Spaces:
Runtime error
Runtime error
Commit
·
a26f15b
1
Parent(s):
8aedac7
Create utils.py
Browse files
utils.py
ADDED
@@ -0,0 +1,154 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn
|
3 |
+
import lightning.pytorch as pl
|
4 |
+
from torchvision import datasets
|
5 |
+
from torch.nn import functional as F
|
6 |
+
from torch.utils.data import DataLoader, Dataset, random_splitim
|
7 |
+
|
8 |
+
class Head(nn.Module):
|
9 |
+
""" one head of self-attention """
|
10 |
+
|
11 |
+
def __init__(self, head_size):
|
12 |
+
super().__init__()
|
13 |
+
self.key = nn.Linear(n_embd, head_size, bias=False)
|
14 |
+
self.query = nn.Linear(n_embd, head_size, bias=False)
|
15 |
+
self.value = nn.Linear(n_embd, head_size, bias=False)
|
16 |
+
self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))
|
17 |
+
|
18 |
+
self.dropout = nn.Dropout(dropout)
|
19 |
+
|
20 |
+
def forward(self, x):
|
21 |
+
B,T,C = x.shape
|
22 |
+
k = self.key(x) # (B,T,C)
|
23 |
+
q = self.query(x) # (B,T,C)
|
24 |
+
# compute attention scores ("affinities")
|
25 |
+
wei = q @ k.transpose(-2,-1) * C**-0.5 # (B, T, C) @ (B, C, T) -> (B, T, T)
|
26 |
+
wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T)
|
27 |
+
wei = F.softmax(wei, dim=-1) # (B, T, T)
|
28 |
+
wei = self.dropout(wei)
|
29 |
+
# perform the weighted aggregation of the values
|
30 |
+
v = self.value(x) # (B,T,C)
|
31 |
+
out = wei @ v # (B, T, T) @ (B, T, C) -> (B, T, C)
|
32 |
+
return out
|
33 |
+
|
34 |
+
class MultiHeadAttention(nn.Module):
|
35 |
+
""" multiple heads of self-attention in parallel """
|
36 |
+
|
37 |
+
def __init__(self, num_heads, head_size):
|
38 |
+
super().__init__()
|
39 |
+
self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
|
40 |
+
self.proj = nn.Linear(n_embd, n_embd)
|
41 |
+
self.dropout = nn.Dropout(dropout)
|
42 |
+
|
43 |
+
def forward(self, x):
|
44 |
+
out = torch.cat([h(x) for h in self.heads], dim=-1)
|
45 |
+
out = self.dropout(self.proj(out))
|
46 |
+
return out
|
47 |
+
|
48 |
+
class FeedFoward(nn.Module):
|
49 |
+
""" a simple linear layer followed by a non-linearity """
|
50 |
+
|
51 |
+
def __init__(self, n_embd):
|
52 |
+
super().__init__()
|
53 |
+
self.net = nn.Sequential(
|
54 |
+
nn.Linear(n_embd, 4 * n_embd),
|
55 |
+
nn.ReLU(),
|
56 |
+
nn.Linear(4 * n_embd, n_embd),
|
57 |
+
nn.Dropout(dropout),
|
58 |
+
)
|
59 |
+
|
60 |
+
def forward(self, x):
|
61 |
+
return self.net(x)
|
62 |
+
|
63 |
+
class Block(nn.Module):
|
64 |
+
""" Transformer block: communication followed by computation """
|
65 |
+
|
66 |
+
def __init__(self, n_embd, n_head):
|
67 |
+
# n_embd: embedding dimension, n_head: the number of heads we'd like
|
68 |
+
super().__init__()
|
69 |
+
head_size = n_embd // n_head
|
70 |
+
self.sa = MultiHeadAttention(n_head, head_size)
|
71 |
+
self.ffwd = FeedFoward(n_embd)
|
72 |
+
self.ln1 = nn.LayerNorm(n_embd)
|
73 |
+
self.ln2 = nn.LayerNorm(n_embd)
|
74 |
+
|
75 |
+
def forward(self, x):
|
76 |
+
x = x + self.sa(self.ln1(x))
|
77 |
+
x = x + self.ffwd(self.ln2(x))
|
78 |
+
return x
|
79 |
+
|
80 |
+
|
81 |
+
