File size: 3,474 Bytes
18c8389 | 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 | import torch
import math
import torch.nn as nn
from torch import Tensor
class TokenEmbedding(nn.Module):
"""Token Embedding lookup layer."""
def __init__(self, vocab_size: int, d_model: int):
super().__init__()
self.embedding = nn.Embedding(vocab_size, d_model, padding_idx=0)
self.d_model = d_model
# scale ตาม "Attention is All You Need" paper
nn.init.normal_(self.embedding.weight, mean=0, std=d_model ** -0.5)
# reset padding_idx เป็น zeros หลัง จาก init
nn.init.constant_(self.embedding.weight.data[0], 0)
def forward(self, input_ids: Tensor) -> Tensor:
# scale embedding ด้วย √d_model ให้ magnitude ไม่จม PE
return self.embedding(input_ids) * math.sqrt(self.d_model) # scale embedding by sqrt(d_model)
class PositionalEncoding(nn.Module):
"""Sinusoidal positional encoding."""
def __init__(self, d_model: int, max_seq_len: int = 512, dropout: float = 0.1):
super().__init__()
self.dropout = nn.Dropout(p=dropout)
# สร้าง PE matrix ขนาด (max_seq_len, d_model)
pe = torch.zeros(max_seq_len, d_model)
position = torch.arange(0, max_seq_len, dtype=torch.float).unsqueeze(1).float() # (max_seq_len, 1)
# div_term: 1 / 10000^(2i/d_model) — คำนวณใน log-space เพื่อ numerical stability
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) # even dimensions
pe[:, 1::2] = torch.cos(position * div_term) # odd dimensions
# register_buffer: ไม่ใช่ parameter (ไม่ถูก update) แต่ติดไปกับ model.state_dict()
# unsqueeze(0) → (1, T, d_model) เพื่อ broadcast กับ batch
self.register_buffer('pe', pe.unsqueeze(0))
def forward(self, x: Tensor) -> Tensor:
# x shape: (B, T, d_model)
# Adds PE to input tensor and applies dropout
x = x+ self.pe[:, :x.size(1), :]
return self.dropout(x)
class ThaiEmbedding(nn.Module):
"""Token + positional embedding รวมกัน พร้อม layer norm."""
def __init__(self, vocab_size: int, d_model: int,
max_seq_len: int = 512, dropout: float = 0.1):
super().__init__()
self.token_emb = TokenEmbedding(vocab_size, d_model)
self.pos_enc = PositionalEncoding(d_model, max_seq_len, dropout)
self.norm = nn.LayerNorm(d_model)
def forward(self, input_ids: Tensor) -> Tensor:
x = self.token_emb(input_ids) # (B, T, d_model)
x = self.pos_enc(x) # บวก PE + dropout
return self.norm(x) # stabilize ก่อนเข้า transformer blocks
if __name__ == "__main__":
emb = ThaiEmbedding(vocab_size=32000, d_model=256)
emb.eval()
ids = torch.randint(1, 32000, (2, 64)) # batch=2, seq_len=64
out = emb(ids)
assert out.shape == (2, 64, 256)
assert not torch.isnan(out).any()
# padding token (id=0) ต้องได้ PE อย่างเดียว ไม่มี token embedding
pad_ids = torch.zeros(1, 10, dtype=torch.long)
pad_out = emb.token_emb(pad_ids)
assert pad_out.abs().sum() == 0, "padding token ต้องเป็น zero vector"
print("embedding OK") |