taesungp
commented
1 year ago The links are missing, so I cannot tell the quality. It sounds like there's some discrepancy between training and testing, like the input image resolution for example. Reducing this gap may help.
Dear authors, thanks for sharing this amazing framework it contains too much parameters its clear that it took a lot of time to develop. Im using pix2pix in watermark removing I created a custom watermark and watermarked a collection of diverse images. after watermarking I resized all the images to multiple specific sizes 10241024,640640,640*480. I also used different watermark weights. each size and watermark weight was tested separately. In the version of bold watermark color (white color) the training results was accaptable but in the inference the model generated extra watermarks not removed the old watermarks. do you suggest Any changes? here is sample of the training and testing command I'm using:
!python train.py --model pix2pix --dataroot /content/gdrive/MyDrive/projects/dataset/rectangle_dataset_v0/dataset --name watermark_remover_pix2pix_rectangle_datasetv1 --direction AtoB --save_epoch_freq 50 --n_epochs 800 --n_epochs_decay 600 --preprocess crop --crop_size 256 --display_id -1!python test.py --dataroot /content/gdrive/MyDrive/pix2pix_official/dataset/dataset/A/test --name watermark_remover_pix2pix_rectangle_datasetv1 --model test --no_dropout --preprocess nonedo you suggest Any changes? Any help is greatly appreciated. @strob @andyli @heaversm @ecoopnet Thanks in advanced.