I get the following errors in logs upon submitting baseline submission
Could not load model. <urlopen error [Errno -3] Temporary failure in name resolution>
raise URLError(err)
urllib.error.URLError: <urlopen error [Errno -3] Temporary failure in name resolution>
any idea on how to resolve? thank you for the help
Hi @polymathAB, thanks for that, we’re actually working on a quick fix that will run without internet. Stay tuned.
Okay thanks for the help. Any updates?
I am facing the same issue in submission. Please help. Thanks.
Yes, we will most likely release an update tomorrow to help people that are attempting to copy the benchmark for actual submissions avoid this issue.
@polymathAB and @uamo Not sure if there’s a version mismatch at play here but I’m having trouble reproducing this error. Can you share your pytorch model .pt files from the benchmark notebook and post a pip freeze to show the versions you have been using?
I am using the following versions on my local -
torch==1.9.0+cu102
pytorch-lightning==1.4.2
segmentation_models_pytorch==0.2.0
I have trained an efficientnet model on my local and saved as ckpt file. The error Traceback shows -
File “/srv/conda/envs/condaenv/lib/python3.9/site-packages/segmentation_models_pytorch/unet/model.py”, line 65, in init
self.encoder = get_encoder(
File “/srv/conda/envs/condaenv/lib/python3.9/site-packages/segmentation_models_pytorch/encoders/init.py”, line 62, in get_encoder
encoder.load_state_dict(model_zoo.load_url(settings[“url”]))
File “/srv/conda/envs/condaenv/lib/python3.9/site-packages/torch/hub.py”, line 524, in load_state_dict_from_url
download_url_to_file(url, cached_file, hash_prefix, progress=progress)
File “/srv/conda/envs/condaenv/lib/python3.9/site-packages/torch/hub.py”, line 394, in download_url_to_file
u = urlopen(req)
File “/srv/conda/envs/condaenv/lib/python3.9/urllib/request.py”, line 214, in urlopen
return opener.open(url, data, timeout)
File “/srv/conda/envs/condaenv/lib/python3.9/urllib/request.py”, line 517, in open
response = self._open(req, data)
File “/srv/conda/envs/condaenv/lib/python3.9/urllib/request.py”, line 534, in _open
result = self._call_chain(self.handle_open, protocol, protocol +
File “/srv/conda/envs/condaenv/lib/python3.9/urllib/request.py”, line 494, in _call_chain
result = func(*args)
File “/srv/conda/envs/condaenv/lib/python3.9/urllib/request.py”, line 1389, in https_open
return self.do_open(http.client.HTTPSConnection, req,
File “/srv/conda/envs/condaenv/lib/python3.9/urllib/request.py”, line 1349, in do_open
raise URLError(err)
urllib.error.URLError: <urlopen error [Errno -3] Temporary failure in name resolution>
ERROR conda.cli.main_run:execute(33): Subprocess for ‘conda run [‘python’, ‘main.py’]’ command failed. (See above for error)
I am also using same versions of the packages as above ^. Also in the baseline notebook, it has been instructed that a checkpoint file has to be saved in the assets sub-dir. In the utility code provided too, it used the PyTorch lighting module’s load_from_checkpoint function.
I suspected that the issue was with using a pretrained resnet34 backbone no? is there any way to include the resnet weights (.pth file) in the assets dir so that we don’t have to download at all? I was looking into pytorch lightning loading pre-trained weights code and it always uses internet for downloading them as far as I could explore
EDIT: my logs for reference :-
EDIT 2: Issue was solved after putting resnet34 file in assets and copying to container in submission. However score received is 0. Is this right? Since validation IoU was >0.44
EDIT 3: According to my logs, now the test_features directory was not read by the main.py at all. Maybe it is a bug in the path definitions in flood_model.py and main.py provided in baselines? Please help as to a resolution
@polymathAB @uamo Please see here for a fully worked example: https://github.com/drivendataorg/floodwater-runtime/tree/main/benchmark-pytorch
We fixed a couple other issues along the way and the blog will be updated within the next few minutes. Thanks for bearing with us on this - there were a lot of layers to dig through in the pytorch hub and segmentation models code to figure out the best way to prevent a download.
thanks a lot for clearing it all up. I think this will make it easier to implement other models for us as competitors too 