search

CSAILVision / IBD

IBD: Interpretable Basis Decomposition for Visual Explanation
51 stars 14 forks source link

Add support to py3.7 and pytorch > 1.0 #2

Open drmeerkat opened 5 years ago

drmeerkat commented 5 years ago
  1. In python3.7 "async" is used by python as a keyword which will result in a syntax error when used as a function parameter. Pytorch cuda semantics change it non_blocking async = True -> non_blocking = True
  2. volatile was removed in pytorch 1.0

    IBD-master/util/feature_operation.py:175: UserWarning: volatile was removed and now has no effect. Use with torch.no_grad(): instead. input_var = V(input,volatile=True)

  3. Connot convert cuda tensor directly to numpy array

Traceback (most recent call last): File "test.py", line 20, in features, _ = fo.feature_extraction(model=model) File "/data4/lmx/tmp/IBD-master/util/feature_operation.py", line 180, in feature_extraction while np.isnan(output.data.max()): File "/data4/lmx/anaconda3/lib/python3.7/site-packages/torch/tensor.py", line 450, in array return self.numpy() TypeError: can't convert CUDA tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

Modify line 180 in feature_operation.py will fix this issue while np.isnan(output.cpu().data.max()):