Fail to load network .pkl file

[Baoren1996]

I tried to download the saved network from the link below:
https://zenodo.org/record/10947091/files/stylesan-xl_cifar10.pkl
however, when I tried to use the following code for evaluation:

python calc_metrics.py --metrics=fid50k_full --network=./stylesan-xl_cifar10.pkl

but got error:

Python==3.8.10
Pytorch==2.4.0+cu121

Can somebody help?

[TakashiShibuyaSony]

Hello, thank you for reaching out to us!

Our recommendation is preparing a Docker container, following the README.md.
https://github.com/sony/san/tree/main/stylesan-xl#requirements

If you don't want to rely on Docker for some reason, it should be good to use the same versions of packages described in the following files.

san/stylesan-xl/requirements.txt

Lines 1 to 20 in 0e52b14

	numpy>=1.20
	click>=8.0
	pillow>=8.3.1
	scipy>=1.7.1
	requests>=2.26.0
	tqdm>=4.62.2
	ninja>=1.10.2
	matplotlib>=3.4.2
	imageio>=2.9.0
	dill>=0.3.4
	psutil>=5.8.0
	regex>=2022.3.15
	pillow>=8.3.1
	imgui>=1.3.0
	glfw>=2.2.0
	pyopengl>=3.1.5
	imageio-ffmpeg>=0.4.3
	pyspng
	ftfy>=6.1.1
	timm==0.4.12

san/stylesan-xl/Dockerfile

Line 5 in 0e52b14

RUN pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu121

This may work.

Thank you.

[Baoren1996]

Hello, thank you for reaching out to us!

Our recommendation is preparing a Docker container, following the README.md. https://github.com/sony/san/tree/main/stylesan-xl#requirements

If you don't want to rely on Docker for some reason, it should be good to use the same versions of packages described in the following files.

san/stylesan-xl/requirements.txt

Lines 1 to 20 in 0e52b14

	numpy>=1.20
	click>=8.0
	pillow>=8.3.1
	scipy>=1.7.1
	requests>=2.26.0
	tqdm>=4.62.2
	ninja>=1.10.2
	matplotlib>=3.4.2
	imageio>=2.9.0
	dill>=0.3.4
	psutil>=5.8.0
	regex>=2022.3.15
	pillow>=8.3.1
	imgui>=1.3.0
	glfw>=2.2.0
	pyopengl>=3.1.5
	imageio-ffmpeg>=0.4.3
	pyspng
	ftfy>=6.1.1
	timm==0.4.12

san/stylesan-xl/Dockerfile

Line 5 in 0e52b14

RUN pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu121

This may work.
Thank you.

Hi @TakashiShibuyaSony, thanks for your instruction which works perfectly for me.

Another question, given the CIFAR model snapshot, is there any way to resume the training from the pkl file? do I also need the stem network .pkl file?

Thanks a lot in advance.

Best,
Baoren

[TakashiShibuyaSony]

Hi @Baoren1996,
You can resume the training by using the following option. You should provide the path of the .pkl file.

san/stylesan-xl/train.py

Line 144 in 0e52b14

@click.option('--resume', help='Resume from given network pickle', metavar='[PATH|URL]', type=str)

Regarding the .pkl file of the stem network, the current code requires it (Our codebase build on that of StyleGAN-XL, and their code requires a stem network even in resuming training). But, I guess you can resume the training without the stem network by reloading the ckpt for resume in the following lines again.

san/stylesan-xl/training/networks_stylegan3_resetting.py

Lines 613 to 614 in 0e52b14

	with dnnlib.util.open_url(path_stem) as f:
	G_stem = legacy.load_network_pkl(f)['G_ema']

And, the succeeding lines should also be modified. In particular, the following lines are not necessary in this case because these lines are for appending new layers in progressive training.

san/stylesan-xl/training/networks_stylegan3_resetting.py

Lines 651 to 673 in 0e52b14

	# initialize new_layers
	last_stem_layer = getattr(self.synthesis, self.synthesis.layer_names[-1])
	fparams = self.compute_superres_filterparams(up_factor, self.img_resolution, last_stem_layer, head_layers)
	self.head_layer_names = []
	for idx in range(head_layers + 1):
	prev = max(idx - 1, 0)
	is_torgb = (idx == head_layers)
	is_critically_sampled = (idx >= head_layers - self.num_critical)
	use_fp16 = (fparams.sampling_rates[idx] * (2 ** self.num_fp16_res) > self.img_resolution)
	layer = SynthesisLayer(
	w_dim=self.w_dim, is_torgb=is_torgb, is_critically_sampled=is_critically_sampled, use_fp16=use_fp16,
	in_channels=int(fparams.channels[prev]), out_channels= int(fparams.channels[idx]),
	in_size=int(fparams.sizes[prev]), out_size=int(fparams.sizes[idx]),
	in_sampling_rate=int(fparams.sampling_rates[prev]), out_sampling_rate=int(fparams.sampling_rates[idx]),
	in_cutoff=fparams.cutoffs[prev], out_cutoff=fparams.cutoffs[idx],
	in_half_width=fparams.half_widths[prev], out_half_width=fparams.half_widths[idx],
	conv_kernel=self.conv_kernel, use_radial_filters=self.use_radial_filters,
	)
	name = f'L{idx+stem_len}_{layer.out_size[0]}_{layer.out_channels}'
	setattr(self.synthesis, name, layer)
	self.synthesis.layer_names.append(name)
	self.head_layer_names.append(name)

Thanks!

[Cwyxx]

Update dill to 0.3.9. It works for me.

[TakashiShibuyaSony]

Hi @Cwyxx, Thank you very much for your helpful information. We updated requirements.txt