RuntimeError: size mismatch (got input: [3], target: [32])

1

import torch
import torch.nn as nn
import torch.nn.functional as F

class BiLSTMWithAttention(nn.Module):
def init(self, embedding_dim=768, hidden_dim=128, num_classes=3):
super(BiLSTMWithAttention, self).init()
self.bilstm = nn.LSTM(input_size=embedding_dim, hidden_size=hidden_dim, num_layers=1, batch_first=True, bidirectional=True)
self.attn = nn.Linear(2 * hidden_dim, 1)
self.dropout = nn.Dropout(0.3)
self.fc1 = nn.Linear(32, 128)
self.fc2 = nn.Linear(128, num_classes)

def forward(self, x):
    # x shape: (batch_size, seq_len, embedding_dim)
    lstm_out, _ = self.bilstm(x)  # lstm_out: (batch, seq_len, 2*hidden_dim)

    attn_scores = self.attn(lstm_out)  # (batch, seq_len, 1)
    attn_weights = torch.softmax(attn_scores, dim=1)  # (batch, seq_len, 1)

    context = torch.sum(attn_weights * lstm_out, dim=1)  # (batch, 2*hidden_dim)

    x = self.dropout(F.relu(self.fc1(context)))  # (batch, 128)
    return self.fc2(x)  # (batch, num_classes)
2

There is a bug in your BiLSTMWithAttention class implementation.

The input to fc1 will be the context tensor, which has the shape of [batch, 2 * hidden_dim]. Hence the in_features to fc1 will be 2 * hidden_dim and not simply 32. No other changes are necessary.

TLDR:
fc1 should be initialized this way in the __init__ method: self.fc1 = nn.Linear(2 * hidden_dim, 128)

Hope this helps.

Kindly enclose your code snippets in code blocks for clarity and ease of reading.