-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
169 lines (140 loc) · 6.46 KB
/
train.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
import os.path
import wandb
import torch
from tqdm import tqdm
import model
from transformers import GPT2TokenizerFast
from torch.utils import data
import numpy as np
tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
def _lazy_file_read(file_obj, chunk_size=1024):
while True:
data = file_obj.read(chunk_size)
if not data:
break
yield data
def pre_tokenize_dataset(path, save_path):
print(f"Running tokenization for {path}")
try:
with open(path, 'r', encoding='utf-8') as file:
arrays = []
for line in tqdm(_lazy_file_read(file)):
tokens = tokenizer(line).data['input_ids']
arrays.append(np.array(tokens, dtype=np.int32))
except KeyboardInterrupt:
pass
finally:
np.save(save_path, np.concatenate(arrays))
print(f"Saved tokenized file to binary {save_path}")
class TinyStoriesDataset(data.IterableDataset):
def __getitem__(self, index):
pass
def __init__(self, tokenized_path, block_size: int, device: str = 'cuda'):
# Each line represents a short story
self.block_size = block_size
self.device = device
self.train = np.load(tokenized_path, mmap_mode='r', allow_pickle=True)
def __iter__(self):
while True:
idx = np.random.randint(0, len(self.train) - self.block_size, 1)[0]
chunk = self.train[idx:idx + self.block_size + 1]
source = torch.tensor(chunk[:-1], device=self.device, dtype=torch.long)
target = torch.tensor(chunk[1:], device=self.device, dtype=torch.long)
yield source, target
def __len__(self):
return len(self.train)
# ----------------------------------------- SETUP ----------------------------------------------------------------------
# ----------------------------------------------------------------------------------------------------------------------
config = {
"BLOCK_SIZE": 128,
"EMB_SIZE": 768,
"N_ATTENTION_HEADS": 4,
"N_DECODER_BLOCKS": 2,
"VOCAB_SIZE": len(tokenizer),
"MAX_OUT_TOKENS": 200,
"EVAL_INTERVAL": 1000,
"EVAL_ITER": 100,
"LR": 3e-4,
"BATCH_SIZE": 32,
"DEVICE": 'cuda' if torch.cuda.is_available() else 'cpu'
}
assert config['EMB_SIZE'] % config['N_ATTENTION_HEADS'] == 0
wandb.init(
project='TinyLM',
config=config
)
text_table = wandb.Table(columns=['epoch', 'loss', 'predicted text'])
@torch.no_grad()
def eval_model(training_model: torch.nn.Module, val_loader: torch.utils.data.DataLoader):
training_model.eval()
losses = torch.zeros(config['EVAL_ITER'])
for k in range(config['EVAL_ITER']):
s_val, t_val = next(iter(val_loader))
val_logits = training_model(s_val)
val_logits: torch.Tensor = val_logits.view(config['BATCH_SIZE'] * config['BLOCK_SIZE'], config['VOCAB_SIZE'])
t_val = t_val.view(config['BATCH_SIZE'] * config['BLOCK_SIZE'])
losses[k] = torch.nn.functional.cross_entropy(val_logits, t_val).item()
training_model.train()
return losses.mean()
@torch.no_grad()
def generate_sample_text(training_model: model.TinyLM, max_tokens: int = 200) -> str:
training_model.eval()
context = torch.zeros((5, config['BLOCK_SIZE']), dtype=torch.long, device=config['DEVICE'])
out_tokens = training_model.generate(context, max_new_tokens=max_tokens)
# Reform to one long piece of text
out_tokens = out_tokens.view(out_tokens.shape[0] * out_tokens.shape[1])
training_model.train()
return tokenizer.decode(out_tokens)
print(f"Loading model on device {config['DEVICE']}")
model = model.TinyLM(emb_dim=config['EMB_SIZE'], block_size=config['BLOCK_SIZE'],
n_att_heads=config['N_ATTENTION_HEADS'], n_decoders=config['N_DECODER_BLOCKS'],
vocab_size=config['VOCAB_SIZE'], device=config['DEVICE'])
optim = torch.optim.Adam(model.parameters(), lr=config['LR'])
if os.path.exists('models/tiny_lm'):
checkpoint = torch.load('models/tiny_lm')
print(f"Loading model from checkpoint, continuing training after {checkpoint['b_idx']} episodes ...")
model.load_state_dict(checkpoint['model_state_dict'])
optim.load_state_dict(checkpoint['optimizer_state_dict'])
model = model.to(config['DEVICE'])
loss_fn = torch.nn.CrossEntropyLoss()
# ----------------------------------------------------------------------------------------------------------------------
TINY_STORY_TRAIN = 'data/TinyStories-train.txt'
TINY_TOKENIZED = 'data/tiny_tokenized.npy'
if not os.path.exists(TINY_TOKENIZED):
pre_tokenize_dataset(TINY_STORY_TRAIN, TINY_TOKENIZED)
train_tiny_stories = TinyStoriesDataset(TINY_TOKENIZED, config['BLOCK_SIZE'], device=config['DEVICE'])
train_loader = data.DataLoader(train_tiny_stories, batch_size=config['BATCH_SIZE'])
# ----------------------------------------------------------------------------------------------------------------------
TINY_STORY_VAL = 'data/TinyStories-valid.txt'
TINY_TOKENIZED_VAL = 'data/tiny_tokenized_val.npy'
if not os.path.exists(TINY_TOKENIZED_VAL):
pre_tokenize_dataset(TINY_STORY_VAL, TINY_TOKENIZED_VAL)
val_tiny_stories = TinyStoriesDataset(TINY_TOKENIZED_VAL, config['BLOCK_SIZE'], device=config['DEVICE'])
val_loader = data.DataLoader(val_tiny_stories, batch_size=config['BATCH_SIZE'])
# ----------------------------------------------------------------------------------------------------------------------
# ----------------------------------------- TRAINING -------------------------------------------------------------------
try:
for b_idx, batch in enumerate(train_loader):
# Inference
sources, targets = batch
logits = model(sources)
logits = logits.view(config['BATCH_SIZE'] * config['BLOCK_SIZE'], config['VOCAB_SIZE'])
targets = targets.view(config['BATCH_SIZE'] * config['BLOCK_SIZE'])
loss = torch.nn.functional.cross_entropy(logits, targets)
wandb.log({"loss": loss})
# Weight update
optim.zero_grad()
loss.backward()
optim.step()
if b_idx % config['EVAL_INTERVAL'] == 0:
val_loss = eval_model(model, val_loader)
generated_text = generate_sample_text(model, max_tokens=config['MAX_OUT_TOKENS'])
print(generated_text)
wandb.log({"val_loss": val_loss})
except KeyboardInterrupt:
pass
finally:
torch.save({'epoch': b_idx,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optim.state_dict()
}, 'models/tiny_lm')