损失函数的选择

以下是word2vec的实现:

%reset -f

import torch
from torch.autograd import Variable
import numpy as np
import torch.functional as F
import torch.nn.functional as F

corpus = [
    'this test',
    'this separate test'
]

def get_input_layer(word_idx):
    x = torch.zeros(vocabulary_size).float()
    x[word_idx] = 1.0
    return x

def tokenize_corpus(corpus):
    tokens = [x.split() for x in corpus]
    return tokens

tokenized_corpus = tokenize_corpus(corpus)

vocabulary = []
for sentence in tokenized_corpus:
    for token in sentence:
        if token not in vocabulary:
            vocabulary.append(token)

word2idx = {w: idx for (idx, w) in enumerate(vocabulary)}
idx2word = {idx: w for (idx, w) in enumerate(vocabulary)}

window_size = 2
idx_pairs = []
# for each sentence
for sentence in tokenized_corpus:
    indices = [word2idx[word] for word in sentence]
    # for each word, threated as center word
    for center_word_pos in range(len(indices)):
        # for each window position
        for w in range(-window_size, window_size + 1):
            context_word_pos = center_word_pos + w
            # make soure not jump out sentence
            if context_word_pos < 0 or context_word_pos >= len(indices) or center_word_pos == context_word_pos:
                continue
            context_word_idx = indices[context_word_pos]
            idx_pairs.append((indices[center_word_pos], context_word_idx))

idx_pairs = np.array(idx_pairs) # it will be useful to have this as numpy array

vocabulary_size = len(vocabulary)

embedding_dims = 4
W1 = Variable(torch.randn(embedding_dims, vocabulary_size).float(), requires_grad=True)
W2 = Variable(torch.randn(vocabulary_size, embedding_dims).float(), requires_grad=True)
num_epochs = 1
learning_rate = 0.001

for epo in range(num_epochs):
    loss_val = 0
    for data, target in idx_pairs:
        x = Variable(get_input_layer(data)).float()
        y_true = Variable(torch.from_numpy(np.array([target])).long())

        z1 = torch.matmul(W1, x)
        z2 = torch.matmul(W2, z1)

        log_softmax = F.log_softmax(z2, dim=0)

        loss = F.nll_loss(log_softmax.view(1,-1), y_true)
        print(float(loss))

        loss_val += loss.data.item()
        loss.backward()
        W1.data -= learning_rate * W1.grad.data
        W2.data -= learning_rate * W2.grad.data

        W1.grad.data.zero_()
        W2.grad.data.zero_()

        print(W1.shape)
        print(W2.shape)

        print(f'Loss at epo {epo}: {loss_val/len(idx_pairs)}')

这张照片是:

0.33185482025146484
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 0.041481852531433105
3.302438735961914
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 0.45428669452667236
2.3144636154174805
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 0.7435946464538574
0.33418864011764526
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 0.7853682264685631
1.0644199848175049
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 0.9184207245707512
0.4970806837081909
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 0.980555810034275
3.2861199378967285
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 1.3913208022713661
6.170125961303711
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 2.16258654743433

修改代码以使用mse_loss change y_true来浮动:

y_true = Variable(torch.from_numpy(np.array([target])).float())

使用mse_损耗:

loss = F.mse_loss(log_softmax.view(1,-1), y_true)

合并更新:

for epo in range(num_epochs):
    loss_val = 0
    for data, target in idx_pairs:
        x = Variable(get_input_layer(data)).float()
        y_true = Variable(torch.from_numpy(np.array([target])).float())

        z1 = torch.matmul(W1, x)
        z2 = torch.matmul(W2, z1)

        log_softmax = F.log_softmax(z2, dim=0)

        loss = F.mse_loss(log_softmax.view(1,-1), y_true)
        print(float(loss))

        loss_val += loss.data.item()
        loss.backward()
        W1.data -= learning_rate * W1.grad.data
        W2.data -= learning_rate * W2.grad.data

        W1.grad.data.zero_()
        W2.grad.data.zero_()

        print(W1.shape)
        print(W2.shape)

        print(f'Loss at epo {epo}: {loss_val/len(idx_pairs)}')

现在的输出是:

41.75048828125
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 5.21881103515625
16.929386138916016
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 7.334984302520752
50.63690948486328
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 13.664597988128662
36.21110534667969
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 18.190986156463623
5.304859638214111
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 18.854093611240387
9.802173614501953
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 20.07936531305313
15.515325546264648
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 22.018781006336212
30.408292770385742
torch.Size([4, 3])
torch.Size([3, 4])
Loss at epo 0: 25.81981760263443

-c:12: UserWarning: Using a target size (torch.Size([1])) that is different to the input size (torch.Size([1, 3])). This will likely lead to incorrect results due to broadcasting. Please ensure they have the same size.

为什么mse损失不如nll损失有效?是否与PyTorch警告有关:

Using a target size (torch.Size([1])) that is different to the input size (torch.Size([1, 3])). This will likely lead to incorrect results due to broadcasting. Please ensure they have the same size.

?