unet ref training preview and inference script

[concrete13377]

after 500 epochs results of training preview do not match results of generated image with inference script

I tried finetuning previous checkpoints trained without data_online_creation_load_size_A option python3 train.py \ --dataroot /datasets/viton_ref/viton_bbox_ref \ --checkpoints_dir /checkpoints \ --name viton_bbox_ref \ --config_json examples/example_ddpm_unetref_viton.json \ --data_online_creation_load_size_A 768 1024 \ --train_continue \

then I tried inference python3 scripts/gen_single_image_diffusion.py \ --model-in-file /checkpoints/viton_bbox_ref/latest_net_G_A.pth \ --img-in /datasets/viton_ref/viton_bbox_ref/trainA/imgs/00000_00.jpg \ --bbox-in /datasets/viton_ref/viton_bbox_ref/trainA/bbox/00000_00.txt \ --ref-in /datasets/viton_ref/viton_bbox_ref/trainA/ref/00000_00.jpg \ --dir-out /checkpoints/viton_bbox_ref/inference_output \ --img-width 128 \ --img-height 128

I also tried inference with 96 128 (did not help improove results) and 512 512 (that approximatelly taking 6 hours)

[royale]

Hello, please could you try with the PR #577 and the following command:

python3 gen_single_image_diffusion.py --model-in-file /checkpoints/viton_bbox_ref/latest_net_G_A.pth --img-in /datasets/viton_ref/viton_bbox_ref/testA/imgs/00013_00.jpg --bbox-in /datasets/viton_ref/viton_bbox_ref/testA/bbox/00013_00.txt --ref-in /datasets/viton_ref/viton_bbox_ref/testA/ref/00017_00.jpg --dir-out /checkpoints/viton_bbox_ref/inference_output

[concrete13377]

it's similiar, it seems that the code that generates samples preview during training and this separate inference script have some drastic differences

[beniz]

This was closed by mistake (automatically from PR).
@concrete13377 can you try with the same image as @royale for a start ?
Some images have bboxes that are larger than the crop, these are corner cases.

This is what we get (from our internal chat):

The fuzzyness is from the resolution upsampling.

Note: models can be finetuned in higher dimensions easily as they are fully convolutional.

[concrete13377]

sure, I haven't notice that the suggested command have different image as an argument, with this one it works fine, thank you for your work!