transformer.py

import copy
import math
import torch
import torch.nn as nn
import torch.nn.functional as fn

class MultiHeadAttention(nn.Module):
    def __init__(self, n_heads, hidden_size, hidden_dropout_prob, attn_dropout_prob, layer_norm_eps):
        super(MultiHeadAttention, self).__init__()
        if hidden_size % n_heads != 0:
            raise ValueError("The hidden size (%d) is not a multiple of the number of attention heads (%d)" % (hidden_size, n_heads))

        self.num_attention_heads = n_heads
        self.attention_head_size = int(hidden_size / n_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(hidden_size, self.all_head_size)
        self.key = nn.Linear(hidden_size, self.all_head_size)
        self.value = nn.Linear(hidden_size, self.all_head_size)

        self.attn_dropout = nn.Dropout(attn_dropout_prob)

        self.dense = nn.Linear(hidden_size, hidden_size)
        self.LayerNorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
        self.out_dropout = nn.Dropout(hidden_dropout_prob)

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(self, input_tensor, attention_mask):
        mixed_query_layer = self.query(input_tensor)
        mixed_key_layer = self.key(input_tensor)
        mixed_value_layer = self.value(input_tensor)

        query_layer = self.transpose_for_scores(mixed_query_layer)
        key_layer = self.transpose_for_scores(mixed_key_layer)
        value_layer = self.transpose_for_scores(mixed_value_layer)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))

        # PLACE1: freeze the attention matrix here

        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
        # [batch_size heads seq_len seq_len] scores
        # [batch_size 1 1 seq_len]
        attention_scores = attention_scores + attention_mask

        # Normalize the attention scores to probabilities.
        # attention_scores = torch.full(attention_scores.shape, 1.0).cuda() # PLACE2: freeze the attention matrix here
        attention_probs = nn.Softmax(dim=-1)(attention_scores)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.

        attention_probs = self.attn_dropout(attention_probs)
        context_layer = torch.matmul(attention_probs, value_layer)

        # PLACE3: freeze the attention matrix here

        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)
        hidden_states = self.dense(context_layer)
        hidden_states = self.out_dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)

        return hidden_states


class FeedForward(nn.Module):
    def __init__(self, hidden_size, inner_size, hidden_dropout_prob, hidden_act, layer_norm_eps):
        super(FeedForward, self).__init__()
        self.dense_1 = nn.Linear(hidden_size, inner_size)
        self.intermediate_act_fn = self.get_hidden_act(hidden_act)

        self.dense_2 = nn.Linear(inner_size, hidden_size)
        self.LayerNorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
        self.dropout = nn.Dropout(hidden_dropout_prob)

    def get_hidden_act(self, act):
        ACT2FN = {
            "gelu": self.gelu,
            "relu": fn.relu,
            "swish": self.swish,
            "tanh": torch.tanh,
            "sigmoid": torch.sigmoid,
        }
        return ACT2FN[act]

    def gelu(self, x):
        return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))

    def swish(self, x):
        return x * torch.sigmoid(x)

    def forward(self, input_tensor):
        hidden_states = self.dense_1(input_tensor)
        hidden_states = self.intermediate_act_fn(hidden_states)

        hidden_states = self.dense_2(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)

        return hidden_states

class TransformerLayer(nn.Module):
    def __init__(self, n_heads, hidden_size, intermediate_size, hidden_dropout_prob, attn_dropout_prob, hidden_act, layer_norm_eps):
        super(TransformerLayer, self).__init__()
        self.multi_head_attention = MultiHeadAttention(n_heads, hidden_size, hidden_dropout_prob, attn_dropout_prob, layer_norm_eps)
        self.feed_forward = FeedForward(hidden_size, intermediate_size, hidden_dropout_prob, hidden_act, layer_norm_eps)
    def forward(self, hidden_states, attention_mask):
        attention_output = self.multi_head_attention(hidden_states, attention_mask)
        feedforward_output = self.feed_forward(attention_output)
        return feedforward_output

class TransformerEncoder(nn.Module):
    def __init__(self, n_layers=2, n_heads=2, hidden_size=64, inner_size=256, hidden_dropout_prob=0.5, attn_dropout_prob=0.5, hidden_act='gelu', layer_norm_eps=1e-12):
        super(TransformerEncoder, self).__init__()
        layer = TransformerLayer(n_heads, hidden_size, inner_size, hidden_dropout_prob, attn_dropout_prob, hidden_act, layer_norm_eps)
        self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(n_layers)])

    def forward(self, hidden_states, attention_mask, output_all_encoded_layers=True):
        """
        Args:
            hidden_states (torch.Tensor): the input of the TransformerEncoder
            attention_mask (torch.Tensor): the attention mask for the input hidden_states
            output_all_encoded_layers (Bool): whether output all transformer layers' output
        Returns:
            all_encoder_layers (list): if output_all_encoded_layers is True, return a list consists of all transformer
            layers' output, otherwise return a list only consists of the output of last transformer layer.
        """
        all_encoder_layers = []
        for layer_module in self.layer:
            hidden_states = layer_module(hidden_states, attention_mask)
            if output_all_encoded_layers:
                all_encoder_layers.append(hidden_states)
        if not output_all_encoded_layers:
            all_encoder_layers.append(hidden_states)
        return all_encoder_layers