how to use the pth.tar file obtained?

[keithyh]

On Ubuntu 16.04,I trained my own data and got .pth.tar files in output folder.Each tar file is 1.1GB,but if I click into one tar file I found that there is an "archive" folder with size 0 bytes.In the "archive" folder there is no .pth file.Therefore I wonder how to use these .pth.tar files.

Please could you tell me.Thank you!

[linhduongtuan]

First, you should use/run clean_checkpoint.py which leads to obtain the model_best_SHA.pth checkpoint file, and then you can use the file like at a hub checkpoint. For example, model = create_model(arg.model, pretrained=arg.pretrained, num_classes=arg.num_classes, checkpoint=arg.checkpoint) #link to your best checkpoint above. Good luck Linh

…

On 30 Dec 2020, at 18:40, mrkind ***@***.***> wrote:  On Ubuntu 16.04,I trained my own data and got .pth.tar files in output folder.Each tar file is 1.1GB,but if I click into one tar file I found that there is an "archive" folder with size 0 bytes.In the "archive" folder there is no .pth file.Therefore I wonder how to use these .pth.tar files. Please could you tell me.Thank you! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[keithyh]

Thank you,Linh!That sounds clear.

[keithyh]

Hi,everyone,I tried to treate .pth.tar file or the .pkl file obtained from extracting the .pth.tar file as the checkpoint,but failed.
That is, both commands below have been tried but both return error message

python clean_checkpoint.py --checkpoint /home/skjmom/timm_pytorch_im_model/pytorch-image-models-master/output/train/checkpoint-9.pth.tar --output ./output/trained.pth --use-ema --clean-aux-bn

python clean_checkpoint.py --checkpoint /home/skjmom/timm_pytorch_im_model/pytorch-image-models-master/output/train/20201228-022118-efficientnet_b7-600/archive/data.pkl --output ./output/trained.pth --use-ema --clean-aux-bn

Both error messages are Unpickling error,one of them looks like:

Traceback (most recent call last):
File "clean_checkpoint.py", line 82, in
main()
File "clean_checkpoint.py", line 40, in main
checkpoint = torch.load(args.checkpoint, map_location='cpu')
File "/home/skjmom/anaconda3/lib/python3.7/site-packages/torch/serialization.py", line 358, in load
return _load(f, map_location, pickle_module)
File "/home/skjmom/anaconda3/lib/python3.7/site-packages/torch/serialization.py", line 532, in _load
magic_number = pickle_module.load(f)
_pickle.UnpicklingError: A load persistent id instruction was encountered,
but no persistent_load function was specified.

[rwightman]

@mrkind you might be switching back and forth between pytorch 1.6/1.7 and older ones, pytorch changed the checkpoint format in newer versions, which causes an 'archive' error if you load with < 1.6. To use with older ones you need to save the checkpoint with an extra arg. If you run the checkpoint clean script in the newer pytorch version, it will output a cleaned checkpoint in the older format that is loadable by Pytorch < 1.6 and newer. Otherwise @linhduongtuan comment's are correct, the clean script is useful to strip out the extra train specific parts of the training checkpoint for hosting/using the model.

[keithyh]

Thank you author,I will have a try later today~I did install pytorch 1.7 in a virtual environment and pytorch 0.4 in the base environment

[keithyh]

Much thanks for you two!I successfully obtained trained_SHA.pth