Repository navigation
Expand file tree
/
Copy pathdecoder.py
More file actions
248 lines (205 loc) · 7.59 KB
/
Copy pathdecoder.py
File metadata and controls
248 lines (205 loc) · 7.59 KB
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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
import json
with open('config.json', 'r', encoding='utf-8') as file:
params = json.load(file)
# required parameters
block_size = params['block_size']
d_model = params['d_model']
n_head = params['n_heads']
n_layers = params['n_layers']
learning_rate = params['learning_rate']
dropout = params['dropout']
norm_eps = params['norm_eps']
import torch
import torch.nn as nn
from torch.nn import functional as F
device = 'cuda' if torch.cuda.is_available() else 'cpu'
class RMSNorm(nn.Module):
"""
using RMS normalization instead of layer norm for better stability while training
Args:
dim (int): dimensions to apply normalizations
eps (float): small floating point no to prevent division by zero
"""
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def _norm(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
def forward(self, x):
output = self._norm(x.float()).type_as(x)
return output * self.weight
class SingleHead(nn.Module):
"""
single head of self attention
generates three copies of 'x' -> Wk, Wq, Wv
`attention matrix = '[Wq . Wk.transpose()]/dim(Wk)**-0.5'`
if masking is True:
[[9, 2, 5, 9, 6] [[9, 0, 0, 0, 0]
[6, 9, 2, 6, 2] [6, 9, 0, 0, 0]
[1, 7, 9, 6, 2] --> [1, 7, 9, 0, 0]
[8, 3, 2, 9, 2] [8, 3, 2, 9, 0]
[2, 1, 2, 6, 9]] [2, 1, 2, 6, 9]]
`outputs = softmax(attention_matrix) . Wv`
Args:
head_size (int): size of each head for attention
d_model (int): model size dimensions
block_size (int): context size
dropout (float): % of dropout connections
"""
def __init__(self,
head_size: int,
d_model: int,
block_size: int,
dropout: float):
super().__init__()
self.key = nn.Linear(d_model, head_size, bias=True)
self.query = nn.Linear(d_model, head_size, bias=True)
self.value = nn.Linear(d_model, head_size, bias=True)
self.dropout = nn.Dropout(dropout)
self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))
def forward(self, x: torch.Tensor, mask: bool= False):
B, T, C = x.shape
key = self.key(x)
query = self.query(x)
scores = torch.matmul(query ,key.transpose(-2, -1)) / (key.shape[-1]**-0.5)
if mask is True:
scores = scores.masked_fill(self.tril[:T, :T] == 0, float('-inf'))
att_mat = F.softmax(scores, dim=-1)
att_mat = self.dropout(att_mat)
value = self.value(x)
output = torch.matmul(att_mat, value)
return output
class MultiHeadAttention(nn.Module):
"""
multiple head of self-attention running in parallel
outputs are then concatenated into a single matrix
Args:
d_model (int): model size dimensions
block_size (int): context size
n_head (int): no of attention head in parallel
dropout (float): % of dropout connections
"""
def __init__(self,
d_model: int,
block_size: int,
n_head : int,
dropout: float):
head_size = d_model // n_head
super().__init__()
self.heads = nn.ModuleList([SingleHead(d_model=d_model, dropout=dropout, block_size=block_size, head_size=head_size) for _ in range(n_head)])
self.projection = nn.Linear(head_size * n_head, d_model)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor, mask: bool):
out = torch.cat([h(x, mask) for h in self.heads], dim=-1)
out = self.dropout(self.projection(out))
del x
return out
class FeedForward(nn.Module):
"""
simple feedforward layer using two linear layers and LeakyRelu as activation function
`LeakyRelu(x) = {x, if x >= 0 ; neg_slope*x, otherwise}`
`LeakyReLU(x) = max(0,x) + negative_slope*min(0,x)`
expansion factor = 5
Args:
d_model (int): model size dimensions
dropout (float): % of dropout connections
"""
def __init__(self, d_model, dropout):
super().__init__()
self.net = nn.Sequential(
nn.Linear(d_model, 5 * d_model),
nn.LeakyReLU(),
nn.Linear(5 * d_model, d_model),
nn.Dropout(dropout),
)
def forward(self, x: torch.Tensor):
return self.net(x)
class DecoderBlock(nn.Module):
"""
simple decoder block with one layer of masked attention followed by self attention layer
and feedforward layer finally
Args:
block_size (int): context size
n_head (int): no of attention head in parallel
norm_eps (float): small floating point no to prevent division by zero
dropout (float): % of dropout connections
"""
def __init__(self, d_model: int,
block_size: int,
n_head: int,
norm_eps: float,
dropout: float):
super().__init__()
self.self_att = MultiHeadAttention(n_head=n_head, d_model=d_model, dropout=dropout, block_size=block_size)
self.ffwd = FeedForward(d_model, dropout)
self.dropout = nn.Dropout(dropout)
self.norm = RMSNorm(d_model, eps=norm_eps)
def forward(self, x: torch.Tensor):
x_out = self.self_att(self.norm(x), mask=True)
x_out = x + self.dropout(x_out)
del x
x = self.self_att(self.norm(x_out), mask=False)
x = x_out + self.dropout(x)
del x_out
x_out = self.ffwd(self.norm(x))
x_out = x + self.dropout(x_out)
del x
return x_out
class GPT(nn.Module):
def __init__(self, vocab_size: int):
super().__init__()
self.block_size = block_size
self.tok_embed = nn.Embedding(vocab_size, d_model)
self.pos_embed = nn.Embedding(block_size, d_model)
self.decoder = nn.Sequential(*[DecoderBlock(n_head=n_head, d_model=d_model, dropout=dropout, norm_eps=norm_eps, block_size=block_size) for _ in range(n_layers)])
self.norm_final = RMSNorm(d_model, eps=norm_eps)
self.linear_final = nn.Linear(d_model, vocab_size)
self.dropout = nn.Dropout(dropout)
self.apply(self._init_weights)
def _init_weights(self, module):
if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
torch.nn.init.zeros_(module.bias.data)
elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(self, idx, targets=None):
B, T = idx.shape
toks = self.tok_embed(idx)
pos = self.pos_embed(torch.arange(T, device=device))
x = toks + pos
x = self.decoder(x)
logits = self.linear_final(self.norm_final(x))
if targets is None:
loss = None
else:
B, T, C = logits.shape
logits = logits.view(B*T, C)
targets = targets.view(B*T)
loss = F.cross_entropy(logits, targets)
return logits, loss
@torch.no_grad()
def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None):
"""
generate new tokens using the trained model
Args:
- idx (Tensor): input tensor representing initial token indices
- max_new_tokens (int): max no of new tokens to generate
- temperature (float): softmax temperature for sampling
- top_k (int): no of top tokens to consider in sampling
Returns:
- generated_tokens (list): list of generated token indices
"""
self.eval()
for _ in range(max_new_tokens):
idx_cond = idx if idx.size(1) <= self.block_size else idx[:, -self.block_size:]
logits, _ = self(idx_cond)
logits = logits[:, -1, :] / temperature
if top_k is not None:
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
logits[logits < v[:, [-1]]] = -float('Inf')
probs = F.softmax(logits, dim=-1)
idx_next = torch.multinomial(probs, num_samples=1)
idx = torch.cat((idx, idx_next), dim=1)
return idx