anvoynov / biggansarewatching Goto Github PK

Hi, I really like your BIgBiGAN2BigGAN code and it's super useful!
However running the code on Windows system will generate the error below, and after quick debugging, I find it's the path format is unrecognizable to hdf5.
'''
KeyError: "Unable to open object (object 'module\Generator\GBlock\conv0\u0:0' doesn't exist)"'
'''
And after using / as separator instead of \\ everything works like charm!
So I suggest a quick and dirty solution to enhance compatibility is to add .replace("\\", "/") to L149, L175, L189 in tf_to_torch

Hi, remarkable work! Congratulations! I am trying to reproduce the segmentation results on CUB dataset. However, I got performance significantly different what is claimed in Table 1. Could you please help me to see if I did something wrong? Or if this result is within the normal range of variation? Many Thanks!

Reproduce results

Commands

BigBiGAN
python train_segmentation.py --out=results/CUB_BigBiGAN --gen_devices=1 --gan_weights=BigGAN/weights/BigBiGAN_x1.pth --z=embeddings/BigBiGAN_ImageNet_z.npy --bg_direction=BigGAN/weights/bg_direction.pth --val_images_dir=data/CUB/test_images --val_masks_dir=data/CUB/test_segmentations

E-BigBiGAN (w/o z-noising)
python train_segmentation.py --out=results/CUB_E_BigBiGAN --gan_weights=BigGAN/weights/BigBiGAN_x1.pth --z=embeddings/BigBiGAN_CUB_train_z.npy --bg_direction=BigGAN/weights/bg_direction.pth --val_images_dir=data/CUB/test_images --val_masks_dir=data/CUB/test_segmentations

E-BigBiGAN (w/ z-noising)
python train_segmentation.py --out=results/CUB_E_BigBiGAN_noise --gen_devices=1 --gan_weights=BigGAN/weights/BigBiGAN_x1.pth --z=embeddings/BigBiGAN_CUB_train_z.npy --z_noise=0.2 --bg_direction=BigGAN/weights/bg_direction.pth --val_images_dir=data/CUB/test_images --val_masks_dir=data/CUB/test_segmentations

For the val_images_dir and val_masks_dir, I just copied all the test images and segmentations to new folders, where test split follows ReDO (Chen et al., 2019) setting.

Environment

Ubuntu 18.04.4 LTS
GeForce GTX TITAN X
torch==1.3.0
torchvision==0.4.1a0+d94043a
scipy==1.3.1
numpy==1.17.3
Pillow==6.2.1
h5py==2.9.0
scikit_image==0.15.0