RuntimeError: PytorchStreamReader failed reading zip archive: failed finding central directory

Question

When I convert a my trained pytorch model to coreml model, I got this error:

File "/Users/lion/Documents/MyLab/web_workspace/sky_replacement/venv/lib/python3.9/site-packages/torch/jit/_serialization.py", line 161, in load
cpp_module = torch._C.import_ir_module(cu, str(f), map_location, _extra_files)

RuntimeError: PytorchStreamReader failed reading zip archive: failed finding central directory

This is my code:

from networks import *
import coremltools as ct

run_device = torch.device("cpu")
net_G = define_G(input_nc=3, output_nc=1, ngf=64, 
netG='coord_resnet50').to(run_device)
checkpoint = torch.load('./model/best_ckpt.pt', map_location=run_device)
net_G.load_state_dict(checkpoint['model_G_state_dict'])
net_G.to(run_device)
net_G.eval()

model = ct.convert('./model/best_ckpt.pt', source='pytorch', inputs=[ct.ImageType()], skip_model_load=True)
model.save("result.mlmodel")

Answer 1

I had this issue because a failed git LFS smudge corrupted the checkpoint. Check the file size/checksum of your ckpt/pth file.

Answer 2

It could be a problem with the PyTorch version and saving mechanism. I had the same problem and solved it by passing the kwarg _use_new_zipfile_serialization=False when saving the model. More details here.