ffhq-e cannot be converted from official model but ffhq-f made it

[XD666]

Setting up TensorFlow plugin "fused_bias_act.cu": Preprocessing... Loading... Done.
Setting up TensorFlow plugin "upfirdn_2d.cu": Preprocessing... Loading... Done.
Traceback (most recent call last):
  File "convert_weight.py", line 214, in <module>
    state_dict = fill_statedict(state_dict, g_ema.vars, size)
  File "convert_weight.py", line 148, in fill_statedict
    convert_torgb(vars, f'G_synthesis/{reso}x{reso}/ToRGB', f'to_rgbs.{i}'),
  File "convert_weight.py", line 101, in update
    raise ValueError(f'Shape mismatch: {v.shape} vs {state_dict[k].shape}')
ValueError: Shape mismatch: torch.Size([1, 3, 256, 1, 1]) vs torch.Size([1, 3, 512, 1, 1])
`

[rosinality]

Added --channel_multiplier arguments to the convert_weight.py at c7c8fc7. Now you can use --channel_multiplier 1 options for converting config-e checkpoints.

[songyoyo]

@XD666 Hi, I'm also tring to convert the weight, but some errors ocurred. Can you share your implementation environment or the converted weight file (ffhq-f)? My e-mail is 610240943@qq.com. THANK YOU VERY MUCH!

[woctezuma]

First, thank you for providing this script to convert weights!

Second, it might be good to mention in the README --channel_multiplier 1 for people using config-e.

I have seen it mentioned in the help of:

!python convert_weight.py -h

but it was not super clear at that time.

[woctezuma]

Similarly, it takes some time to check the code and understand a few things which are not super clear from the command-line help.

usage: convert_weight.py [-h] --repo REPO [--gen] [--disc]
                         [--channel_multiplier CHANNEL_MULTIPLIER]
                         PATH

positional arguments:
  PATH

optional arguments:
  -h, --help            show this help message and exit
  --repo REPO
  --gen
  --disc
  --channel_multiplier CHANNEL_MULTIPLIER

For instance, that the script outputs an image with 3 rows:

• images generated with TensorFlow, PyTorch, and the difference between the two,

stylegan2-pytorch/convert_weight.py

Line 273 in d1ef862

img_concat = torch.cat((img_tf, img_pt, img_diff), dim=0)

or that:

• --repo is used to provide a path to the official StyleGAN2 repository with the dnnlib/ folder,
• PATH is the absolute path to the TensorFlow weights.
• --gen is a boolean flag to convert the generator,
• --disc is a boolean flag to convert the discriminator.

[rosinality]

@woctezuma Yes...I will add some helps to command line arguments. I am quite lazy about that. 😅

[woctezuma]

I can understand that. It is not the most interesting part of the work. However, your converter tool does more things than I would have expected, so I think that providing a bit of clarity helps shine some light on the optional (yet useful) stuff that you did. :)

[rosinality]

@woctezuma Added cli argument helps at f938922.

[woctezuma]

Thank you very much! This will be super helpful!