Mirrors
/
EveryDream2trainer
mirror of https://github.com/victorchall/EveryDream2trainer.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

better undersized log file

This commit is contained in:
Victor Hall 2023-01-20 16:23:56 -05:00
parent 1c2708dc63
commit 36ece59660
2 changed files with 2 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
data/data_loader.py
View File

@ -231,7 +231,7 @@ class DataLoaderMultiAspect():
target_wh = min(self.aspects, key=lambda aspects:abs(aspects[0]/aspects[1] - image_aspect)) target_wh = min(self.aspects, key=lambda aspects:abs(aspects[0]/aspects[1] - image_aspect))
if not self.has_scanned: if not self.has_scanned:
if width * height < target_wh[0] * target_wh[1]: if width * height < target_wh[0] * target_wh[1]:
undersized_images.append(f" *** {pathname} with size: {width},{height} is smaller than target size: {target_wh}, consider using larger images") undersized_images.append(f" {pathname}, size: {width},{height}, target size: {target_wh}")
image_train_item = ImageTrainItem(image=None, caption=caption, target_wh=target_wh, pathname=pathname, flip_p=flip_p) image_train_item = ImageTrainItem(image=None, caption=caption, target_wh=target_wh, pathname=pathname, flip_p=flip_p)
@ -251,7 +251,7 @@ class DataLoaderMultiAspect():
with open(underized_log_path, "w") as undersized_images_file: with open(underized_log_path, "w") as undersized_images_file:
undersized_images_file.write(f" The following images are smaller than the target size, consider removing or sourcing a larger copy:") undersized_images_file.write(f" The following images are smaller than the target size, consider removing or sourcing a larger copy:")
for undersized_image in undersized_images: for undersized_image in undersized_images:
undersized_images_file.write(undersized_image) undersized_images_file.write(f"{undersized_image}\n")
return decorated_image_train_items return decorated_image_train_items

10
train.py
View File

@ -675,8 +675,6 @@ def main(args):
logging.info(f" {Fore.GREEN}batch_size: {Style.RESET_ALL}{Fore.LIGHTGREEN_EX}{args.batch_size}{Style.RESET_ALL}") logging.info(f" {Fore.GREEN}batch_size: {Style.RESET_ALL}{Fore.LIGHTGREEN_EX}{args.batch_size}{Style.RESET_ALL}")
logging.info(f" {Fore.GREEN}epoch_len: {Fore.LIGHTGREEN_EX}{epoch_len}{Style.RESET_ALL}") logging.info(f" {Fore.GREEN}epoch_len: {Fore.LIGHTGREEN_EX}{epoch_len}{Style.RESET_ALL}")
#scaler = torch.cuda.amp.GradScaler()
scaler = GradScaler( scaler = GradScaler(
enabled=args.amp, enabled=args.amp,
init_scale=2**17.5, init_scale=2**17.5,
@ -686,13 +684,8 @@ def main(args):
) )
logging.info(f" Grad scaler enabled: {scaler.is_enabled()} (amp mode)") logging.info(f" Grad scaler enabled: {scaler.is_enabled()} (amp mode)")
epoch_pbar = tqdm(range(args.max_epochs), position=0, leave=True) epoch_pbar = tqdm(range(args.max_epochs), position=0, leave=True)
epoch_pbar.set_description(f"{Fore.LIGHTCYAN_EX}Epochs{Style.RESET_ALL}") epoch_pbar.set_description(f"{Fore.LIGHTCYAN_EX}Epochs{Style.RESET_ALL}")
# steps_pbar = tqdm(range(epoch_len), position=1, leave=True)
# steps_pbar.set_description(f"{Fore.LIGHTCYAN_EX}Steps{Style.RESET_ALL}")
epoch_times = [] epoch_times = []
global global_step global global_step
@ -781,11 +774,8 @@ def main(args):
param.grad *= grad_scale param.grad *= grad_scale
if ((global_step + 1) % args.grad_accum == 0) or (step == epoch_len - 1): if ((global_step + 1) % args.grad_accum == 0) or (step == epoch_len - 1):
# if args.amp:
scaler.step(optimizer) scaler.step(optimizer)
scaler.update() scaler.update()
# else:
# optimizer.step()
optimizer.zero_grad(set_to_none=True) optimizer.zero_grad(set_to_none=True)
lr_scheduler.step() lr_scheduler.step()