eliot
commited on
Commit
•
787be42
1
Parent(s):
7048965
training file and weights
Browse files- bigram.py +172 -0
- transformer.pth +3 -0
bigram.py
ADDED
@@ -0,0 +1,172 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
import torch
|
3 |
+
import argparse
|
4 |
+
from torch.nn import functional as F
|
5 |
+
import time
|
6 |
+
from attention_head import AttentionHead,Head, MultiHeadAttention, TransFormerBlock
|
7 |
+
torch.manual_seed(1337)
|
8 |
+
|
9 |
+
def get_batch(batch_size, dataset, block_size):
|
10 |
+
sample = torch.randint(high=len(dataset)- (block_size +1), size = (batch_size, 1))
|
11 |
+
xb = torch.zeros(batch_size,block_size, dtype=torch.long)
|
12 |
+
yb = torch.zeros(batch_size,block_size, dtype=torch.long)
|
13 |
+
for idx, sample_index in enumerate(sample):
|
14 |
+
xb[idx,:] = dataset[sample_index:sample_index+block_size]
|
15 |
+
yb[idx,:] = dataset[sample_index+1:sample_index+block_size+1]
|
16 |
+
return xb, yb
|
17 |
+
|
18 |
+
@torch.no_grad()
|
19 |
+
def eval(model, batch_size, block_size, dataset):
|
20 |
+
xb, yb = get_batch(batch_size, dataset, block_size)
|
21 |
+
logits, loss = model(xb, yb)
|
22 |
+
return loss.item()
|
23 |
+
|
24 |
+
def train(model, optimizer, batch_size, block_size, train_ds, val_ds, steps):
|
25 |
+
sumloss = 0
|
26 |
+
for _ in range(1,steps+1):
|
27 |
+
xb, yb = get_batch(batch_size, train_ds, block_size)
|
28 |
+
logits, loss = model(xb, yb)
|
29 |
+
sumloss += loss.item()
|
30 |
+
optimizer.zero_grad(set_to_none=True)
|
31 |
+
loss.backward()
|
32 |
+
optimizer.step()
|
33 |
+
if _ % 1000 == 0:
|
34 |
+
val_loss = eval(model, 30, block_size, val_ds,)
|
35 |
+
print(f"step {_} || train loss: {sumloss/1000} , val loss: {val_loss}")
|
36 |
+
|
37 |
+
sumloss = 0
|
38 |
+
|
39 |
+
class Transformer(torch.nn.Module):
|
40 |
+
def __init__(self,vocab_size,n_tf=3, block_size=8,token_embed_dim=16) -> None:
|
41 |
+
super().__init__()
|
42 |
+
self.block_size=block_size
|
43 |
+
self.token_embedding_table = torch.nn.Embedding(vocab_size, token_embed_dim)
|
44 |
+
self.positional_embedding = torch.nn.Embedding(block_size, token_embed_dim)
|
45 |
+
self.tf_blocks = torch.nn.Sequential(
|
46 |
+
*[TransFormerBlock(token_embed_dim, block_size, 16, 8) for _ in range(n_tf)]
|
47 |
+
)
|
48 |
+
self.lm_head = torch.nn.Linear(128, vocab_size)
|
49 |
+
def forward(self, idx, targets=None):
|
50 |
+
B,T=idx.shape
|
51 |
+
token_embed = self.token_embedding_table(idx)
|
52 |
+
positional_embed = self.positional_embedding(torch.arange(T))
|
53 |
+
x = token_embed+positional_embed
|
54 |
+
x= self.tf_blocks(x)
|
55 |
+
logits = self.lm_head(x)
|
56 |
+
|
57 |
+
if targets is None:
|
58 |
+
loss = None
|
59 |
+
else:
|
60 |
+
B, T, C = logits.shape
|
61 |
+
logits = logits.view(B*T, C)
|
62 |
+
targets = targets.view(B*T)
|
63 |
+
loss = F.cross_entropy(logits, targets)
|
64 |
+
return logits, loss
|
65 |
+
def generate(self, idx, max_new_tokens):
|
66 |
+
# idx is (B, T) array of indices in the current context
|
67 |
+
for _ in range(max_new_tokens):
|
68 |
+
# get the predictions
|
69 |
+
logits, loss = self(idx[:, -self.block_size:])
|
70 |
+
# focus only on the last time step
|
71 |
+
logits = logits[:, -1, :] # becomes (B, C)
|
72 |
+
# apply softmax to get probabilities
|
73 |
+
probs = F.softmax(logits, dim=-1) # (B, C)
|
74 |
+
# sample from the distribution
|
75 |
+
idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
|
76 |
+
# append sampled index to the running sequence
|
77 |
+
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
|
78 |
+
return idx
|
79 |
+
class BigramLanguageModel(torch.nn.Module):
|
80 |
+
def __init__(self, vocab_size,block_size=8,token_embed_dim=16):
|
81 |
+
super().__init__()
|
82 |
+
self.token_embedding_table = torch.nn.Embedding(vocab_size, token_embed_dim)
|
83 |
+
self.positional_embedding = torch.nn.Embedding(block_size, token_embed_dim)
|
84 |
+
self.attention_head = MultiHeadAttention(n_embed=token_embed_dim,
|
85 |
+
timesteps=block_size,
|
86 |
+
head_size=token_embed_dim//4, # does head size have to == token embed_dim / n heads? I think it does
|
87 |
+
n_heads=4) # (in = (B, T, C), out = B,T,C)
