UnboundLocalError occurred when not using onnx and with --max_data_loader_n_workers

[frodo821]

I have been experiencing an issue with tag_images_by_wd14_tagger.py.
I ran finetune/tag_images_by_wd14_tagger.py by using the options shown below:

$ accelerate launch "sd-scripts/finetune/tag_images_by_wd14_tagger.py" \
>   --batch_size=8 --general_threshold=0.35 --character_threshold=0.35 \
>   --caption_extension=".txt" --caption_separator="," \
>   --model="SmilingWolf/wd-v1-4-convnextv2-tagger-v2" \
>   --max_data_loader_n_workers="2" --debug --remove_underscore \
>   --frequency_tags "data/images"

Then I got this UnboundLocalError:

Traceback (most recent call last):
  File "/workspace/sd-scripts/finetune/tag_images_by_wd14_tagger.py", line 514, in <module>
    main(args)
  File "/workspace/sd-scripts/finetune/tag_images_by_wd14_tagger.py", line 333, in main
    data = torch.utils.data.DataLoader(
UnboundLocalError: local variable 'torch' referenced before assignment
Traceback (most recent call last):
  File "/opt/conda/bin/accelerate", line 8, in <module>
    sys.exit(main())
  File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/accelerate_cli.py", line 46, in main
    args.func(args)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/launch.py", line 1075, in launch_command
    simple_launcher(args)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/launch.py", line 681, in simple_launcher
    raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)

I have investigated why the error occurs, and I noticed an unnecessary extra import of torch is made here:

sd-scripts/finetune/tag_images_by_wd14_tagger.py

Lines 113 to 116 in 71e2c91

	# モデルを読み込む
	if args.onnx:
	import torch
	import onnx

I suspect this extra import statement makes torch a local, unbound variable, and it is only available when run with the --onnx option.

torch is already imported this line:

sd-scripts/finetune/tag_images_by_wd14_tagger.py

Line 8 in 71e2c91

import torch

And the import statement on line 115 must be removed.