无法获得正确的尺寸 - 用于文本分类的 CNN
Can't get dimensions right - CNN for text classification
这是我的 CNN class:
class CNN(nn.Module):
def __init__(
self,
vocab_size,
emb_dim,
out_channels,
kernel_sizes,
dropout,
):
super().__init__()
self.embedding = nn.Embedding(vocab_size, emb_dim)
self.conv_0 = nn.Conv2d(in_channels=1, out_channels=out_channels, kernel_size=(kernel_sizes[0], emb_dim), 2)
self.conv_1 = nn.Conv2d(in_channels=1, out_channels=out_channels, kernel_size=(kernel_sizes[1], emb_dim), 2)
self.conv_2 = nn.Conv2d(in_channels=1, out_channels=out_channels, kernel_size=(kernel_sizes[2], emb_dim), 2)
self.fc = nn.Linear(len(kernel_sizes) * out_channels, 1)
self.dropout = nn.Dropout(dropout)
def forward(self, text):
embedded = self.embedding(text)
print('embedded', embedded.shape)
embedded = embedded.unsqueeze(1) # may be reshape here
print('embedded', embedded.shape)
conved_0 = F.relu(self.conv_0(embedded)).squeeze(3) # may be reshape here
print('conved_0', conved_0.shape)
conved_1 = F.relu(self.conv_1(embedded)).squeeze(3) # may be reshape here
print('conved_1', conved_1.shape)
conved_2 = F.relu(self.conv_2(embedded)).squeeze(3) # may be reshape here
print('conved_2', conved_2.shape)
pooled_0 = F.max_pool1d(conved_0, conved_0.shape[2]).squeeze(2)
print('pooled_0', pooled_0.shape)
pooled_1 = F.max_pool1d(conved_1, conved_1.shape[2]).squeeze(2)
print('pooled_1', pooled_1.shape)
pooled_2 = F.max_pool1d(conved_2, conved_2.shape[2]).squeeze(2)
print('pooled_2', pooled_2.shape)
cat = self.dropout(torch.cat((pooled_0, pooled_1, pooled_2), dim=1))
print('cat', cat.shape)
return self.fc(cat)
变量：
kernel_sizes = [3, 4, 5]
vocab_size = len(TEXT.vocab)
out_channels = 64
dropout = 0.2
dim = 300
model = CNN(vocab_size=vocab_size, emb_dim=dim, out_channels=out_channels,
kernel_sizes=kernel_sizes, dropout=dropout)
和培训：
import numpy as np
min_loss = np.inf
cur_patience = 0
for epoch in range(1, max_epochs + 1):
train_loss = 0.0
model.train()
pbar = tqdm(enumerate(train_iter), total=len(train_iter), leave=False)
pbar.set_description(f"Epoch {epoch}")
for it, batch in pbar:
#YOUR CODE GOES HERE
opt.zero_grad()
input = batch.text[0].to(device)
output = model(input)
train_loss = loss_func(output, batch.label)
train_loss.backward()
opt.step()
train_loss /= len(train_iter)
val_loss = 0.0
model.eval()
pbar = tqdm(enumerate(valid_iter), total=len(valid_iter), leave=False)
pbar.set_description(f"Epoch {epoch}")
for it, batch in pbar:
# YOUR CODE GOES HERE
input = batch.text[0].to(device)
output = model(input)
val_loss = loss_fn(output, batch.label)
val_loss /= len(valid_iter)
if val_loss < min_loss:
min_loss = val_loss
best_model = model.state_dict()
else:
cur_patience += 1
if cur_patience == patience:
cur_patience = 0
break
print('Epoch: {}, Training Loss: {}, Validation Loss: {}'.format(epoch, train_loss, val_loss))
model.load_state_dict(best_model)
我收到这个错误：
RuntimeError: Expected 4-dimensional input for 4-dimensional weight [64, 1, 3, 300], but got 3-dimensional input of size [894, 1, 300] instead
在这一行中：
---> 32 conved_0 = F.relu(self.conv_0(embedded)).squeeze(3)
我试过使用 Conv1d，但在尺寸方面仍然存在问题。有人可以解释一下我应该在这里修复什么来训练网络吗？
编辑：
This is my class but with Conv1d:
class CNN(nn.Module):
def __init__(
self,
vocab_size,
emb_dim,
out_channels,
kernel_sizes,
dropout,
):
super().__init__()
self.embedding = nn.Embedding(vocab_size, emb_dim)
