I tried to transfer weights from custom model, which i trained using STYLEGAN2_ADA tf repo.
However, after I convert weight, and try to load weight with torch.load, I get
RuntimeError: Invalid magic number; corrupt file?
To Reproduce
Steps to reproduce the behavior:
Disclaimer : I am on docker enviornment
Expected behavior
I understand loading should be done, so i am guessing whether custom weights cannot be transferred.
I have done other manipulation(ex. projecting query image using projector.py in TF repo), so weight itself works on TF enviornment
Screenshots
If applicable, add screenshots to help explain your problem.
Desktop (please complete the following information):
OS: Linux Ubuntu 1.8.0
PyTorch version : torch 1.8.0
CUDA toolkit version : CUDA 11.2
NVIDIA driver version : 460.32.03
GPU : Irrelevant as I'm not using GPU, but Quadro RTX 8000
Docker: Yes, pytorch/pytorch:1.7.0-cuda11.0-cudnn8-runtime
(Then i upgraded torch to 1.8.0 as this repo requires torch 1.7.1 and over)
Inspecting legacy.py showed it was written to convert weights downloaded from URL.
I'm searching my ways to convert existing weight, but at least the original code is meant to not convert custom weights
Describe the bug
I tried to transfer weights from custom model, which i trained using STYLEGAN2_ADA tf repo. However, after I convert weight, and try to load weight with torch.load, I get RuntimeError: Invalid magic number; corrupt file?
To Reproduce Steps to reproduce the behavior: Disclaimer : I am on docker enviornment
Then with transferred pkl, I tried to load weight
Then, I got error below
Expected behavior I understand loading should be done, so i am guessing whether custom weights cannot be transferred. I have done other manipulation(ex. projecting query image using projector.py in TF repo), so weight itself works on TF enviornment Screenshots If applicable, add screenshots to help explain your problem.
Desktop (please complete the following information):