Improve HF integration

[NielsRogge]

This PR showcases the PyTorchModelHubMixin class which you could use to push and reload the AllTracker models to and from the hub.

Here's the model: https://huggingface.co/nielsr/alltracker.

Usage is as follows:

from nets.net34 import Net

model = Net.from_pretrained("nielsr/alltracker")

Benefits:

• download stats work
• the weights are serialized in the safetensors format (along with a config.json)
• you can tweak the model card, but there's a default one
• each checkpoint gets pushed to a separate model repository (best practice).

Here's the script I used to push one particular checkpoint:

from nets.net34 import Net
from huggingface_hub import hf_hub_download
import torch

model = Net(seqlen=16)
    
# equip the model with the weights from the hub
filepath = hf_hub_download(repo_id="aharley/alltracker", filename="alltracker_tiny.pth")
state_dict = torch.load(filepath, map_location="cpu")["model"]
model.load_state_dict(state_dict)

# push the model to the hub
model.push_to_hub("nielsr/alltracker-tiny")

# load the model from the hub
model = Net.from_pretrained("nielsr/alltracker-tiny")

[aharley]

This is nice, but I will want to use the model from here:
https://huggingface.co/aharley/alltracker/

[NielsRogge]

Ok, in that case you could add a from_pretrained method like this one: https://huggingface.co/docs/huggingface_hub/guides/integrations#frompretrained. It can do the following behind the scenes:

from huggingface_hub import hf_hub_download

filepath = hf_hub_download(repo_id="aharley/alltracker", filename="alltracker.pth")

Would you be able to integrate this yourself or would you like me to open a PR?