import torch import torch.nn as nn import math class InputEmbedding(nn.Module): def __init__(self, d_model: int, vocab_size: int) -> None: super().__init__() self.d_model = d_model self.vocab_size = vocab_size self.embedding = nn.Embedding(self.vocab_size, d_model) def forward(self, x): return self.embedding(x)*math.sqrt(self.d_model) class PositionalEncoding(nn.Module): def __init__(self, d_model: int, sequence_length: int, dropout: float) -> None: super().__init__() self.d_model = d_model self.sequence_length = sequence_length self.dropout = nn.Dropout(dropout) pe = torch.zeros(sequence_length, d_model) position = torch.arange(0, sequence_length, dtype=torch.float).unsqueeze(1) frequency_term = torch.exp(torch.arange(0, d_model, 2, dtype=torch.float) * (-math.log(10000.0) / d_model)) pe[:, 0::2] = torch.sin(position*frequency_term) pe[:, 1::2] = torch.cos(position*frequency_term) pe = pe.unsqueeze(0) # add batch dimention self.register_buffer('pe', pe) def forward(self, x): pe = self.pe.detach() # Detach the positional encoding tensor x = x + pe[:, :x.shape[1], :] return self.dropout(x) class LayerNormalization(nn.Module): def __init__(self, eps: float = 10**-6) -> None: super().__init__() self.eps = eps self.alpha = nn.Parameter(torch.ones(1)) self.beta = nn.Parameter(torch.zeros(1)) def forward(self, x): mean = x.mean(dim = -1, keepdim = True) std = x.std(dim = -1, keepdim = True) return self.alpha*(x - mean)/(std + self.eps) + self.beta class FeedForwardBlock(nn.Module): def __init__(self, d_model: int, dff: int, dropout: float): super().__init__() self.linear_1 = nn.Linear(d_model, dff) self.dropout = nn.Dropout(dropout) self.linear_2 = nn.Linear(dff, d_model) def forward(self, x): return self.linear_2(self.dropout(torch.relu(self.linear_1(x)))) class MultiheadAttentionBlock(nn.Module): def __init__(self, d_model: int, num_heads: int, dropout: float): super().__init__() self.d_model = d_model self.num_heads = num_heads assert d_model%num_heads == 0, "num heads does not divide d_model" self.d_k = d_model // num_heads self.Wq = nn.Linear(d_model, d_model) # vec to query self.Wk = nn.Linear(d_model, d_model) # vec to key self.Wv = nn.Linear(d_model, d_model) # vec to value self.dropout = nn.Dropout(dropout) self.Wo = nn.Linear(d_model, d_model) @staticmethod def attention(query, key, value, dropout: nn.Dropout, mask = None): # attention matrix scores = torch.matmul(query, key.transpose(-2, -1))/math.sqrt(query.shape[-1]) if mask is not None: scores = scores.masked_fill(mask == 0, -1e9) scores = torch.softmax(scores, dim = -1) if dropout is not None: scores = dropout(scores) return torch.matmul(scores, value), scores # return the output of the head as well as attention matrix for visualization def forward(self, q, k, v, mask): Q = self.Wq(q) K = self.Wk(k) V = self.Wv(v) # divide the input vectors into different heads Q = Q.view(Q.shape[0], Q.shape[1], self.num_heads, self.d_k).transpose(1,2) K = K.view(K.shape[0], K.shape[1], self.num_heads, self.d_k).transpose(1,2) V = V.view(V.shape[0], V.shape[1], self.num_heads, self.d_k).transpose(1,2) x, self.attention_scores = MultiheadAttentionBlock.attention(Q, K, V, self.dropout, mask) # print(f"shapes of attentions: {x.shape[0]} {x.shape[1]} {x.shape[2]} {x.shape[3]}") x = x.transpose(1,2).contiguous().view(x.shape[0], -1, self.num_heads*self.d_k) return self.Wo(x) class ResidualConnection(nn.Module): def __init__(self, dropout: float): super().__init__() self.norm = LayerNormalization() self.dropout = nn.Dropout(dropout) def forward(self, x, sublayer): return x + self.dropout(sublayer(self.norm(x))) class EncoderBlock(nn.ModuleList): def __init__(self, self_attention_block: MultiheadAttentionBlock, feed_forward_block: FeedForwardBlock, dropout: float): super().__init__() self.self_attention_block = self_attention_block self.feed_forward_block = feed_forward_block self.residual_connections = nn.ModuleList([ResidualConnection(dropout) for _ in range(2)]) def forward(self, x, src_mask): x = self.residual_connections[0](x, lambda x: self.self_attention_block(x, x, x, src_mask)) x = self.residual_connections[1](x, lambda x: self.feed_forward_block(x)) return x class Encoder(nn.Module): def __init__(self, layers: nn.ModuleList) -> None: super().__init__() self.layers = layers self.norm = LayerNormalization() def forward(self, x, src_mask): for layer in self.layers: x = layer(x, src_mask) return self.norm(x) class ProjectionLayer(nn.Module): def __init__(self, d_model: int, vocab_size: int): super().__init__() self.linear = nn.Linear(d_model, vocab_size) def forward(self, x): return torch.log_softmax(self.linear(x), dim = -1) class DecoderOnlyTransformer(nn.Module): def __init__(self, encoder: Encoder, tgt_embed: InputEmbedding, tgt_pos: PositionalEncoding, projection_layer: ProjectionLayer) -> None: super().__init__() self.encoder = encoder self.tgt_embed = tgt_embed self.tgt_pos = tgt_pos self.projection_layer = projection_layer def decode(self, tgt, tgt_mask): tgt = self.tgt_embed(tgt) tgt = self.tgt_pos(tgt) return self.encoder(tgt, tgt_mask) def project(self, x): return self.projection_layer(x) def build_decoder_only_transformer(vocab_size: int, seq_len: int, d_model:int = 512, N:int = 6, h:int = 8, dropout:float = 0.1, dff:int = 2048): # embedding layers embed = InputEmbedding(d_model, vocab_size) # positional encodings pos = PositionalEncoding(d_model, seq_len, dropout) encoder_blocks = [] for _ in range(N): encoder_self_attention_block = MultiheadAttentionBlock(d_model, h, dropout) feed_fwd_block = FeedForwardBlock(d_model, dff, dropout) encoder_block = EncoderBlock(encoder_self_attention_block, feed_fwd_block, dropout) encoder_blocks.append(encoder_block) encoder = Encoder(nn.ModuleList(encoder_blocks)) projection_layer = ProjectionLayer(d_model, vocab_size) transformer = DecoderOnlyTransformer(encoder, embed, pos, projection_layer) for p in transformer.parameters(): if p.dim() > 1: nn.init.xavier_uniform_(p) return transformer