Question

我想修改“ torch.nn.functional”中的某些函数，然后使用“ importlib.reload”重新加载新的“ torch.nn.functional”。但是我遇到运行时错误：function'avg_pool2d'已经有一个文档字符串。

我正在尝试对励磁bp（https://github.com/greydanus/excitationbp）进行一些工作，并且有一个名为“ eb.use_eb（）”的操作可以修改“ torch.nn.functional”。函数eb.use_eb（）用于修改“ torch.nn.linear”，“ torch.nn.functional.conv1d”等。我尝试修改utils.py中包括许多eb.use_eb（）的代码。。当我想运行原始的“ torch.nn.linear”时，必须运行eb.use_eb（False）。但是模块已加载到文件的开头（导入torch.nn.funtional）。因此，我想重新加载模块“ torch.nn.funtional”。当我使用“ importlib.reload”时，遇到错误。

eb.use_eb(False,True)
importlib.reload(torch)
importlib.reload(torch.nn)
importlib.reload(torch.nn.functional) 
print("gradient_evidence:"+str(gradient_evidence))
return torch.autograd.grad(contr_h_, target_h_, grad_outputs=gradient_evidence)[0]

RuntimeError         Traceback (most recent call last)
ipython-input-15-29d0fca29882 in module()
----> 1 prob_inputs_one = eb.excitation_backprop(model, inputs, prob_outputs_one, contrastive=2)
      2 #pdb.set_trace()
      3 prob_inputs_true = eb.excitation_backprop(model, inputs, prob_outputs_true, contrastive=2)

~/code/excitationbp/excitationbp-master/excitationbp/utils.py in excitation_backprop(model, inputs, prob_outputs, contrastive, target_layer)
     75         importlib.reload(torch)
     76         importlib.reload(torch.nn)
---> 77         importlib.reload(torch.nn.functional)
     78 
     79         gradient_evidence = torch.autograd.grad(top_h_, contr_h_, grad_outputs=prob_outputs.clone())[0]

~/anaconda3/envs/pytorch0.3/lib/python3.6/importlib/__init__.py in reload(module)
    164         target = module
    165         spec = module.__spec__ = _bootstrap._find_spec(name, pkgpath, target)
--> 166         _bootstrap._exec(spec, module)
    167         # The module may have replaced itself in sys.modules!
    168         return sys.modules[name]

~/anaconda3/envs/pytorch0.3/lib/python3.6/importlib/_bootstrap.py in _exec(spec, module)

~/anaconda3/envs/pytorch0.3/lib/python3.6/importlib/_bootstrap_external.py in exec_module(self, module)

~/anaconda3/envs/pytorch0.3/lib/python3.6/importlib/_bootstrap.py in _call_with_frames_removed(f, *args, **kwds)

~/anaconda3/envs/pytorch0.3/lib/python3.6/site-packages/torch/nn/functional.py in <module>()
    286     count_include_pad: when True, will include the zero-padding in th
    287         averaging calculation. Default: ``True``
--> 288 """)
    289 
    290 avg_pool3d = _add_docstr(torch._C._nn.avg_pool3d, r"""

RuntimeError: function 'avg_pool2d' already has a docstring

重新加载“ torch.nn.functioanl”时会出错吗？

0 个答案: