pytorch：GRU 無法就地更新 hidden_state

Question

使用 pytorch 實現 GRU 網絡時遇到問題：

我的代碼如下：

import torch

class GRU_model(torch.nn.Module):
    def __init__(self, device):
        super(GRU_model, self).__init__()
        self.h = torch.randn((1,1,5), device=device, dtype=torch.float)

        self.GRU_1 = torch.nn.GRU(input_size=5, hidden_size=5)

    def forward(self, a):
        output, self.h = self.GRU_1(a, self.h)
        return output

if __name__ == '__main__':
    learn_rate=1e-4
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = GRU_model(device).to(device=device)
    optimizer = torch.optim.Adam(model.parameters(), lr=learn_rate)

    for i in range(10):
        a = torch.randn((1, 1, 5), device=device, dtype=torch.float)
        output = model(a)
        loss = (a - output).mean()

        optimizer.zero_grad()
        loss.backward(retain_graph=True)
        optimizer.step()

我收到這樣的錯誤：

Traceback (most recent call last):
  File "C:/Users/Administrator_/Desktop/Graduation_Project/MIDI_Music_style_transfer/GRU_toy_in-place_hidden_states_change/main.py", line 40, in <module>
    loss.backward(retain_graph=True)
  File "C:\Users\Administrator_\AppData\Local\Programs\Python\Python37\lib\site-packages\torch\tensor.py", line 245, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "C:\Users\Administrator_\AppData\Local\Programs\Python\Python37\lib\site-packages\torch\autograd\__init__.py", line 147, in backward
    allow_unreachable=True, accumulate_grad=True)  # allow_unreachable flag
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [15, 5]] is at version 2; expected version 1 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

我只想在一個紀元后更新 GRU 中的 hidden_state ，但它不起作用！

如果您能幫上忙，我將不勝感激！

Answer 1

在 PyTorch 中，為一個 epoch 中的每次迭代創建計算圖。 在每次迭代中，我們執行前向傳遞，計算 output w.r.t 到網絡參數的導數，並更新參數以適應給定的示例。 執行反向傳遞后，圖形將被釋放以保存 memory。 在下一次迭代中，將創建一個全新的圖並准備好進行反向傳播。

因為計算圖將在第一次反向傳遞后默認釋放，如果您嘗試第二次在同一個圖上進行反向操作，您將遇到錯誤。 這就是彈出以下錯誤消息的原因：

RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time

資源

在您的情況下，在指定retain_graph=True后，您會看到：

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [15, 5]] is at version 2; expected version 1 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

當您嘗試在前向傳遞中更新self.h時會出現此問題。 您沒有inplace修改它，因為它是 grad 計算所必需的。 來源這個應該可以工作：

import torch


class GRU_model(torch.nn.Module):
    def __init__(self, device):
        super(GRU_model, self).__init__()
        self.GRU_1 = torch.nn.GRU(input_size=5, hidden_size=5)

    def forward(self, a, h):
        output, hh = self.GRU_1(a, h)
        return output, hh


if __name__ == '__main__':
    learn_rate = 1e-4
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = GRU_model(device).to(device=device)
    model.train()
    optimizer = torch.optim.Adam(model.parameters(), lr=learn_rate)

    def mse(output, target):
        loss = torch.mean((output - target)**2)
        return loss

    for i in range(10):
        a = torch.randn((1, 1, 5), device=device, dtype=torch.float, requires_grad=True)
        target = torch.randn((1, 1, 5), device=device)
        h = torch.randn((1, 1, 5), device=device)

        optimizer.zero_grad()
        output, h = model(a, h)
        loss = mse(output, target)

        loss.backward(retain_graph=True)
        optimizer.step()

Answer 2

真正的問題是 hidden_state 不應該參與梯度反向傳播的計算。 因此，只需添加一行self.h = self.h.detach()如下，肯定會解決問題：

    def forward(self, a):
        output, self.h = self.GRU_1(a, self.h)
        self.h = self.h.detach()
        return output