Transformer Encoder implemented by Pytorch

Attention.py

import torch
from torch import nn
import torch.nn.functional as F
from math import sqrt

class AttentionHead(nn.Module):
    def __init__(self, embed_dim, head_dim):
        super().__init__()
        self.q = nn.Linear(embed_dim, head_dim)
        self.k = nn.Linear(embed_dim, head_dim)
        self.v = nn.Linear(embed_dim, head_dim)

    def forward(self, query, key, value, mask=None):
        query, key, value = self.q(query), self.k(key), self.v(value)
        scores = torch.bmm(query, key.transpose(1, 2)) / sqrt(query.size(-1))
        if mask is not None:
            scores = scores.masked_fill(mask == 0, -float("inf"))
        weights = F.softmax(scores, dim=-1)
        return torch.bmm(weights, value)

class MultiHeadAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        embed_dim = config.hidden_size
        num_heads = config.num_attention_heads
        head_dim = embed_dim // num_heads
        self.heads = nn.ModuleList(
            [AttentionHead(embed_dim, head_dim) for _ in range(num_heads)]
        )
        self.output_linear = nn.Linear(embed_dim, embed_dim)

    def forward(self, query, key, value, mask=None, query_mask=None, key_mask=None):
        if query_mask is not None and key_mask is not None:
            mask = torch.bmm(query_mask.unsqueeze(-1), key_mask.unsqueeze(1))
        x = torch.cat([h(query, key, value, mask) for h in self.heads], dim=-1)
        x = self.output_linear(x)
        return x

class FeedForward(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.linear_1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.linear_2 = nn.Linear(config.intermediate_size, config.hidden_size)
        self.gelu = nn.GELU()
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, x):
        x = self.linear_1(x)
        x = self.gelu(x)
        x = self.linear_2(x)
        x = self.dropout(x)
        return x

class TransformerEncoderLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.layer_norm_1 = nn.LayerNorm(config.hidden_size)
        self.layer_norm_2 = nn.LayerNorm(config.hidden_size)
        self.attention = MultiHeadAttention(config)
        self.feed_forward = FeedForward(config)

    def forward(self, x, mask=None):
        # Apply layer normalization and then copy input into query, key, value
        hidden_state = self.layer_norm_1(x)
        # Apply attention with a skip connection
        x = x + self.attention(hidden_state, hidden_state, hidden_state, mask=mask)
        # Apply feed-forward layer with a skip connection
        x = x + self.feed_forward(self.layer_norm_2(x))
        return x

class Embeddings(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.token_embeddings = nn.Embedding(config.vocab_size,
                                             config.hidden_size)
        self.position_embeddings = nn.Embedding(config.max_position_embeddings,
                                                config.hidden_size)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=1e-12)
        self.dropout = nn.Dropout()

    def forward(self, input_ids):
        # Create position IDs for input sequence
        seq_length = input_ids.size(1)
        position_ids = torch.arange(seq_length, dtype=torch.long).unsqueeze(0)
        # Create token and position embeddings
        token_embeddings = self.token_embeddings(input_ids)
        position_embeddings = self.position_embeddings(position_ids)
        # Combine token and position embeddings
        embeddings = token_embeddings + position_embeddings
        embeddings = self.layer_norm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings

class TransformerEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.embeddings = Embeddings(config)
        self.layers = nn.ModuleList(
            [TransformerEncoderLayer(config) for _ in range(config.num_hidden_layers)]
        )

    def forward(self, x, mask=None):
        x = self.embeddings(x)
        for layer in self.layers:
            x = layer(x, mask)
        return x

if __name__ == '__main__':
    from transformers import AutoConfig
    from transformers import AutoTokenizer

    model_ckpt = "bert-base-uncased"
    tokenizer = AutoTokenizer.from_pretrained(model_ckpt)
    config = AutoConfig.from_pretrained(model_ckpt)

    text = "time flies like an arrow"
    inputs = tokenizer(text, return_tensors="pt", add_special_tokens=False)

    encoder = TransformerEncoder(config)
    print(encoder(inputs.input_ids).size())

why do you not use n+1 dimension for 'multihead'-attention against primitive attention, but not multihead wrapperd attention-head?

why do you not use n+1 dimension for 'multihead'-attention against primitive attention, but not multihead wrapperd attention-head?

不清楚你说的什么意思，原版的多头注意力就是把多个头的结果简单拼接，不会多出头维度。

why do you not use n+1 dimension for 'multihead'-attention against primitive attention, but not multihead wrapperd attention-head?
这两种实现方式是不一样的。上面的代码更符合论文定义。多一个维度的那种，前向计算的结果和本文的实现都不相同。