class GPTLanguageModel(nn.Module):
|
82 |
+
def __init__(self):
|
83 |
+
super().__init__()
|
84 |
+
# each token directly reads off the logits for the next token from a lookup table
|
85 |
+
self.token_embedding_table = nn.Embedding(vocab_size, n_embd)
|
86 |
+
self.position_embedding_table = nn.Embedding(block_size, n_embd)
|
87 |
+
self.blocks = nn.Sequential(*[Block(n_embd, n_head=n_head) for _ in range(n_layer)])
|
88 |
+
self.ln_f = nn.LayerNorm(n_embd) # final layer norm
|
89 |
+
self.lm_head = nn.Linear(n_embd, vocab_size)
|
90 |
+
|
91 |
+
def forward(self, idx, targets=None):
|
92 |
+
B, T = idx.shape
|
93 |
+
#print(idx.device)
|
94 |
+
# idx and targets are both (B,T) tensor of integers
|
95 |
+
tok_emb = self.token_embedding_table(idx) # (B,T,C)
|
96 |
+
pos_emb = self.position_embedding_table(torch.arange(T, device=device)) # (T,C)
|
97 |
+
x = tok_emb + pos_emb # (B,T,C)
|
98 |
+
x = self.blocks(x) # (B,T,C)
|
99 |
+
x = self.ln_f(x) # (B,T,C)
|
100 |
+
logits = self.lm_head(x) # (B,T,vocab_size)
|
101 |
+
|
102 |
+
if targets is None:
|
103 |
+
loss = None
|
104 |
+
else:
|
105 |
+
B, T, C = logits.shape
|
106 |
+
logits = logits.view(B*T, C)
|
107 |
+
targets = targets.view(B*T)
|
108 |
+
loss = F.cross_entropy(logits, targets)
|
109 |
+
|
110 |
+
return logits, loss
|
111 |
+
|
112 |
+
def generate(self, idx, max_new_tokens):
|
113 |
+
# idx is (B, T) array of indices in the current context
|
114 |
+
for _ in range(max_new_tokens):
|
115 |
+
# crop idx to the last block_size tokens
|
116 |
+
idx_cond = idx[:, -block_size:].to(device)
|
117 |
+
# get the predictions
|
118 |
+
logits, loss = self(idx_cond)
|
119 |
+
# focus only on the last time step
|
120 |
+
logits = logits[:, -1, :] # becomes (B, C)
|
121 |
+
# apply softmax to get probabilities
|
122 |
+
probs = F.softmax(logits, dim=-1) # (B, C)
|
123 |
+
# sample from the distribution
|
124 |
+
idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
|
125 |
+
# append sampled index to the running sequence
|
126 |
+
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
|
127 |
+
return idx
|
128 |
+
|
129 |
+
class GPTLM(pl.LightningModule):
|
130 |
+
def __init__(self):
|
131 |
+
super().__init__()
|
132 |
+
self.model = GPTLanguageModel()
|
133 |
+
|
134 |
+
def forward(self, idx, targets=None):
|
135 |
+
return self.model(idx, targets)
|
136 |
+
|
137 |
+
def process_step(self, batch):
|
138 |
+
xb, yb = batch
|
139 |
+
logits, loss = self(xb, yb)
|
140 |
+
return(logits, loss)
|
141 |
+
|
142 |
+
def training_step(self, batch, batch_idx):
|
143 |
+
_, loss = self.process_step(batch)
|
144 |
+
self.log('train_loss', loss, on_step=True, on_epoch=True, prog_bar=True, logger=True)
|
145 |
+
return loss
|
146 |
+
|
147 |
+
def validation_step(self, batch, batch_idx):
|
148 |
+
_, loss = self.process_step(batch)
|
149 |
+
self.log('val_loss', loss, on_epoch=True, prog_bar=True, logger=True)
|
150 |
+
return loss
|
151 |
+
|
152 |
+
def configure_optimizers(self):
|
153 |
+
optimizer = torch.optim.AdamW(self.parameters(), lr=learning_rate)
|
154 |
+
return optimizer
|