-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_proteins.py
263 lines (209 loc) · 7.84 KB
/
train_proteins.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
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
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
"""
Training a simple NanoGPT model on protein sequences
Adapted from Karpathy's NanoGPT (https://github.com/karpathy/nanoGPT) with some changes:
* Make it trainable on multiple isolated sequences of variable lengths,
using padding and masking the transformer's weights accordingly.
* Use PyTorch Lightning for the training loop
* Vocabulary is amino acids
The original PDB sequence dataset comes from
https://ftp.wwpdb.org/pub/pdb/derived_data/pdb_seqres.txt.gz
and has been preprocessed by `preprocess_pdb_seqres.py`.
"""
import torch
import random
import numpy as np
import pytorch_lightning as pl
from torch.utils.data import Dataset, DataLoader
from collections import Counter
from pytorch_lightning.callbacks import TQDMProgressBar, ModelCheckpoint
from pytorch_lightning.loggers import TensorBoardLogger
from pytorch_lightning import seed_everything
import os
from nano_transformer import NanoTransformer, BigramModel, PLNanoTransformer
seed_everything(1337)
PROT_FNAME = "data/prot_seqs.txt"
# We leave the possibility to benchmark against a simple bigram model
# True: train NanoTransformer, False: train BigramModel
USE_TRANSFORMER = True
# hyperparameters
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
BATCH_SIZE = 32 if DEVICE == "cpu" else 64
BLOCK_SIZE = 8 if DEVICE == "cpu" else 384 # context size
N_EMBD = 16 if DEVICE == "cpu" else 384 # called d_model in paper
N_BLOCKS = 2 if DEVICE == "cpu" else 6 # number N of transformer blocks
NUM_HEADS = 2 if DEVICE == "cpu" else 6 # nr attention heads
DROPOUT = 0.2
MAX_ITERS = 75000
EVAL_INTERVAL = 5000
LEARNING_RATE = 1e-3 if DEVICE == "cpu" else 3e-4
EVAL_ITERS = 500
PRECISION = 32 if DEVICE == "cpu" else 16
NUM_WORKERS = 0 if DEVICE == "cpu" else 4 # number of workers for dataloaders
CHECKPOINT_EVERY_STEP = 1000
print(f"device: {DEVICE}")
# N_EMBD must be divisible by NUM_HEADS
assert N_EMBD % NUM_HEADS == 0, "N_EMBD must be divisible by NUM_HEADS"
""" Read file
"""
with open(PROT_FNAME, "r") as f:
lines = f.readlines()
""" Define amino acid vocabulary encoding
"""
pad = "!" # padding character
flat_text = [c for line in lines for c in line] # all proteins concatenated
chars = sorted(set(flat_text)) + [pad]
vocab_size = len(chars)
stoi = {c: i for i, c in enumerate(chars)}
itos = {i: c for i, c in enumerate(chars)}
counter = Counter(flat_text)
# plt.bar(counter.keys(), counter.values(), log=True)
def encode(s):
return [stoi[c] for c in s]
def decode(i):
return "".join([itos[ii] for ii in i])
def encode_pad(s, block_size):
encoding = encode(s)
return encoding + [stoi[pad]] * max(0, block_size + 1 - len(encoding))
""" Train / val split
"""
# There is one line per protein.
# We shuffle for train/val split (in place).
random.shuffle(lines)
n = int(0.9 * len(lines))
train_lines = lines[:n]
val_lines = lines[n:]
""" Define dataset.
We need to define a custome PyTorch Dataset which disambiguates
between line indices and subsequence start indices.
"""
class LineDataset(Dataset):
"""
A dataset where sampling is uniform over all lines.
All lines are sampled with the same frequency.
"""
def __init__(self, lines):
self.lines = lines
nr_samples_per_line = [max(1, len(line) - BLOCK_SIZE) for line in self.lines]
self.num_samples = sum(nr_samples_per_line)
def __len__(self):
return self.num_samples
def __getitem__(self, idx):
# select a line uniformly at random and then a subsequence uniformly at random
line_idx = torch.randint(len(self.lines), size=(1,)).item()
line = self.lines[line_idx]
line_encoded = encode_pad(line, BLOCK_SIZE)
start_idx = torch.randint(len(line_encoded) - BLOCK_SIZE, size=(1,)).item()
end_idx = start_idx + BLOCK_SIZE
x = torch.tensor(line_encoded[start_idx:end_idx], dtype=torch.long)
y = torch.tensor(line_encoded[start_idx + 1 : end_idx + 1], dtype=torch.long)
length = torch.tensor(min(len(line), BLOCK_SIZE), dtype=torch.long)
if USE_TRANSFORMER:
# transformer uses length to mask out the padded tokens
return x, y, length
else:
return x, y
class LengthAwareLineDataset(Dataset):
"""
A dataset where sampling is uniform over all possible subsequences.