self.conv_0 = nn.Conv1d(in_channels=1, out_channels=out_channels, kernel_size=kernel_sizes[0])
self.conv_1 = nn.Conv1d(in_channels=1, out_channels=out_channels, kernel_size=kernel_sizes[1])
self.conv_2 = nn.Conv1d(in_channels=1, out_channels=out_channels, kernel_size=kernel_sizes[2])
self.fc = nn.Linear(len(kernel_sizes) * out_channels, 1)
self.dropout = nn.Dropout(dropout)
def forward(self, text):
embedded = self.embedding(text)
print('embedded', embedded.shape)
embedded = embedded.unsqueeze(1) # may be reshape here
print('embedded', embedded.shape)
conved_0 = F.relu(self.conv_0(embedded)) # may be reshape here
print('conved_0', conved_0.shape)
conved_1 = F.relu(self.conv_1(embedded)) # may be reshape here
print('conved_1', conved_1.shape)
conved_2 = F.relu(self.conv_2(embedded)) # may be reshape here
print('conved_2', conved_2.shape)
pooled_0 = F.max_pool1d(conved_0, conved_0.shape[2]).squeeze(2)
print('pooled_0', pooled_0.shape)
pooled_1 = F.max_pool1d(conved_1, conved_1.shape[2]).squeeze(2)
print('pooled_1', pooled_1.shape)
pooled_2 = F.max_pool1d(conved_2, conved_2.shape[2]).squeeze(2)
print('pooled_2', pooled_2.shape)
cat = self.dropout(torch.cat((pooled_0, pooled_1, pooled_2), dim=1))
print('cat', cat.shape)
return self.fc(cat)
维度输出：
embedded torch.Size([1115, 300])
embedded torch.Size([1115, 1, 300])
conved_0 torch.Size([1115, 64, 298])
conved_1 torch.Size([1115, 64, 297])
conved_2 torch.Size([1115, 64, 296])
pooled_0 torch.Size([1115, 64])
pooled_1 torch.Size([1115, 64])
pooled_2 torch.Size([1115, 64])
cat torch.Size([1115, 192])
错误：
ValueError: Target size (torch.Size([128])) must be the same as input size (torch.Size([1115, 1]))
我缺少的是我将 batch_first 参数设置为 True，其中 SWAPED batch_size 和 seq_len。将其设置为 False 后，一切正常。
这是我的 CNN class:
class CNN(nn.Module):
def __init__(
self,
vocab_size,
emb_dim,
out_channels,
kernel_sizes,
dropout,
):
super().__init__()
self.embedding = nn.Embedding(vocab_size, emb_dim)
self.conv_0 = nn.Conv2d(in_channels=1, out_channels=out_channels, kernel_size=(kernel_sizes[0], emb_dim), 2)
self.conv_1 = nn.Conv2d(in_channels=1, out_channels=out_channels, kernel_size=(kernel_sizes[1], emb_dim), 2)
self.conv_2 = nn.Conv2d(in_channels=1, out_channels=out_channels, kernel_size=(kernel_sizes[2], emb_dim), 2)
self.fc = nn.Linear(len(kernel_sizes) * out_channels, 1)
self.dropout = nn.Dropout(dropout)
def forward(self, text):
embedded = self.embedding(text)
print('embedded', embedded.shape)
embedded = embedded.unsqueeze(1) # may be reshape here
print('embedded', embedded.shape)
conved_0 = F.relu(self.conv_0(embedded)).squeeze(3) # may be reshape here
print('conved_0', conved_0.shape)
conved_1 = F.relu(self.conv_1(embedded)).squeeze(3) # may be reshape here
print('conved_1', conved_1.shape)
conved_2 = F.relu(self.conv_2(embedded)).squeeze(3) # may be reshape here
print('conved_2', conved_2.shape)
pooled_0 = F.max_pool1d(conved_0, conved_0.shape[2]).squeeze(2)
print('pooled_0', pooled_0.shape)
pooled_1 = F.max_pool1d(conved_1, conved_1.shape[2]).squeeze(2)
print('pooled_1', pooled_1.shape)
pooled_2 = F.max_pool1d(conved_2, conved_2.shape[2]).squeeze(2)
print('pooled_2', pooled_2.shape)
cat = self.dropout(torch.cat((pooled_0, pooled_1, pooled_2), dim=1))
print('cat', cat.shape)
return self.fc(cat)
变量：
kernel_sizes = [3, 4, 5]
vocab_size = len(TEXT.vocab)
out_channels = 64
dropout = 0.2
dim = 300
