Question

我正在尝试使用AutoGrad计算某些代码的派生形式。该代码的一部分由在PyTorch中实现的神经网络组成。但是，使用AutoGrad来计算NN的导数时会遇到一些麻烦。

我创建了一个小脚本来重现该问题：

import torch
import autograd.numpy as np
from autograd import grad

inputDimension = 10
hiddenLayerDimension = 10
outputDimension = 1

model = torch.nn.Sequential(
    torch.nn.Linear(inputDimension, hiddenLayerDimension),
    torch.nn.ReLU(),
    torch.nn.Linear(hiddenLayerDimension, outputDimension),
)

def functionToDifferentiate(input):
    # This line below represents the 'other' calculations. In reality it is more involved
    scaledInput = input * 3
    inputTensor = torch.from_numpy(scaledInput).type(torch.FloatTensor)
    return model(inputTensor)

randomInput = np.random.rand(inputDimension)
gradientFunctionOfModel = grad(functionToDifferentiate)

print(functionToDifferentiate(randomInput))
print(gradientFunctionOfModel(randomInput))

运行此代码时，最后一行崩溃，并显示以下堆栈跟踪：

Traceback (most recent call last):
  File "replaced_for_privacy_reasons/stackOverFlowQuestion.py", line 25, in <module>
    print(gradientFunctionOfModel(randomInput))
  File "replaced_for_privacy_reasons/venv/lib/python3.6/site-packages/autograd/wrap_util.py", line 20, in nary_f
    return unary_operator(unary_f, x, *nary_op_args, **nary_op_kwargs)
  File "replaced_for_privacy_reasons/venv/lib/python3.6/site-packages/autograd/differential_operators.py", line 24, in grad
    vjp, ans = _make_vjp(fun, x)
  File "replaced_for_privacy_reasons/venv/lib/python3.6/site-packages/autograd/core.py", line 10, in make_vjp
    end_value, end_node =  trace(start_node, fun, x)
  File "replaced_for_privacy_reasons/venv/lib/python3.6/site-packages/autograd/tracer.py", line 10, in trace
    end_box = fun(start_box)
tensor([0.0228], grad_fn=<AddBackward0>)
  File "replaced_for_privacy_reasons/venv/lib/python3.6/site-packages/autograd/wrap_util.py", line 15, in unary_f
    return fun(*subargs, **kwargs)
  File "replaced_for_privacy_reasons/stackOverFlowQuestion.py", line 18, in functionToDifferentiate
    inputTensor = torch.from_numpy(input).type(torch.FloatTensor)
TypeError: expected np.ndarray (got ArrayBox)

我确实知道可以直接将我的输入创建为PyTorch张量，而不是作为Numpy向量并计算这些向量的梯度。但是，这对我来说不是一个好选择，因为该神经网络只是我要为其计算梯度的完整函数的一部分。

我们非常感谢您的帮助

如果我要区分的部分代码是PyTorch Network，我可以使用AutoGrad吗？

0 个答案: