Spaces:
Sleeping
Sleeping
Upload 7 files
Browse files- block.py +21 -0
- data_utils.py +49 -0
- feedforward.py +17 -0
- gpt_config.py +16 -0
- gpt_language_model.py +66 -0
- head.py +33 -0
- multi_head_attention.py +19 -0
block.py
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
|
3 |
+
from multi_head_attention import MultiHeadAttention
|
4 |
+
from feedforward import FeedFoward
|
5 |
+
|
6 |
+
class Block(nn.Module):
|
7 |
+
""" Transformer block: communication followed by computation """
|
8 |
+
|
9 |
+
def __init__(self, n_embd, n_head):
|
10 |
+
# n_embd: embedding dimension, n_head: the number of heads we'd like
|
11 |
+
super().__init__()
|
12 |
+
head_size = n_embd // n_head
|
13 |
+
self.sa = MultiHeadAttention(n_head, head_size)
|
14 |
+
self.ffwd = FeedFoward(n_embd)
|
15 |
+
self.ln1 = nn.LayerNorm(n_embd)
|
16 |
+
self.ln2 = nn.LayerNorm(n_embd)
|
17 |
+
|
18 |
+
def forward(self, x):
|
19 |
+
x = x + self.sa(self.ln1(x))
|
20 |
+
x = x + self.ffwd(self.ln2(x))
|
21 |
+
return x
|
data_utils.py
ADDED
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
with open('data/input.txt', 'r', encoding='utf-8') as f:
|
4 |
+
text = f.read()
|
5 |
+
|
6 |
+
# here are all the unique characters that occur in this text
|
7 |
+
chars = sorted(list(set(text)))
|
8 |
+
vocab_size = len(chars)
|
9 |
+
# create a mapping from characters to integers
|
10 |
+
stoi = { ch:i for i,ch in enumerate(chars) }
|
11 |
+
itos = { i:ch for i,ch in enumerate(chars) }
|
12 |
+
encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers
|
13 |
+
decode = lambda l: ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string
|
14 |
+
|
15 |
+
# Train and test splits
|
16 |
+
data = torch.tensor(encode(text), dtype=torch.long)
|
17 |
+
n = int(0.9*len(data)) # first 90% will be train, rest val
|
18 |
+
train_data = data[:n]
|
19 |
+
val_data = data[n:]
|
20 |
+
|
21 |
+
'''
|
22 |
+
'''
|
23 |
+
def get_train_data():
|
24 |
+
return train_data
|
25 |
+
|
26 |
+
'''
|
27 |
+
'''
|
28 |
+
def get_val_data():
|
29 |
+
return val_data
|
30 |
+
|
31 |
+
'''
|
32 |
+
'''
|
33 |
+
def get_data():
|
34 |
+
return data
|
35 |
+
|
36 |
+
'''
|
37 |
+
'''
|
38 |
+
def get_encoder():
|
39 |
+
return encode
|
40 |
+
|
41 |
+
'''
|
42 |
+
'''
|
43 |
+
def get_decoder():
|
44 |
+
return decode
|
45 |
+
|
46 |
+
'''
|
47 |
+
'''
|
48 |
+
def get_vocab_size():
|
49 |
+
return vocab_size
|
feedforward.py
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import gpt_config as config
|
3 |
+
|
4 |
+
class FeedFoward(nn.Module):
|
5 |
+
""" a simple linear layer followed by a non-linearity """
|
6 |
+
|
7 |
+
def __init__(self, n_embd):
|
8 |
+
super().__init__()
|
9 |
+
self.net = nn.Sequential(
|
10 |
+
nn.Linear(n_embd, 4 * n_embd),
|
11 |
+
nn.ReLU(),
|
12 |
+
nn.Linear(4 * n_embd, n_embd),
|
13 |
+
nn.Dropout(config.dropout),
|
14 |
+
)
|
15 |
+
|
16 |
+
def forward(self, x):
|
17 |
+
return self.net(x)
|
gpt_config.py
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
# hyperparameters
|
4 |
+
batch_size = 64 # how many independent sequences will we process in parallel?
|
5 |
+
block_size = 256 # what is the maximum context length for predictions?
