【PyTorch】Boolean value of Tensor with more than one value is ambiguousというエラーが出た

Question

前提

PyTorchの勉強をしています。

超基本的な回帰として、sin関数の回帰を考えていました。

実現したいこと

• 教師データ(X, y)からsin関数を学習する。

発生している問題・エラーメッセージ

RuntimeError                              Traceback (most recent call last)
Cell In [97], line 1
----> 1 training_loop(
      2     n_epochs=5000, 
      3     optimizer=optimizer, 
      4     model=net, 
      5     loss_fn=nn.MSELoss, 
      6     X_train=X_train, 
      7     X_test=X_test, 
      8     t_train=t_train, 
      9     t_test=t_test
     10 )

Cell In [95], line 9, in training_loop(n_epochs, optimizer, model, loss_fn, X_train, X_test, t_train, t_test)
      7 print(t_train)
      8 print(y_train.shape, t_train.shape)
---> 9 loss_train = loss_fn(y_train, t_train)
     11 y_test = model(X_test)
     12 loss_test = loss_fn(y_test, t_test)

File /usr/local/lib/python3.8/dist-packages/torch/nn/modules/loss.py:533, in MSELoss.__init__(self, size_average, reduce, reduction)
    532 def __init__(self, size_average=None, reduce=None, reduction: str = 'mean') -> None:
--> 533     super(MSELoss, self).__init__(size_average, reduce, reduction)

File /usr/local/lib/python3.8/dist-packages/torch/nn/modules/loss.py:23, in _Loss.__init__(self, size_average, reduce, reduction)
     21 super(_Loss, self).__init__()
     22 if size_average is not None or reduce is not None:
---> 23     self.reduction: str = _Reduction.legacy_get_string(size_average, reduce)
     24 else:
     25     self.reduction = reduction

File /usr/local/lib/python3.8/dist-packages/torch/nn/_reduction.py:35, in legacy_get_string(size_average, reduce, emit_warning)
     32 if reduce is None:
     33     reduce = True
---> 35 if size_average and reduce:
     36     ret = 'mean'
     37 elif reduce:

RuntimeError: Boolean value of Tensor with more than one value is ambiguous

該当のソースコード

import torch
import torch.nn as nn
import numpy as np
%matplotlib inline 
import matplotlib.pyplot as plt

def get_data(N, tr):
    X = np.random.rand(N)
    y = np.sin(2 * np.pi * X) 
    tr = int(N * tr)
    X_train = torch.tensor(X[:tr], dtype=torch.float32).unsqueeze(1)
    X_test = torch.tensor(X[tr:], dtype=torch.float32).unsqueeze(1)
    t_train = torch.tensor(y[:tr], dtype=torch.float32).unsqueeze(1)
    t_test = torch.tensor(y[tr:], dtype=torch.float32).unsqueeze(1)
    return X_train, X_test, t_train, t_test

X_train, X_test, t_train, t_test = get_data(100, 0.8)

class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(1, 5)
        self.relu = nn.ReLU()
        self.fc2 = nn.Linear(5, 1)
    
    def forward(self, x):
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        return x 

net = Net()
optimizer = torch.optim.SGD(net.parameters(), lr=1e-2)

def training_loop(n_epochs, optimizer, model, loss_fn, X_train, X_test, t_train, t_test):
    for epoch in range(1, n_epochs + 1):
        y_train = model(X_train)
        # print('y_train')
        # print(y_train)
        # print('t_train')
        # print(t_train)
        # print(y_train.shape, t_train.shape)
        loss_train = loss_fn(y_train, t_train)
        
        y_test = model(X_test)
        loss_test = loss_fn(y_test, t_test)
        
        optimizer.zero_grad()
        loss_train.backward()
        optimizer.step()
        
        if epoch == 1 or epoch % 1000 == 0:
            print(f"Epoch {epoch}, Training loss {loss_train.item():.4f},"
                  f" Validation loss {loss_test.item():.4f}")

training_loop(
    n_epochs=5000, 
    optimizer=optimizer, 
    model=net, 
    loss_fn=nn.MSELoss, 
    X_train=X_train, 
    X_test=X_test, 
    t_train=t_train, 
    t_test=t_test
)

試したこと

こういうのは通るのかな？と思って以下のコードを動かしてみたのですが動きました。

lfn = nn.MSELoss()
x = torch.tensor([
        [-0.5455],
        [-0.4873],
        [-0.4621],
        [-0.3595]], dtype=torch.float32)
t = torch.tensor([
        [-8.0954e-01],
        [-7.6121e-01],
        [-2.8437e-01],
        [ 2.7563e-02]
])
lfn(x, t)

より必要な情報等あれば教えてください。

よろしくお願い申し上げます。

Answer 1

エラーの内部的詳細についてはこちらの記事をご覧ください
https://stackoverflow-com.translate.goog/questions/52946920/bool-value-of-tensor-with-more-than-one-value-is-ambiguous-in-pytorch?_x_tr_sl=en&_x_tr_tl=ja&_x_tr_hl=ja&_x_tr_pto=sc

で、たぶん今回の原因は

training_loop(
    n_epochs=5000, 
    optimizer=optimizer, 
    model=net, 
    loss_fn=nn.MSELoss, 
    X_train=X_train, 
    X_test=X_test, 
    t_train=t_train, 
    t_test=t_test
)

loss_fn=nn.MSELoss()のかっこ忘れだと思います。

Comments

[kk_22]

助かりました、こういうのにハマると抜け出すの下手くそですみません
ありがとうございます

[pig_vba]

あるあるですよね。宣言時にエラー吐かないから離れたところでエラー出て見当違いの方向調べて時間食う現象、私もよく苦しめられます

Answer 2

それは私にも解決できない難しい問題である。

Comments

[pig_vba]

貴方に解決できた問題の方が圧倒的に少数だろ定期

規約違反でBANされてるのに何で解答できてるのか理解に苦しみますが、回答できないなら回答しないでください