File size: 5,719 Bytes
8920c6e |
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 |
# -*- coding: utf-8 -*-
"""
Created on Fri Sep 13 19:23:54 2024
This script defines the LWM model architecture.
@author: Sadjad Alikhani
"""
#%%
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
#%%
class LayerNormalization(nn.Module):
def __init__(self, d_model: int, eps: float = 1e-6) -> None:
super().__init__()
self.eps = eps
self.alpha = nn.Parameter(torch.ones(d_model))
self.bias = nn.Parameter(torch.zeros(d_model))
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.bias
class Embedding(nn.Module):
def __init__(self, element_length, d_model, max_len=513):
super().__init__()
self.element_length = element_length
self.d_model = d_model
self.proj = nn.Linear(element_length, d_model)
self.pos_embed = nn.Embedding(max_len, d_model)
self.norm = LayerNormalization(d_model)
def forward(self, x):
seq_len = x.size(1)
pos = torch.arange(seq_len, dtype=torch.long, device=x.device)
pos_encodings = self.pos_embed(pos)
tok_emb = self.proj(x.float())
embedding = tok_emb + pos_encodings
return self.norm(embedding)
class ScaledDotProductAttention(nn.Module):
def __init__(self, d_k):
super().__init__()
self.d_k = d_k
def forward(self, Q, K, V):
scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(self.d_k)
attn = F.softmax(scores, dim=-1)
context = torch.matmul(attn, V)
return context, attn
class MultiHeadAttention(nn.Module):
def __init__(self, d_model, n_heads, dropout):
super().__init__()
self.d_k = d_model // n_heads
self.d_v = d_model // n_heads
self.n_heads = n_heads
self.W_Q = nn.Linear(d_model, self.d_k * n_heads)
self.W_K = nn.Linear(d_model, self.d_k * n_heads)
self.W_V = nn.Linear(d_model, self.d_v * n_heads)
self.linear = nn.Linear(n_heads * self.d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.scaled_dot_attn = ScaledDotProductAttention(self.d_k)
def forward(self, Q, K, V):
residual, batch_size = Q, Q.size(0)
q_s = self.W_Q(Q).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
k_s = self.W_K(K).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
v_s = self.W_V(V).view(batch_size, -1, self.n_heads, self.d_v).transpose(1, 2)
context, attn = self.scaled_dot_attn(q_s, k_s, v_s)
output = context.transpose(1, 2).contiguous().view(batch_size, -1, self.n_heads * self.d_v)
output = self.linear(output)
return residual + self.dropout(output), attn
class PoswiseFeedForwardNet(nn.Module):
def __init__(self, d_model, d_ff, dropout):
super().__init__()
self.fc1 = nn.Linear(d_model, d_ff)
self.fc2 = nn.Linear(d_ff, d_model)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
return self.fc2(self.dropout(F.relu(self.fc1(x))))
class EncoderLayer(nn.Module):
def __init__(self, d_model, n_heads, d_ff, dropout):
super().__init__()
self.enc_self_attn = MultiHeadAttention(d_model, n_heads, dropout)
self.pos_ffn = PoswiseFeedForwardNet(d_model, d_ff, dropout)
self.norm1 = LayerNormalization(d_model)
self.norm2 = LayerNormalization(d_model)
def forward(self, enc_inputs):
# Self-Attention with Add & Norm
attn_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs)
attn_outputs = self.norm1(enc_inputs + attn_outputs) # Add & Norm
# Feed-Forward with Add & Norm
ff_outputs = self.pos_ffn(attn_outputs)
enc_outputs = self.norm2(attn_outputs + ff_outputs) # Add & Norm
return enc_outputs, attn
class lwm(nn.Module):
def __init__(self, element_length=32, d_model=128, n_layers=12, max_len=513, n_heads=8, dropout=0.1):
super().__init__()
self.embedding = Embedding(element_length, d_model, max_len)
self.layers = nn.ModuleList(
[EncoderLayer(d_model, n_heads, d_model*4, dropout) for _ in range(n_layers)]
)
self.linear = nn.Linear(d_model, d_model)
self.norm = LayerNormalization(d_model)
embed_weight = self.embedding.proj.weight
_, n_dim = embed_weight.size()
self.decoder = nn.Linear(d_model, n_dim, bias=False)
self.decoder_bias = nn.Parameter(torch.zeros(n_dim))
@classmethod
def from_pretrained(cls, ckpt_name='model_weights.pth', device='cuda'):
model = cls().to(device)
model.load_state_dict(torch.load(ckpt_name, map_location=device))
print(f"Model loaded successfully from {ckpt_name}")
return model
def forward(self, input_ids, masked_pos=None):
# Step 1: Embedding
output = self.embedding(input_ids)
attention_maps = []
# Step 2: Pass through Encoder Layers
for layer in self.layers:
output, attn = layer(output)
attention_maps.append(attn)
# If masked_pos is provided, perform masked token prediction
if masked_pos is not None:
masked_pos = masked_pos.long()[:, :, None].expand(-1, -1, output.size(-1))
h_masked = torch.gather(output, 1, masked_pos)
h_masked = self.norm(F.relu(self.linear(h_masked)))
logits_lm = self.decoder(h_masked) + self.decoder_bias
return logits_lm, output, attention_maps
else:
return output, attention_maps
|