
|
|
|
|
def run(task, config, _):
|
|
kernel = TAFGEngineKernel(config)
|
|
run_kernel(task, config, kernel)
|