As a result, sequences are (much) more likely to come from longer lines.
We don't use it as it biases the lengths of generated sequences.
"""
def __init__(self, lines):
self.lines = sorted(lines, key=len, reverse=True)
nr_samples_per_line = [len(line) - BLOCK_SIZE for line in self.lines]
self.num_samples = sum(nr_samples_per_line)
self.line_idx_to_cumulative_nr_samples = np.array(
[0] + np.cumsum(nr_samples_per_line).tolist()
)
def __len__(self):
return self.num_samples
def __getitem__(self, idx):
# infer line idx from idx, using binary search over line_idx_to_cumulative_nr_samples
# np.searchsorted() returns the index of the first element in the array that is greater than the value
line_idx = (
np.searchsorted(self.line_idx_to_cumulative_nr_samples, idx, side="right")
- 1
)
start_idx = idx - self.line_idx_to_cumulative_nr_samples[line_idx]
line = self.lines[line_idx]
line_encoded = encode_pad(line, BLOCK_SIZE)
x = torch.tensor(
line_encoded[start_idx : start_idx + BLOCK_SIZE], dtype=torch.long
)
y = torch.tensor(
line_encoded[start_idx + 1 : start_idx + BLOCK_SIZE + 1], dtype=torch.long
)
length = torch.tensor(min(len(line), BLOCK_SIZE), dtype=torch.long)
if USE_TRANSFORMER:
# transformer uses length to mask out the padded tokens
return x, y, length
else:
return x, y
if USE_TRANSFORMER:
inner_model = NanoTransformer(
vocab_size=vocab_size,
block_size=BLOCK_SIZE,
n_embd=N_EMBD,
n_blocks=N_BLOCKS,
num_heads=NUM_HEADS,
dropout=DROPOUT,
)
else:
inner_model = BigramModel(vocab_size=vocab_size)
nano_prot_gpt = PLNanoTransformer(inner_model, LEARNING_RATE)
train_dataset = LineDataset(train_lines)
train_loader = DataLoader(
dataset=train_dataset, batch_size=BATCH_SIZE, num_workers=NUM_WORKERS, shuffle=True
)
val_dataset = LineDataset(val_lines)
val_loader = DataLoader(
dataset=val_dataset, batch_size=BATCH_SIZE, num_workers=NUM_WORKERS, shuffle=False
)
# configure tensorboard logging
logger = TensorBoardLogger("tensorboard_logs", name="nano_prot_gpt")
# log hyperparameters
if USE_TRANSFORMER:
logger.log_hyperparams(
{
"block_size": BLOCK_SIZE,
"n_embd": N_EMBD,
"n_blocks": N_BLOCKS,
"num_heads": NUM_HEADS,
"dropout": DROPOUT,
"max_iters": MAX_ITERS,
"learning_rate": LEARNING_RATE,
"batch_size": BATCH_SIZE,
"precision": PRECISION,
"is bigram": False,
}
)
else:
logger.log_hyperparams(
{
"max_iters": MAX_ITERS,
"learning_rate": LEARNING_RATE,
"batch_size": BATCH_SIZE,
"is_bigram": True,
}
)
# configure model checkpointing
checkpoint_callback = ModelCheckpoint(
save_top_k=5,
monitor="val_loss",
mode="min",
every_n_train_steps=CHECKPOINT_EVERY_STEP,
filename="nano_prot_gpt-{epoch:02d}-{step:07d}-{val_loss:.4f}",
)
trainer = pl.Trainer(
max_steps=MAX_ITERS,
val_check_interval=EVAL_INTERVAL,
limit_val_batches=EVAL_ITERS,
callbacks=[TQDMProgressBar(refresh_rate=40), checkpoint_callback],
precision=PRECISION,
logger=logger,
accelerator=DEVICE,
)
trainer.fit(
model=nano_prot_gpt, train_dataloaders=train_loader, val_dataloaders=val_loader
)
print("Best model checkpoint:", checkpoint_callback.best_model_path)