Spaces:
Sleeping
Sleeping
File size: 7,491 Bytes
c06baa5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 |
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
|