为什么将输入和模型强制转换为float16不起作用？

Question

我正在尝试将输入和深度学习模型更改为flaot16，因为我使用的是T4 GPU，它们在fp16上的运行速度更快。 这是代码的一部分：我首先有我的模型，然后制作了一些虚拟数据点，以期弄清楚首先要弄清楚的数据转换（我在整个批次中都运行了该错误，并且得到了相同的错误）。

model = CRNN().to(device)
model = model.type(torch.cuda.HalfTensor)

data_recon = torch.from_numpy(data_recon)
data_truth = torch.from_numpy(data_truth)

dummy = data_recon[0:1,:,:,:,:] # Gets just one batch
dummy = dummy.to(device)
dummy = dummy.type(torch.cuda.HalfTensor)

model(dummy)

这是我得到的错误：

> --------------------------------------------------------------------------- 
RuntimeError                              Traceback (most recent call
> last) <ipython-input-27-1fe8ecc524aa> in <module>
> ----> 1 model(dummy)
> 
> /opt/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py
> in __call__(self, *input, **kwargs)
>     491             result = self._slow_forward(*input, **kwargs)
>     492         else:
> --> 493             result = self.forward(*input, **kwargs)
>     494         for hook in self._forward_hooks.values():
>     495             hook_result = hook(self, input, result)
> 
> <ipython-input-12-06f39f9304a1> in forward(self, inputs, test)
>      57 
>      58             net['t%d_x0'%(i-1)] = net['t%d_x0'%(i-1)].view(times, batch, self.filter_size, width,
> height)
> ---> 59             net['t%d_x0'%i] = self.bcrnn(inputs, net['t%d_x0'%(i-1)], test)
>      60             net['t%d_x0'%i] = net['t%d_x0'%i].view(-1, self.filter_size, width, height)
>      61 
> 
> /opt/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py
> in __call__(self, *input, **kwargs)
>     491             result = self._slow_forward(*input, **kwargs)
>     492         else:
> --> 493             result = self.forward(*input, **kwargs)
>     494         for hook in self._forward_hooks.values():
>     495             hook_result = hook(self, input, result)
> 
> <ipython-input-11-b687949e9ce5> in forward(self, inputs,
> input_iteration, test)
>      31         hidden = initial_hidden
>      32         for i in range(times):
> ---> 33             hidden = self.CRNN(inputs[i], input_iteration[i], hidden)
>      34             output_forward.append(hidden)
>      35         output_forward = torch.cat(output_forward)
> 
> /opt/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py
> in __call__(self, *input, **kwargs)
>     491             result = self._slow_forward(*input, **kwargs)
>     492         else:
> --> 493             result = self.forward(*input, **kwargs)
>     494         for hook in self._forward_hooks.values():
>     495             hook_result = hook(self, input, result)
> 
> <ipython-input-10-15c0b221226b> in forward(self, inputs,
> hidden_iteration, hidden)
>      23     def forward(self, inputs, hidden_iteration, hidden):
>      24         in_to_hid = self.i2h(inputs)
> ---> 25         hid_to_hid = self.h2h(hidden)
>      26         ih_to_ih = self.ih2ih(hidden_iteration)
>      27 
> 
> /opt/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py
> in __call__(self, *input, **kwargs)
>     491             result = self._slow_forward(*input, **kwargs)
>     492         else:
> --> 493             result = self.forward(*input, **kwargs)
>     494         for hook in self._forward_hooks.values():
>     495             hook_result = hook(self, input, result)
> 
> /opt/anaconda3/lib/python3.7/site-packages/torch/nn/modules/conv.py in
> forward(self, input)
>     336                             _pair(0), self.dilation, self.groups)
>     337         return F.conv2d(input, self.weight, self.bias, self.stride,
> --> 338                         self.padding, self.dilation, self.groups)
>     339 
>     340 
> 
> RuntimeError: Input type (torch.cuda.FloatTensor) and weight type
> (torch.cuda.HalfTensor) should be the same

Answer 1

检查您对CRNN的实现。 我的猜测是您在模型中存储了“隐藏”状态张量，但不是作为“缓冲区”而是作为常规张量存储。 因此，将模型转换为float16时，隐藏状态仍为float32并导致此错误。

尝试将隐藏状态存储为模块中的register_buffer有关更多信息，请参见register_buffer ）。
另外，您可以通过重载模型的.to .to()方法将模块中的任何成员张量显式转换为float16。