How to load a torchvision model from disk?

Question

I'm trying to follow the solution from the top answer here to load an object detection model from the .pth file.

os.environ['TORCH_HOME'] = '../input/torchvision-fasterrcnn-resnet-50/' #setting the environment variable
model = detection.fasterrcnn_resnet50_fpn(pretrained=False).to(DEVICE)

I get the following error

NotADirectoryError: [Errno 20] Not a directory: '../input/torchvision-fasterrcnn-resnet-50/fasterrcnn_resnet50_fpn_coco-258fb6c6.pth/hub'

google did not reveal an answer to the error and I don't exactly know what it means except for the obvious (that folder 'hub' is missing).

Do I have to unpack or create a folder? I have tried loading the weights but I get the same error message.

this is how I load the model

model = detection.fasterrcnn_resnet50_fpn(pretrained=True)
checkpoint = torch.load('../input/torchvision-fasterrcnn-resnet-50/model.pth.tar')
model.load_state_dict(checkpoint['state_dict'])

thank you for your help!

Full Error Trace:

gaierror: [Errno -3] Temporary failure in name resolution

During handling of the above exception, another exception occurred:

URLError                                  Traceback (most recent call last)
/tmp/ipykernel_42/1218627017.py in <module>
      1 # to load
----> 2 model = detection.fasterrcnn_resnet50_fpn(pretrained=True)
      3 checkpoint = torch.load('../input/torchvision-fasterrcnn-resnet-50/model.pth.tar')
      4 model.load_state_dict(checkpoint['state_dict'])

/opt/conda/lib/python3.7/site-packages/torchvision/models/detection/faster_rcnn.py in fasterrcnn_resnet50_fpn(pretrained, progress, num_classes, pretrained_backbone, trainable_backbone_layers, **kwargs)
    360     if pretrained:
    361         state_dict = load_state_dict_from_url(model_urls['fasterrcnn_resnet50_fpn_coco'],
--> 362                                               progress=progress)
    363         model.load_state_dict(state_dict)
    364     return model

/opt/conda/lib/python3.7/site-packages/torch/hub.py in load_state_dict_from_url(url, model_dir, map_location, progress, check_hash, file_name)
    553             r = HASH_REGEX.search(filename)  # r is Optional[Match[str]]
    554             hash_prefix = r.group(1) if r else None
--> 555         download_url_to_file(url, cached_file, hash_prefix, progress=progress)
    556 
    557     if _is_legacy_zip_format(cached_file):

/opt/conda/lib/python3.7/site-packages/torch/hub.py in download_url_to_file(url, dst, hash_prefix, progress)
    423     # certificates in older Python
    424     req = Request(url, headers={"User-Agent": "torch.hub"})
--> 425     u = urlopen(req)
    426     meta = u.info()
    427     if hasattr(meta, 'getheaders'):

/opt/conda/lib/python3.7/urllib/request.py in urlopen(url, data, timeout, cafile, capath, cadefault, context)
    220     else:
    221         opener = _opener
--> 222     return opener.open(url, data, timeout)
    223 
    224 def install_opener(opener):

/opt/conda/lib/python3.7/urllib/request.py in open(self, fullurl, data, timeout)
    523             req = meth(req)
    524 
--> 525         response = self._open(req, data)
    526 
    527         # post-process response

/opt/conda/lib/python3.7/urllib/request.py in _open(self, req, data)
    541         protocol = req.type
    542         result = self._call_chain(self.handle_open, protocol, protocol +
--> 543                                   '_open', req)
    544         if result:
    545             return result

/opt/conda/lib/python3.7/urllib/request.py in _call_chain(self, chain, kind, meth_name, *args)
    501         for handler in handlers:
    502             func = getattr(handler, meth_name)
--> 503             result = func(*args)
    504             if result is not None:
    505                 return result

/opt/conda/lib/python3.7/urllib/request.py in https_open(self, req)
   1391         def https_open(self, req):
   1392             return self.do_open(http.client.HTTPSConnection, req,
-> 1393                 context=self._context, check_hostname=self._check_hostname)
   1394 
   1395         https_request = AbstractHTTPHandler.do_request_

/opt/conda/lib/python3.7/urllib/request.py in do_open(self, http_class, req, **http_conn_args)
   1350                           encode_chunked=req.has_header('Transfer-encoding'))
   1351             except OSError as err: # timeout error
-> 1352                 raise URLError(err)
   1353             r = h.getresponse()
   1354         except:

URLError: <urlopen error [Errno -3] Temporary failure in name resolution>

Answer 1

If you are loading a pretrained network, you don't need to load the model from torchvision pretrained (as in pretrained by torchvision on ImageNet using pretrained=True). You have two options:

1. Either set pretrained=False and load you weights using:

   checkpoint = torch.load('../input/torchvision-fasterrcnn-resnet-50/model.pth.tar')
   model.load_state_dict(checkpoint['state_dict'])
   

2. Or if you decide to change TORCH_HOME (which is not ideal) you need to keep the same directory structure Torchvision has which would be:

   inputs/hub/checkpoints/fasterrcnn_resnet50_fpn_coco-258fb6c6.pth 
   

   In practice, you wouldn't change TORCH_HOME just to load one model.

Answer 2

I found the solution digging deep into github, to the problem, which is a little hidden.

detection.() has a default argument besides pretrained, it's called pretrained_backbone which by default is set to true, which if True sets the models to download from a dictionary path of urls.

this will work:

detection.fasterrcnn_resnet50_fpn(pretrained=False, pretrained_backbone = False, num_classes = 91).

then load the model as usual. num_classes is expected, in the docs it's a default = 91 but in github i saw it as None, which is why I added it here for saftey.