[discussion] Recommend a different file extension for models (.PTH is a special extension for Python)

[vadimkantorov]

*.pth files are used by Python to list additional package search paths: https://docs.python.org/3/library/site.html

The pth files will be loaded as text files by Python interpreter. At some point when I had some PyTorch model pth file placed along with the sources, it caused a hang of Python at startup (it was trying to parse the big binary file as a list of paths).

Maybe just *.pt?

[t-vi]

As far as I can tell, .pt is used in many bits anyway, e.g. https://pytorch.org/tutorials/advanced/cpp_export.html , even if I have seen .pth (or even .pth.tar, when it wasn't a tar) in the wild.
But yes, I agree that standardizing on something not colliding with basic Python functionality is a good thing.

[vadimkantorov]

I think the downloadable torchvision models have pth extension

[vadimkantorov]

e.g. Intel's distiller uses the strange .pth.tar as well: https://nervanasystems.github.io/distiller/model_zoo/index.html

[vadimkantorov]

I think this is especially pertinent with a new announcement of Torch Hub:

https://pytorch.org/docs/master/hub.html and https://github.com/pytorch/vision/blob/master/hubconf.py both mention *.pth files

@soumith

[soumith]

sure, we can change our models to .pt, I have no reservations.
Do you might sending PRs, or pointing out where all you noticed the .pth recommendations so that we can change them?

[vadimkantorov]

@soumith Sure! I'll find all occurences and paste the pointers here :)

[vadimkantorov]

An incomplete yet list (so far searched on github pytorch, torchvision, examples):

1. pytorch/test/onnx/test_pytorch_onnx_caffe2.py

   Lines 96 to 108 in 6fccca4

   	model_urls = {
   	'alexnet': 'https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth',
   	'dcgan_b': 'https://s3.amazonaws.com/pytorch/test_data/export/netG_bedroom_epoch_1-0649e76b.pth',
   	'dcgan_f': 'https://s3.amazonaws.com/pytorch/test_data/export/netG_faces_epoch_49-d86035a6.pth',
   	'densenet121': 'https://download.pytorch.org/models/densenet121-d66d3027.pth',
   	'inception_v3_google': 'https://download.pytorch.org/models/inception_v3_google-1a9a5a14.pth',
   	'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
   	'srresNet': 'https://s3.amazonaws.com/pytorch/demos/srresnet-e10b2039.pth',
   	'super_resolution': 'https://s3.amazonaws.com/pytorch/test_data/export/superres_epoch100-44c6958e.pth',
   	'squeezenet1_0': 'https://download.pytorch.org/models/squeezenet1_0-a815701f.pth',
   	'squeezenet1_1': 'https://download.pytorch.org/models/squeezenet1_1-f364aa15.pth',
   	'vgg16': 'https://download.pytorch.org/models/vgg16-397923af.pth',
   	'vgg19': 'https://download.pytorch.org/models/vgg19-dcbb9e9d.pth',

2. pytorch/torch/utils/model_zoo.py

   Lines 28 to 52 in c47f680

   	# matches bfd8deac from resnet18-bfd8deac.pth
   	HASH_REGEX = re.compile(r'-([a-f0-9]*)\.')
   	def load_url(url, model_dir=None, map_location=None, progress=True):
   	r"""Loads the Torch serialized object at the given URL.
   	If the object is already present in `model_dir`, it's deserialized and
   	returned. The filename part of the URL should follow the naming convention
   	``filename-<sha256>.ext`` where ``<sha256>`` is the first eight or more
   	digits of the SHA256 hash of the contents of the file. The hash is used to
   	ensure unique names and to verify the contents of the file.
   	The default value of `model_dir` is ``$TORCH_HOME/models`` where
   	``$TORCH_HOME`` defaults to ``~/.torch``. The default directory can be
   	overridden with the ``$TORCH_MODEL_ZOO`` environment variable.
   	Args:
   	url (string): URL of the object to download
   	model_dir (string, optional): directory in which to save the object
   	map_location (optional): a function or a dict specifying how to remap storage locations (see torch.load)
   	progress (bool, optional): whether or not to display a progress bar to stderr
   	Example:
   	>>> state_dict = torch.utils.model_zoo.load_url('https://s3.amazonaws.com/pytorch/models/resnet18-5c106cde.pth')

