File size: 6,660 Bytes
c6ee1cf | 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 | import torch.nn as nn
import torch
class RoPE(nn.Module):
def __init__(self, dim, max_seq_len=1024):
super().__init__()
theta = 10000
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
positions = torch.arange(max_seq_len)
angles = torch.outer(positions, freqs)
self.register_buffer(
"cos",
angles.cos()
)
self.register_buffer(
"sin",
angles.sin()
)
def rotate_half(self, x):
x1 = x[..., ::2]
x2 = x[..., 1::2]
return torch.stack(
(-x2,x1),
dim=-1
).flatten(-2)
def forward(self, x):
seq_len = x.shape[-2]
cos = self.cos[:seq_len]
sin = self.sin[:seq_len]
cos = torch.repeat_interleave(cos, 2, dim=-1)
sin = torch.repeat_interleave(sin, 2, dim=-1)
return x*cos + self.rotate_half(x)*sin
class MHA(nn.Module):
def __init__(self, d_model, num_heads):
super().__init__()
super().__init__()
self.num_heads = num_heads
self.d_model = d_model
self.Q = nn.Linear(d_model, d_model)
self.K = nn.Linear(d_model, d_model)
self.V = nn.Linear(d_model, d_model)
self.W_o = nn.Linear(d_model, d_model)
self.d_k = d_model // num_heads
self.rope = RoPE(self.d_k)
def forward(self, x, kv_input=None, mask=None):
batch_size = x.size(0)
# x.shape = (2, 3, 64) 64 dictates that there will be 64 Q, K and V PER TOKEN
if kv_input is None:
kv_input = x
Q = self.Q(x) # we do q_1 = XWq = 'W_1'*'e1' + ... + 'W_64'*'e64' W_i being the weight at i ... until we comp q_64, if no kv_input, it means we're doing encoding
K = self.K(kv_input) # ,, ,,
V = self.V(kv_input) # ,, ,,
# We do this to ensure all Q,K and V's matrices have access to the input embeddings
# Manual length extraction to avoid mismatch in size from encoder/decoder
q_seq_length = x.shape[1]
kv_seq_length = kv_input.shape[1]
# We then split all this data through different heads so each one goes for a raneg window for each token (split ranges -> specialized heads)
Q = Q.view(batch_size, q_seq_length, self.num_heads, self.d_k)
K = K.view(batch_size, kv_seq_length, self.num_heads, self.d_k)
V = V.view(batch_size, kv_seq_length, self.num_heads, self.d_k)
# So we get (batch, heads, seq_len, d_k) last 2 dims are the ones that get multiplied by Pytorch
# The reason we put seq_len, d_k is because we gotta multiply them indep. of how many heads or batches, what matters is the consistent output per head per batch
# Also, since we're computing relationships between tokens, there must be seq_len and d_k, so we can compare all tokens with all tokens for every head in every batch
Q = Q.transpose(1,2)
K = K.transpose(1,2)
V = V.transpose(1,2)
Q = self.rope(Q)
K = self.rope(K)
scores = Q @ K.transpose(-2, -1) / self.d_k**0.5 #we swap d_k with seq_len so we get #seq_len * #seq_len matrix as a result
if mask is not None:
scores += mask
attn_weights = torch.softmax(scores, dim=-1)
output = attn_weights @ V # contiguous vector of shape (batch, heads, seq, d_k)
output = output.transpose(1,2).contiguous().view(batch_size, q_seq_length, self.d_model) # transposing makes it breaks contiguousy, so we add .contiguous() and then .view() to reshape it with concatenation
output = self.W_o(output) # expressing the new shape correctly accross heads
return output
class MLP(nn.Module):
def __init__(self, d_model, d_ff):
super().__init__()
self.fc1 = nn.Linear(d_model, d_ff)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(d_ff, d_model)
def forward(self, x):
x = self.fc1(x)
x = self.relu(x)
x = self.fc2(x)
return x
class EncoderBlock(nn.Module):
def __init__(self, d_model, num_heads, d_ff):
super().__init__()
self.mha = MHA(d_model, num_heads)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.mlp = MLP(d_model, d_ff)
def forward(self, x):
normed = self.norm1(x)
attn_out = self.mha(normed)
x = x + attn_out
normed2 = self.norm2(x)
mlp_out = self.mlp(normed2)
x = x + mlp_out
return x
class DecoderBlock(nn.Module):
def __init__(self, d_model, num_heads, d_ff):
super().__init__()
self.self_attn = MHA(d_model, num_heads)
self.cross_attn = MHA(d_model, num_heads)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
self.mlp = MLP(d_model, d_ff)
def forward(self, x, encoder_output):
causal_mask = torch.triu(torch.full((x.shape[1], x.shape[1]), float('-inf'), device=x.device), diagonal=1)
normed = self.norm1(x)
attn_out = self.self_attn(normed, mask = causal_mask)
x = x + attn_out
normed2 = self.norm2(x)
cross_out = self.cross_attn(normed2, encoder_output)
x = x + cross_out
normed3 = self.norm3(x)
mlp_out = self.mlp(normed3)
x = x + mlp_out
return x
class Transformer(nn.Module):
def __init__(self, src_vocab_size, tgt_vocab_size, d_model, num_heads, d_ff):
super().__init__()
self.src_embedding = nn.Embedding(src_vocab_size, d_model)
self.tgt_embedding = nn.Embedding(tgt_vocab_size, d_model)
self.encoder = nn.ModuleList([EncoderBlock(d_model, num_heads, d_ff) for _ in range(3)])
self.decoder = nn.ModuleList([DecoderBlock(d_model, num_heads, d_ff) for _ in range(3)])
self.output_proj = nn.Linear(d_model, tgt_vocab_size)
def forward(self, src_tokens, tgt_tokens):
src = self.src_embedding(src_tokens)
tgt = self.tgt_embedding(tgt_tokens)
for block in self.encoder:
src = block(src)
encoder_output = src
for block in self.decoder:
tgt = block(tgt, encoder_output)
logits = self.output_proj(tgt)
return logits
|