pkl to pt conversion

[vedantdere]

How to convert pkl file to pt ?

[woctezuma]

stylegan3-editing/models/stylegan3/legacy.py

Lines 300 to 301 in e2a63fc

	def convert_network_pickle(source, dest, force_fp16):
	"""Convert legacy network pickle into the native PyTorch format.

[woctezuma]

The function does not support StyleGAN3 checkpoints, I believe.

stylegan3-editing/models/stylegan3/legacy.py

Lines 303 to 304 in e2a63fc

	The tool is able to load the main network configurations exported using the TensorFlow version of StyleGAN2 or StyleGAN2-ADA.
	It does not support e.g. StyleGAN2-ADA comparison methods, StyleGAN2 configs A-D, or StyleGAN1 networks.

[hqnicolas]

How did you create the weights .PT Files For StyleGan3?
https://github.com/yuval-alaluf/stylegan3-editing#stylegan3-encoder
testing rosinality pkl to pt

[yuval-alaluf]

I believe this is what you guys want:

path = "/path/to/your/sg3.pkl"

with open(path, "rb") as f:
    decoder = pickle.load(f)['G_ema'].cuda()

state_dict = decoder.state_dict()
torch.save(state_dict, "/path/to/your/sg3.pt") 

Let me know if this works as I haven't tested it out yet

[yuval-alaluf]

@hotnikq , I'm not sure I follow here.
Please open a new issue if you are having a problem running something. And please provide a clear, detailed summary of what you're trying to run.

[yuval-alaluf]

@hotnikq , you're saying the code I posted above did not work for you on the stylegan3-r-ffhqu-256x256.pkl file?

[yuval-alaluf]

How did you create the weights .PT Files For StyleGan3?

I have answered the question above and provided the code that we used to convert pkl files to pt files. I am really not sure why you are still trying to use the legacy.py file after it was noted above that this file is not relevant for SG3 models.
And @hotnikq , it is quite difficult to follow these threads with so many comments. Please try to limit your comments when possible to make it easier for myself and others to follow these threads.

[hqnicolas]

I believe this is what you guys want:

path = "/path/to/your/sg3.pkl"

with open(path, "rb") as f:
    decoder = pickle.load(f)['G_ema'].cuda()

state_dict = decoder.state_dict()
torch.save(state_dict, "/path/to/your/sg3.pt") 

Let me know if this works as I haven't tested it out yet

Testing now !

[hqnicolas]

import pickle
from enum import Enum
from pathlib import Path
from typing import Optional

import torch

checkpoint_path = "pretrained_models/stylegan3-r-ffhqu-256x256.pkl"

print(f"Loading StyleGAN3 generator from path: {checkpoint_path}")
with open(checkpoint_path, "rb") as f:
	decoder = pickle.load(f)['G_ema'].cuda()
print('Done!')
       
state_dict = decoder.state_dict()
torch.save(state_dict, "pretrained_models/stylegan3-r-ffhqu-256x256.pt")

print('Done!')

This code works just fine the topic ends here!

now i need to adapt the generator model to math the checkpoint
RuntimeError: Error(s) in loading state_dict for Generator:
size mismatch for synthesis.L3_52_1024.up_filter: copying a param with shape torch.Size([24]) from checkpoint, the shape in current model is torch.Size([12]).

[yuval-alaluf]

I am happy to see that the code I gave you above works fine and that we can now close this issue.
If you have an issue that is unrelated to this, please open a new issue with a concise yet clear explanation of the problem you are facing