-
Notifications
You must be signed in to change notification settings - Fork 0
/
UT_model.py
192 lines (150 loc) · 8.25 KB
/
UT_model.py
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 math
import logging
import torch
import torch.nn as nn
import torch.nn.functional as F
from ACT import ACTModule
logger = logging.getLogger(__name__)
class MultiHeadCausalAttention(nn.Module):
def __init__(self, embed_size, heads):
super(MultiHeadCausalAttention, self).__init__()
self.embed_size = embed_size
self.heads = heads
# Ensure the embedding size is divisible by the number of heads for equal division
assert embed_size % heads == 0, "Embedding size needs to be divisible by heads"
# Define linear transformations for queries, keys, and values
self.values = nn.Linear(embed_size, embed_size)
self.keys = nn.Linear(embed_size, embed_size)
self.queries = nn.Linear(embed_size, embed_size)
# Output projection layer
self.proj_out = nn.Linear(embed_size, embed_size)
def forward(self, x):
# Extract dimensions: batch size, sequence length, and embedding size
batch_size, seq_len, embed_size = x.size()
# Process queries, keys, values: split into 'heads' number of heads, each with head dimension (embed_size/heads)
# Using .view we divide queries, keys and values for each head
## Before transpose: [batch_size, seq_len, heads, embed_size/heads] | After transpose: [batch_size, heads, seq_len, embed_size/heads]
all_queries = self.queries(x).view(batch_size, seq_len, self.heads, embed_size//self.heads).transpose(1,2)
all_keys = self.keys(x).view(batch_size, seq_len, self.heads, embed_size//self.heads).transpose(1,2)
all_values = self.values(x).view(batch_size, seq_len, self.heads, embed_size//self.heads).transpose(1,2)
queries_keys = (all_queries @ all_keys.transpose(-1,-2)) * (1/math.sqrt(all_keys.size(-1)))
# Create a causal mask to mask out future tokens (prevent attending to future positions)
causal_mask = torch.triu(torch.ones(seq_len, seq_len) * float('-inf'), diagonal=1).to(queries_keys.device)
# Apply the causal mask by adding it to the scaled dot-product scores
masked_queries_keys = queries_keys + causal_mask[None, None, :, :]
attn_score = F.softmax(masked_queries_keys, dim=-1)
out = (attn_score @ all_values).transpose(1,2).contiguous().view(batch_size, seq_len, embed_size)
# Apply the final linear projection layer
out = self.proj_out(out) # Transform back to original embed size
return out
class TransitionFunction(nn.Module):
"""TransitionFunction Block is the position-wise feed-forward, just becuase it's named so
in the universal tranformer, I adopted it here"""
def __init__(self, embed_size, ff_size):
super().__init__()
self.fc1 = nn.Linear(embed_size, ff_size)
self.fc2 = nn.Linear(ff_size, embed_size)
def forward(self, x):
x = F.relu(self.fc1(x))
x = self.fc2(x)
return x
class GPTDecoderLayer(nn.Module):
def __init__(self, embed_size, heads, ff_size, attn_dropout_rate=0.1, resid_dropout_rate=0.1):
super().__init__()
# Pre-attention layer normalisation
self.norm_attn = nn.LayerNorm(embed_size)
# Pre-feed-forward layer normalisation
self.norm_ff = nn.LayerNorm(embed_size)
# Causal attention mechanism
self.attn_dropout = nn.Dropout(attn_dropout_rate)
self.attention = MultiHeadCausalAttention(embed_size, heads)
# Feed-forward network
self.feed_forward = TransitionFunction(embed_size, ff_size)
#residual dropout
self.resid_dropout = nn.Dropout(resid_dropout_rate)
def forward(self, x):
attn = self.attn_dropout(self.attention(self.norm_attn(x)))
x = self.resid_dropout(x + attn)
ff = self.feed_forward(self.norm_ff(x))
x = self.resid_dropout(x + ff)
return x
class UTModel(nn.Module):
def __init__(self, vocab_size, max_context, embed_size, ff_size, num_layers, num_heads, act=True, embed_dropout_rate=0.1, resid_dropout_rate=0.1, attn_dropout_rate=0.1):
super(UTModel, self).__init__()
self.embedding = nn.Embedding(vocab_size, embed_size)
self.embed_dropout = nn.Dropout(embed_dropout_rate)
self.act = act
self.max_context = max_context
# Below I fixed time encoding length to 1000, I don't to test beyond this limit
assert self.max_context < 10000, "Please modify so max_contextcan be less than 10000, just becuase of time encoding limit!. "
self.time_encoding = nn.Parameter(torch.ones(1, 10000, embed_size))
if self.act:
self.position_encoding = nn.Parameter(torch.ones(1, num_layers, embed_size))
if self.act:
self.transformation_fn = GPTDecoderLayer(embed_size, num_heads, ff_size, attn_dropout_rate=attn_dropout_rate, resid_dropout_rate=resid_dropout_rate)
self.act_module = ACTModule(embed_size, max_hop=num_layers)
else:
