missing files for celeba preprocessing

[romain-rsr]

missing files for celeba preprocessing

[romain-rsr]

we were able to download the flower dataset and apply your flower preprocessing file on it, yet when we try to train the file we get this error :

Traceback (most recent call last):
  File "train_gan.py", line 87, in <module>
    generator = model.Generator(dparams).to(device)
  File "x/xingseg/models/generator.py", line 152, in __init__
    if 'celeba' in hyper_paras['class_name'] or 'taichi' in hyper_paras['class_name'] or 'flower' in hyper_paras['class_name'] or 'cub' in hyper_paras['class_name']:
KeyError: 'class_name'

apparently a 'class_name' key was missing from the dict fed into the Generator in train_gan.py :

generator = model.Generator(dparams={   'z_dim': args.z_dim, 
            'n_keypoints': args.n_keypoints, 
            'n_per_kp': args.n_per_kp,
            'n_embedding': args.n_embedding,
            'image_size': args.image_size,
            'feature_map_sizes': args.feature_map_sizes,
            'feature_map_channels': args.feature_map_channels,
            'single_final': args.single_final, 
            'use_linear': args.use_linear,
            'smaller_init_mask': args.smaller_init_mask
}).to(device)

we then added this line in the dict : 'class_name':args.class_name,

but then we got this error :

RuntimeError: Caught RuntimeError in replica 0 on device 0.
Original Traceback (most recent call last):
  File "x/envs/xingseg/lib/python3.8/site-packages/torch/nn/parallel/parallel_apply.py", line 64, in _worker
    output = module(*input, **kwargs)
  File "x/micromamba/envs/xingseg/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl
    return forward_call(*input, **kwargs)
  File "/x/xingseg/models/generator.py", line 298, in forward
    out_batch = self.gen_foreground(out_batch)
  File "/x/xingseg/models/generator.py", line 403, in gen_foreground
    heatmaps = current_kp_mask.unsqueeze(2) * kp_emb.unsqueeze(-1).unsqueeze(-1)
RuntimeError: The size of tensor a (63) must match the size of tensor b (8) at non-singleton dimension 1

Can you please help us to identify the meaning of this size mismatch, and how to resolve it ?

Thanks in advance
Romain

[xingzhehe]

Hi Romain,

Thanks for your interests in our work.

1. I double checked and did find the files in https://github.com/xingzhehe/GANSeg/tree/main/data/celeba_wild_raw/MAFL. Please ping me again if you cannot find them.

2. Thanks for the missing keys! I have added it to the file.

3. As for the mismatch dimensions. I ran the experiments and didn't get any errors. It would be great if you could tell me which hyperparameters you changed. Since pytorch has a weird bug in loading weights, I have to hard code the structure from line 149-225 in models/generator.py as what they were during training, so that it can take the pre-trained weights. I am so sorry for the inconvenience, and wish pytorch to fix that issue soon.

[romain-rsr]

Thanks for your reactivity and these explanations, we were finally able to run the celeba example without any problem.