3. https://github.com/pytorch/pytorch/blob/5734e9677564743fc4000cfb955fb42046689be9/docs/source/hub.rst

4. https://github.com/pytorch/vision/blob/8f943d4e0c380cb0a5587b6e0e032932576fabea/torchvision/models/vgg.py#L12-L19

5. https://github.com/pytorch/vision/blob/71182bc1ea27652f9952f6d60d8b27e408fc940e/torchvision/models/resnet.py#L10-L14

6. https://github.com/pytorch/vision/blob/c7e9bd3006b0144fd1a94724f08122f673fe3587/hubconf.py#L48-L62

7. https://github.com/pytorch/vision/blob/d5637696eba298f96a5fda44c6462f97ad1f987c/torchvision/models/densenet.py#L12-L15

8. https://github.com/pytorch/vision/blob/dc0238b82f0df5c44ec9878cb41011d1852a7afd/torchvision/models/squeezenet.py#L11-L12

9. https://github.com/pytorch/vision/blob/1fb0ccf71620d113cb72696b2eb8317b3e252cbb/torchvision/models/alexnet.py#L9

10. https://github.com/pytorch/vision/blob/85369e3a315697be7e167f303d44f6b69d46c8ee/torchvision/models/inception.py#L12

11. https://github.com/pytorch/examples/blob/29c2ed8ca6dc36fc78a3e74a5908615619987863/dcgan/README.md

12. https://github.com/pytorch/examples/blob/29c2ed8ca6dc36fc78a3e74a5908615619987863/super_resolution/README.md

13. https://github.com/pytorch/examples/blob/2fc0211d30b808f049ab7e7f4990858cf2ac471f/fast_neural_style/neural_style/neural_style.py#L107-L217

14. https://github.com/pytorch/examples/blob/64f829ce495dad43392451c7431ae26eeee39bad/dcgan/main.py#L260-L261

15. https://github.com/pytorch/examples/blob/29c2ed8ca6dc36fc78a3e74a5908615619987863/super_resolution/main.py#L75

16. https://github.com/pytorch/examples/blob/29c2ed8ca6dc36fc78a3e74a5908615619987863/fast_neural_style/README.md

17. https://github.com/pytorch/examples/blob/15e27719d75e35358555a27215665c797999740f/imagenet/main.py#L349-L352

[vadimkantorov]

If *.pt is reserved for zipballs from saved JIT'ted models, it may be needed to recommend a different extension for raw saved tensors (preferably not *.pth or fake *.tar)

[nzmora]

Hi,
Any updates on this? It's a rather trivial issue, but it would be nice to have a "standard" and meaningful file extension for the PyTorch checkpoint files.
Thanks!

[soumith]

we can go with *.ptc. We haven't had time to actually do the task though.

[vadimkantorov]

@soumith *.ptc for both pickle format (from torch.save and state_dict) and zip format from JIT?

[soumith]

maybe .pt for pickle format and .ptc (pytorch compiled) for JIT

[vadimkantorov]

One alternative more verbose option: *.torch.pkl, *.torch.zip, *.torch.h5

[soumith]

i think that's too long

[vadimkantorov]

@soumith Another option: *.pt and *.ptz (hints that it is a collection of multiple things, like npz).

[ain-soph]

Hi, any update on this?
My current library still uses .pth to save models and .pt to save tensors. Let me know the standard if it's finally determined, so that I could apply it on my library.

I don't recommend .ptz if we don't have a torch.savez function and the same style as numpy np.savez(file_path, key1=value1,key2=value2). Do we plan to have it?

[KOLANICH]

👍 for chained extensions not concealing the underlying format. Just .zip and .tar conveys not enough info about what is inside them (it can be pickle, for me pickle === "I cannot accept that").