pretrained densenet does not work on CPU

[lopuhin]

When I try to initialise pretrained densenet121, I get the error AssertionError: Torch not compiled with CUDA enabled (while other networks work). Using http://download.pytorch.org/whl/torch-0.2.0.post1-cp36-cp36m-macosx_10_7_x86_64.whl and torchvision-0.1.9 on CPython 3.6.2 on OS X 10.12.6.

Here is the full traceback:

In [2]: for m in [resnet50, inception_v3, densenet121]:
   ...:     m(pretrained=True)
   ...:
venv/lib/python3.6/site-packages/torchvision/models/inception.py:65: UserWarning: src is not broadcastable to dst, but they have the same number of elements.  Falling back to deprecated pointwise behavior.
  m.weight.data.copy_(values)
Downloading: "https://download.pytorch.org/models/densenet121-241335ed.pth" to /Users/kostia/.torch/models/densenet121-241335ed.pth
100.0%
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-2-28619da1d19b> in <module>()
      1 for m in [resnet50, inception_v3, densenet121]:
----> 2     m(pretrained=True)
      3

venv/lib/python3.6/site-packages/torchvision/models/densenet.py in densenet121(pretrained, **kwargs)
     26                      **kwargs)
     27     if pretrained:
---> 28         model.load_state_dict(model_zoo.load_url(model_urls['densenet121']))
     29     return model
     30

venv/lib/python3.6/site-packages/torch/utils/model_zoo.py in load_url(url, model_dir, map_location)
     56         hash_prefix = HASH_REGEX.search(filename).group(1)
     57         _download_url_to_file(url, cached_file, hash_prefix)
---> 58     return torch.load(cached_file, map_location=map_location)
     59
     60

venv/lib/python3.6/site-packages/torch/serialization.py in load(f, map_location, pickle_module)
    229         f = open(f, 'rb')
    230     try:
--> 231         return _load(f, map_location, pickle_module)
    232     finally:
    233         if new_fd:

venv/lib/python3.6/site-packages/torch/serialization.py in _load(f, map_location, pickle_module)
    377     unpickler = pickle_module.Unpickler(f)
    378     unpickler.persistent_load = persistent_load
--> 379     result = unpickler.load()
    380
    381     deserialized_storage_keys = pickle_module.load(f)

venv/lib/python3.6/site-packages/torch/cuda/__init__.py in _lazy_new(cls, *args, **kwargs)
    264 @staticmethod
    265 def _lazy_new(cls, *args, **kwargs):
--> 266     _lazy_init()
    267     # We need this method only for lazy init, so we can remove it
    268     del _CudaBase.__new__

venv/lib/python3.6/site-packages/torch/cuda/__init__.py in _lazy_init()
     82         raise RuntimeError(
     83             "Cannot re-initialize CUDA in forked subprocess. " + msg)
---> 84     _check_driver()
     85     torch._C._cuda_init()
     86     torch._C._cuda_sparse_init()

venv/lib/python3.6/site-packages/torch/cuda/__init__.py in _check_driver()
     49 def _check_driver():
     50     if not hasattr(torch._C, '_cuda_isDriverSufficient'):
---> 51         raise AssertionError("Torch not compiled with CUDA enabled")
     52     if not torch._C._cuda_isDriverSufficient():
     53         if torch._C._cuda_getDriverVersion() == 0:

AssertionError: Torch not compiled with CUDA enabled

[alykhantejani]

I think the files we're serialized with the old method and now can't be properly read with the new serialization. I've tried loading, saving and re-loading on a CUDA enabled machine - which seems to work.

I will try to upload these new files somewhere sensible tomorrow so that the pre-trained locations can be updated.

cc: @fmassa

[alykhantejani]

I will re-generate the models with the new serialization today and then update the URLs in torchvision.

[soumith]

fixed in master by @alykhantejani

[hassanshallal]

Hi Folks, I am facing the same exact issue/error whenever I try to load any densenet modeL. I don't have any issue loading any other pre-trained model. I tried to download the densenet models manually from 489d6e7 and load them using torch.load, but I still get the error: AssertionError: (AssertionError('Torch not compiled with CUDA enabled',), <class 'torch.cuda.FloatTensor'>, ()). I am using macOS Sierra version 10.12.6.

I would really appreciate any help resolving this issue. Here is the full traceback:

`---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
in ()
----> 1 torch.load('densenet121-241335ed.pth')

/Users/hassanshallal/anaconda/lib/python2.7/site-packages/torch/serialization.pyc in load(f, map_location, pickle_module)
229 f = open(f, 'rb')
230 try:
--> 231 return _load(f, map_location, pickle_module)
232 finally:
233 if new_fd:

/Users/hassanshallal/anaconda/lib/python2.7/site-packages/torch/serialization.pyc in _load(f, map_location, pickle_module)
377 unpickler = pickle_module.Unpickler(f)
378 unpickler.persistent_load = persistent_load
--> 379 result = unpickler.load()
380
381 deserialized_storage_keys = pickle_module.load(f)

/Users/hassanshallal/anaconda/lib/python2.7/site-packages/torch/cuda/init.pyc in _lazy_new(cls, *args, **kwargs)
264 @staticmethod
265 def _lazy_new(cls, *args, **kwargs):
--> 266 _lazy_init()
267 # We need this method only for lazy init, so we can remove it
268 del _CudaBase.new

/Users/hassanshallal/anaconda/lib/python2.7/site-packages/torch/cuda/init.pyc in _lazy_init()
82 raise RuntimeError(
83 "Cannot re-initialize CUDA in forked subprocess. " + msg)
---> 84 _check_driver()
85 torch._C._cuda_init()
86 torch._C._cuda_sparse_init()

/Users/hassanshallal/anaconda/lib/python2.7/site-packages/torch/cuda/init.pyc in _check_driver()
49 def _check_driver():
50 if not hasattr(torch._C, '_cuda_isDriverSufficient'):
---> 51 raise AssertionError("Torch not compiled with CUDA enabled")
52 if not torch._C._cuda_isDriverSufficient():
53 if torch._C._cuda_getDriverVersion() == 0:

AssertionError: (AssertionError('Torch not compiled with CUDA enabled',), <class 'torch.cuda.FloatTensor'>, ())
`

Thanks

[soumith]

@hassanshallal update your torchvision with: pip install -U torchvision

[hassanshallal]

Thanks a lot for the feedback, I can now upload and use the densenet pre-trained models.