|
6 |
+
max_iters = 10000
|
7 |
+
eval_interval = 500
|
8 |
+
learning_rate = 3e-4
|
9 |
+
# device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
10 |
+
device = 'mps'
|
11 |
+
eval_iters = 200
|
12 |
+
n_embd = 384
|
13 |
+
n_head = 6
|
14 |
+
n_layer = 6
|
15 |
+
dropout = 0.2
|
16 |
+
vocab_size = 65
|
gpt_language_model.py
ADDED
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from torch.nn import functional as F
|
4 |
+
|
5 |
+
import gpt_config as config
|
6 |
+
from block import Block
|
7 |
+
|
8 |
+
class GPTLanguageModel(nn.Module):
|
9 |
+
|
10 |
+
def __init__(self):
|
11 |
+
super().__init__()
|
12 |
+
# each token directly reads off the logits for the next token from a lookup table
|
13 |
+
self.token_embedding_table = nn.Embedding(config.vocab_size, config.n_embd)
|
14 |
+
self.position_embedding_table = nn.Embedding(config.block_size, config.n_embd)
|
15 |
+
self.blocks = nn.Sequential(*[Block(config.n_embd, n_head=config.n_head) for _ in range(config.n_layer)])
|
16 |
+
self.ln_f = nn.LayerNorm(config.n_embd) # final layer norm
|
17 |
+
self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
|
18 |
+
|
19 |
+
# better init, not covered in the original GPT video, but important, will cover in followup video
|
20 |
+
self.apply(self._init_weights)
|
21 |
+
|
22 |
+
def _init_weights(self, module):
|
23 |
+
if isinstance(module, nn.Linear):
|
24 |
+
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
|
25 |
+
if module.bias is not None:
|
26 |
+
torch.nn.init.zeros_(module.bias)
|
27 |
+
elif isinstance(module, nn.Embedding):
|
28 |
+
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
|
29 |
+
|
30 |
+
def forward(self, idx, targets=None):
|
31 |
+
B, T = idx.shape
|
32 |
+
|
33 |
+
# idx and targets are both (B,T) tensor of integers
|
34 |
+
tok_emb = self.token_embedding_table(idx) # (B,T,C)
|
35 |
+
pos_emb = self.position_embedding_table(torch.arange(T, device=config.device)) # (T,C)
|
36 |
+
x = tok_emb + pos_emb # (B,T,C)
|
37 |
+
x = self.blocks(x) # (B,T,C)
|
38 |
+
x = self.ln_f(x) # (B,T,C)
|
39 |
+
logits = self.lm_head(x) # (B,T,vocab_size)
|
40 |
+
|
41 |
+
if targets is None:
|
42 |
+
loss = None
|
43 |
+
else:
|
44 |
+
B, T, C = logits.shape
|
45 |
+
logits = logits.view(B*T, C)
|
46 |
+
targets = targets.view(B*T)
|
47 |
+
loss = F.cross_entropy(logits, targets)
|
48 |
+
|
49 |
+
return logits, loss
|
50 |
+
|
51 |
+
def generate(self, idx, max_new_tokens):
|
52 |
+
# idx is (B, T) array of indices in the current context
|
53 |
+
for _ in range(max_new_tokens):
|
54 |
+
# crop idx to the last block_size tokens
|
55 |
+
idx_cond = idx[:, -config.block_size:]
|
56 |
+
# get the predictions
|
57 |
+
logits, loss = self(idx_cond)
|
58 |
+
# focus only on the last time step
|
59 |
+
logits = logits[:, -1, :] # becomes (B, C)
|
60 |
+
# apply softmax to get probabilities
|
61 |
+
probs = F.softmax(logits, dim=-1) # (B, C)
|
62 |
+
# sample from the distribution
|
63 |
+
idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
|
64 |
+
# append sampled index to the running sequence
|
65 |
+
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
|
66 |
+
return idx
|
head.py
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from torch.nn import functional as F
|
4 |
+
|
5 |
+
import gpt_config as config
|
6 |
+
|
7 |
+
class Head(nn.Module):
|
8 |
+
""" one head of self-attention """
|
9 |
+
|
10 |
+
def __init__(self, head_size):
|
11 |
+
super().__init__()
|
12 |
+
self.key = nn.Linear(config.n_embd, head_size, bias=False)
|
13 |
+
self.query = nn.Linear(config.n_embd, head_size, bias=False)
|
14 |
+
self.value = nn.Linear(config.n_embd, head_size, bias=False)
|
15 |
+
self.register_buffer('tril', torch.tril(torch.ones(config.block_size, config.block_size)))
|
16 |
+
|
17 |
+
self.dropout = nn.Dropout(config.dropout)
|
18 |
+
|
19 |
+
def forward(self, x):
|
20 |
+
# input of size (batch, time-step, channels)
|
21 |
+
# output of size (batch, time-step, head size)
|
22 |
+
B,T,C = x.shape
|
23 |
+
k = self.key(x) # (B,T,hs)
|
24 |
+
q = self.query(x) # (B,T,hs)
|
25 |
+
# compute attention scores ("affinities")
|
26 |
+
wei = q @ k.transpose(-2,-1) * k.shape[-1]**-0.5 # (B, T, hs) @ (B, hs, T) -> (B, T, T)
|
27 |
+
wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T)
|
28 |
+
wei = F.softmax(wei, dim=-1) # (B, T, T)
|
29 |
+
wei = self.dropout(wei)
|
30 |
+
# perform the weighted aggregation of the values
|
31 |
+
v = self.value(x) # (B,T,hs)
|
32 |
+
out = wei @ v # (B, T, T) @ (B, T, hs) -> (B, T, hs)
|
33 |
+
return out
|
multi_head_attention.py
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
|
4 |
+
import gpt_config as config
|
5 |
+
from head import Head
|
6 |
+
|
7 |
+
class MultiHeadAttention(nn.Module):
|
8 |
+
""" multiple heads of self-attention in parallel """
|
9 |
+
|
10 |
+
def __init__(self, num_heads, head_size):
|
11 |
+
super().__init__()
|
12 |
+
self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
|
13 |
+
self.proj = nn.Linear(head_size * num_heads, config.n_embd)
|
14 |
+
self.dropout = nn.Dropout(config.dropout)
|
15 |
+
|
16 |
+
def forward(self, x):
|
17 |
+
out = torch.cat([h(x) for h in self.heads], dim=-1)
|
18 |
+
out = self.dropout(self.proj(out))
|
19 |
+
return out
|