# Create a ModuleList of GPTDecoderLayer instances.
self.layers = nn.ModuleList([GPTDecoderLayer(embed_size, num_heads, ff_size) for _ in range(num_layers)])
self.unembedding = nn.Linear(embed_size, vocab_size)
self.apply(self._init_weights)
logger.info("Number of parameters: %e",sum(p.numel() for p in self.parameters()))
def _init_weights(self, module):
if isinstance(module, (nn.Linear, nn.Embedding)):
nn.init.normal_(module.weight, std=0.02)
if isinstance(module, (nn.Linear)):
if module.bias is not None:
if not getattr(module.bias, "_no_reinit", False):
nn.init.zeros_(module.bias)
elif isinstance(module, nn.LayerNorm):
module.weight.data.fill_(1.0)
module.bias.data.zero_()
def UT_optimizer(self, train_config):
decay = set()
whitelist_weight_modules = (nn.Linear)
for mn, m in self.named_modules():
for pn, _ in m.named_parameters():
fpn = "%s.%s" %(mn, pn) if mn else pn
if pn.endswith('weight') and isinstance(m, whitelist_weight_modules):
decay.add(fpn)
param_dict = {pn: p for pn, p in self.named_parameters()}
optim_groups = [
{"params":[param_dict[pn] for pn in sorted(list(decay))], "weight_decay": train_config.weight_decay},
{"params":[param_dict[pn] for pn in param_dict.keys() if pn not in sorted(list(decay))], "weight_decay": 0.0}
]
optimizer = torch.optim.AdamW(optim_groups, lr=train_config.learning_rate, betas=train_config.AdamwBetas)
return optimizer
def forward(self, input_ids, targets=None):
x = self.embed_dropout(self.embedding(input_ids))
if not self.act:
x = x + self.time_encoding[:,:input_ids.shape[1],:]
if self.act:
x, (remainders, n_updates) = self.act_module(x, self.time_encoding, self.position_encoding, self.transformation_fn)
else:
for layer in self.layers:
x = layer(x)
logits = self.unembedding(x)
loss=None
if targets is not None:
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
return (logits, loss, (remainders, n_updates)) if self.act else (logits, loss)
def generate(self, idx, num_tokens):
tokens_len = idx.shape[1]
token_ponder_time = {}
generate_ponder_time = []
for i in range(num_tokens):
block_idx = idx[: , -self.max_context:]
if self.act:
logits, _, (_, n_updates) = self.forward(block_idx)
generate_ponder_time.append(n_updates.mean().item())
else:
logits, _ = self.forward(block_idx)
logits = logits.reshape(1, tokens_len, -1)[:,-1,:]
probs = F.softmax(logits, dim=1)
next_token = torch.multinomial(probs, num_samples=1)
idx = torch.cat((idx, next_token), dim=1)
if tokens_len < self.max_context:
tokens_len+=1
if self.act:
token_ponder_time[i] = (next_token.item(), n_updates.mean().item())
return (idx, sum(generate_ponder_time)/len(generate_ponder_time), token_ponder_time) if self.act else idx