model = CNN(vocab_size=vocab_size, emb_dim=dim, out_channels=out_channels,
kernel_sizes=kernel_sizes, dropout=dropout)
和培训：
import numpy as np
min_loss = np.inf
cur_patience = 0
for epoch in range(1, max_epochs + 1):
train_loss = 0.0
model.train()
pbar = tqdm(enumerate(train_iter), total=len(train_iter), leave=False)
pbar.set_description(f"Epoch {epoch}")
for it, batch in pbar:
#YOUR CODE GOES HERE
opt.zero_grad()
input = batch.text[0].to(device)
output = model(input)
train_loss = loss_func(output, batch.label)
train_loss.backward()
opt.step()
train_loss /= len(train_iter)
val_loss = 0.0
model.eval()
pbar = tqdm(enumerate(valid_iter), total=len(valid_iter), leave=False)
pbar.set_description(f"Epoch {epoch}")
for it, batch in pbar:
# YOUR CODE GOES HERE
input = batch.text[0].to(device)
output = model(input)
val_loss = loss_fn(output, batch.label)
val_loss /= len(valid_iter)
if val_loss < min_loss:
min_loss = val_loss
best_model = model.state_dict()
else:
cur_patience += 1
if cur_patience == patience:
cur_patience = 0
break
print('Epoch: {}, Training Loss: {}, Validation Loss: {}'.format(epoch, train_loss, val_loss))
model.load_state_dict(best_model)
我收到这个错误：
RuntimeError: Expected 4-dimensional input for 4-dimensional weight [64, 1, 3, 300], but got 3-dimensional input of size [894, 1, 300] instead
在这一行中：
---> 32 conved_0 = F.relu(self.conv_0(embedded)).squeeze(3)
我试过使用 Conv1d，但在尺寸方面仍然存在问题。有人可以解释一下我应该在这里修复什么来训练网络吗？
编辑：
This is my class but with Conv1d:
class CNN(nn.Module):
def __init__(
self,
vocab_size,
emb_dim,
out_channels,
kernel_sizes,
dropout,
):
super().__init__()
self.embedding = nn.Embedding(vocab_size, emb_dim)
self.conv_0 = nn.Conv1d(in_channels=1, out_channels=out_channels, kernel_size=kernel_sizes[0])
self.conv_1 = nn.Conv1d(in_channels=1, out_channels=out_channels, kernel_size=kernel_sizes[1])
self.conv_2 = nn.Conv1d(in_channels=1, out_channels=out_channels, kernel_size=kernel_sizes[2])
self.fc = nn.Linear(len(kernel_sizes) * out_channels, 1)
self.dropout = nn.Dropout(dropout)
def forward(self, text):
embedded = self.embedding(text)
print('embedded', embedded.shape)
embedded = embedded.unsqueeze(1) # may be reshape here
print('embedded', embedded.shape)
conved_0 = F.relu(self.conv_0(embedded)) # may be reshape here
print('conved_0', conved_0.shape)
conved_1 = F.relu(self.conv_1(embedded)) # may be reshape here
print('conved_1', conved_1.shape)
conved_2 = F.relu(self.conv_2(embedded)) # may be reshape here
print('conved_2', conved_2.shape)
pooled_0 = F.max_pool1d(conved_0, conved_0.shape[2]).squeeze(2)
print('pooled_0', pooled_0.shape)
pooled_1 = F.max_pool1d(conved_1, conved_1.shape[2]).squeeze(2)
print('pooled_1', pooled_1.shape)
pooled_2 = F.max_pool1d(conved_2, conved_2.shape[2]).squeeze(2)
print('pooled_2', pooled_2.shape)
cat = self.dropout(torch.cat((pooled_0, pooled_1, pooled_2), dim=1))
print('cat', cat.shape)
return self.fc(cat)
维度输出：
embedded torch.Size([1115, 300])
embedded torch.Size([1115, 1, 300])
conved_0 torch.Size([1115, 64, 298])
conved_1 torch.Size([1115, 64, 297])
conved_2 torch.Size([1115, 64, 296])
pooled_0 torch.Size([1115, 64])
pooled_1 torch.Size([1115, 64])
pooled_2 torch.Size([1115, 64])
cat torch.Size([1115, 192])
错误：
ValueError: Target size (torch.Size([128])) must be the same as input size (torch.Size([1115, 1]))
我缺少的是我将 batch_first 参数设置为 True，其中 SWAPED batch_size 和 seq_len。将其设置为 False 后，一切正常。