File size: 7,112 Bytes
5129187
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# model/transformer_explained.py
"""
Tiny Transformer language model (educational).
Components:
 - PositionalEncoding: sinusoidal positional encodings (buffered)
 - MultiHeadSelfAttention: returns attn weights optionally
 - FeedForward: MLP with GELU
 - TransformerBlock: attention + add&norm + FFN + add&norm
 - TinyTransformerLM: token embeddings, pos enc, stacked blocks, LM head
"""

import math
from typing import Optional, Tuple

import torch
import torch.nn as nn
import torch.nn.functional as F


class PositionalEncoding(nn.Module):
    """Sinusoidal positional encoding as in "Attention is All You Need".
    Stored as a buffer (not learned). Adds positional encodings to token embeddings.
    """

    def __init__(self, d_model: int, max_len: int = 2048):
        super().__init__()
        pe = torch.zeros(max_len, d_model)  # (max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)  # (max_len, 1)
        div_term = torch.exp(
            torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)
        )  # (d_model/2,)
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)  # (1, max_len, d_model)
        self.register_buffer("pe", pe)  # not a parameter

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        x: (batch, seq_len, d_model)
        returns: x + pe[:, :seq_len, :]
        """
        seq_len = x.size(1)
        return x + self.pe[:, :seq_len, :].to(x.device)


class MultiHeadSelfAttention(nn.Module):
    """
    Multi-head self-attention.
    Optionally returns attention weights for visualization.

    Input shapes:
      x: (batch, seq_len, d_model)
    Output:
      out: (batch, seq_len, d_model)
    Optional:
      attn: (batch, num_heads, seq_len, seq_len)
    """

    def __init__(self, d_model: int, num_heads: int, dropout: float = 0.0):
        super().__init__()
        assert d_model % num_heads == 0, "d_model must be divisible by num_heads"
        self.d_model = d_model
        self.num_heads = num_heads
        self.d_k = d_model // num_heads
        # single linear for qkv then split
        self.qkv_proj = nn.Linear(d_model, d_model * 3, bias=False)
        self.out_proj = nn.Linear(d_model, d_model, bias=False)
        self.attn_dropout = nn.Dropout(dropout)
        self.softmax = nn.Softmax(dim=-1)

    def forward(
        self, x: torch.Tensor, mask: Optional[torch.Tensor] = None, return_attn: bool = False
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        """
        x: (batch, seq_len, d_model)
        mask: (batch, 1, seq_len, seq_len) or (batch, seq_len) causal mask etc.
        return_attn: if True, also return attention weights
        """
        B, S, D = x.shape
        # project and split into q,k,v
        qkv = self.qkv_proj(x)  # (B, S, 3*D)
        qkv = qkv.view(B, S, 3, self.num_heads, self.d_k)
        q, k, v = qkv.unbind(dim=2)  # each: (B, S, num_heads, d_k)

        # transpose to (B, num_heads, S, d_k)
        q = q.transpose(1, 2)
        k = k.transpose(1, 2)
        v = v.transpose(1, 2)

        # scaled dot-product attention
        # attn_scores: (B, num_heads, S, S)
        attn_scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)

        if mask is not None:
            # mask should be broadcastable to (B, num_heads, S, S)
            attn_scores = attn_scores.masked_fill(mask == 0, float("-inf"))

        attn = self.softmax(attn_scores)  # (B, num_heads, S, S)
        attn = self.attn_dropout(attn)
        # attn @ v -> (B, num_heads, S, d_k)
        out = torch.matmul(attn, v)
        # transpose & combine heads -> (B, S, D)
        out = out.transpose(1, 2).contiguous().view(B, S, D)
        out = self.out_proj(out)  # final linear

        if return_attn:
            return out, attn
        return out, None


class FeedForward(nn.Module):
    """Position-wise feed-forward network."""

    def __init__(self, d_model: int, d_ff: int, dropout: float = 0.1):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(d_model, d_ff),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(d_ff, d_model),
            nn.Dropout(dropout),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x)


class TransformerBlock(nn.Module):
    """A single Transformer block: MHSA -> Add&Norm -> FFN -> Add&Norm"""

    def __init__(self, d_model: int, num_heads: int, d_ff: int, dropout: float = 0.1):
        super().__init__()
        self.ln1 = nn.LayerNorm(d_model)
        self.attn = MultiHeadSelfAttention(d_model, num_heads, dropout)
        self.ln2 = nn.LayerNorm(d_model)
        self.ff = FeedForward(d_model, d_ff, dropout)

    def forward(
        self, x: torch.Tensor, mask: Optional[torch.Tensor] = None, return_attn: bool = False
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        # Pre-norm style: ln -> attn -> add
        z = self.ln1(x)
        attn_out, attn_weights = self.attn(z, mask=mask, return_attn=return_attn)
        x = x + attn_out
        # FFN
        z2 = self.ln2(x)
        ff_out = self.ff(z2)
        x = x + ff_out
        if return_attn:
            return x, attn_weights
        return x, None


class TinyTransformerLM(nn.Module):
    """
    Tiny Transformer language model for educational training/experiments.
    Not tokenizer-dependent; expects token ids.
    """

    def __init__(
        self,
        vocab_size: int,
        d_model: int = 256,
        n_layers: int = 4,
        num_heads: int = 4,
        d_ff: int = 1024,
        max_len: int = 512,
        dropout: float = 0.1,
    ):
        super().__init__()
        self.vocab_size = vocab_size
        self.tok_emb = nn.Embedding(vocab_size, d_model)
        self.pos_enc = PositionalEncoding(d_model, max_len=max_len)
        self.layers = nn.ModuleList(
            [TransformerBlock(d_model, num_heads, d_ff, dropout) for _ in range(n_layers)]
        )
        self.ln_f = nn.LayerNorm(d_model)
        self.head = nn.Linear(d_model, vocab_size, bias=False)  # logits head

    def forward(
        self, input_ids: torch.LongTensor, mask: Optional[torch.Tensor] = None, return_attn_layer: Optional[int] = None
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        """
        input_ids: (B, S)
        returns: logits (B, S, vocab_size)
        if return_attn_layer is an int, it will return attention weights from that layer (heads)
        """
        B, S = input_ids.shape
        x = self.tok_emb(input_ids)  # (B, S, d_model)
        x = self.pos_enc(x)
        attn_weights = None
        for idx, layer in enumerate(self.layers):
            if return_attn_layer is not None and idx == return_attn_layer:
                x, attn_weights = layer(x, mask=mask, return_attn=True)
            else:
                x, _ = layer(x, mask=mask, return_attn=False)
        x = self.ln_f(x)
        logits = self.head(x)  # (B, S, vocab_size)
        return logits, attn_weights