|
88 |
+
self.lm_head = torch.nn.Linear(token_embed_dim, vocab_size) # (in B, T, C, out = B, T, C, performs linear on C)
|
89 |
+
self.block_size = block_size
|
90 |
+
def forward(self, idx, targets=None):
|
91 |
+
B, T = idx.shape
|
92 |
+
# idx and targets are both (B,T) tensor of integers
|
93 |
+
token_embedding = self.token_embedding_table(idx) # (B,T, in), (B,T,embed_dim out)
|
94 |
+
positional_embedding = self.positional_embedding(torch.arange(T,dtype=torch.long)) # (T, embed_dim)
|
95 |
+
x = token_embedding + positional_embedding # (B,T,embed_dim)
|
96 |
+
x = self.attention_head(x) # (B,T,embed_dim)
|
97 |
+
logits = self.lm_head(x)
|
98 |
+
if targets is None:
|
99 |
+
loss = None
|
100 |
+
else:
|
101 |
+
B, T, C = logits.shape
|
102 |
+
logits = logits.view(B*T, C)
|
103 |
+
targets = targets.view(B*T)
|
104 |
+
loss = F.cross_entropy(logits, targets)
|
105 |
+
return logits, loss
|
106 |
+
|
107 |
+
def generate(self, idx, max_new_tokens):
|
108 |
+
# idx is (B, T) array of indices in the current context
|
109 |
+
for _ in range(max_new_tokens):
|
110 |
+
# get the predictions
|
111 |
+
logits, loss = self(idx[:, -self.block_size:])
|
112 |
+
# focus only on the last time step
|
113 |
+
logits = logits[:, -1, :] # becomes (B, C)
|
114 |
+
# apply softmax to get probabilities
|
115 |
+
probs = F.softmax(logits, dim=-1) # (B, C)
|
116 |
+
# sample from the distribution
|
117 |
+
idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
|
118 |
+
# append sampled index to the running sequence
|
119 |
+
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
|
120 |
+
return idx
|
121 |
+
def main():
|
122 |
+
########################
|
123 |
+
#PARAMS#################
|
124 |
+
batch_size = 32
|
125 |
+
block_size= 128
|
126 |
+
n_embed = 128
|
127 |
+
n_tf = 3
|
128 |
+
n_heads=8
|
129 |
+
head_size=16
|
130 |
+
vocab_size=65
|
131 |
+
########################
|
132 |
+
parser = argparse.ArgumentParser(
|
133 |
+
description='Train a bigram language model'
|
134 |
+
)
|
135 |
+
parser.add_argument('-c', '--cont', action='store_true',)
|
136 |
+
parser.add_argument('-e', '--eval', action='store_true',)
|
137 |
+
parser.add_argument('-v', '--verbose',action='store_true')
|
138 |
+
text = open('input.txt').read()
|
139 |
+
characters = sorted(list(set(text)))
|
140 |
+
decoder = dict(enumerate(characters))
|
141 |
+
encoder = {v: k for k, v in decoder.items()}
|
142 |
+
encode = lambda x: encoder[x]
|
143 |
+
decode = lambda x: decoder[x]
|
144 |
+
text_tensor = torch.tensor([encode(c) for c in text])
|
145 |
+
train_tensor = text_tensor[:int(len(text_tensor) * 0.8)]
|
146 |
+
val_tensor = text_tensor[int(len(text_tensor) * 0.8):]
|
147 |
+
model = Transformer(vocab_size=vocab_size, n_tf=n_tf,block_size=block_size, token_embed_dim=n_embed)
|
148 |
+
if parser.parse_args().verbose:
|
149 |
+
print(model)
|
150 |
+
num_params: int = sum(p.numel() for p in model.parameters() if p.requires_grad)
|
151 |
+
print('parameters:', num_params)
|
152 |
+
# if -c is passed we will load the model from the file
|
153 |
+
if parser.parse_args().cont:
|
154 |
+
state_dict = torch.load('transformer.pth')
|
155 |
+
model.load_state_dict(state_dict)
|
156 |
+
optimizer = torch.optim.Adam(model.parameters(), lr=3e-5)
|
157 |
+
s = time.time()
|
158 |
+
if not parser.parse_args().eval:
|
159 |
+
try:
|
160 |
+
train(model, optimizer, batch_size=batch_size, block_size=block_size, train_ds=train_tensor, val_ds=val_tensor,steps= 100000)
|
161 |
+
except KeyboardInterrupt:
|
162 |
+
torch.save(model.state_dict(), 'transformer.pth')
|
163 |
+
exit()
|
164 |
+
if parser.parse_args().verbose:
|
165 |
+
print('training time: ', time.time() - s)
|
166 |
+
torch.save(model.state_dict(), 'transformer.pth')
|
167 |
+
model.eval()
|
168 |
+
print(''.join([decode(c) for c in model.generate(torch.zeros(1,32, dtype=torch.long), 1000)[0].tolist()[32:]]))
|
169 |
+
# 2.57 adam
|
170 |
+
if __name__ == '__main__':
|
171 |
+
main()
|
172 |
+
|
transformer.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b89e10dd4ab50a8ae82d6340d7cddc6c4953035194a30c871c0ea9ee90ab0848
|
3 |
+
size 2543221
|