-
Notifications
You must be signed in to change notification settings - Fork 128
/
train.py
201 lines (165 loc) · 6.31 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
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
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
"""
This module trains a cost model with a GNN on a LLVM-IR transition database
predicting some output reward (the default is instruction count).
Example usage:
$ python train_cost_model.py --num_epoch 10 --batch_size 16 --dataset_size 64
"""
import collections
import io
import logging
import pickle
import sys
import tarfile
import time
from pathlib import Path
from threading import Lock
import numpy as np
import torch
from absl import app, flags
from fasteners import InterProcessLock
from torch.utils.data import DataLoader
import compiler_gym.util.flags.nproc # noqa flag definition
from compiler_gym.util.download import download
from compiler_gym.util.filesystem import extract_tar
from compiler_gym.util.runfiles_path import cache_path, transient_cache_path
from compiler_gym.util.timer import Timer, humanize_duration
from .compiler_gym_dataset import CompilerGymDataset
from .model import GNNEncoder
flags.DEFINE_integer(
"dataset_size", -1, "How large should the dataset be, -1 if no constraint"
)
flags.DEFINE_integer("num_epoch", 100, "Number of epochs for training")
flags.DEFINE_integer("batch_size", 4, "Number of epochs for training")
flags.DEFINE_string(
"db",
"https://dl.fbaipublicfiles.com/compiler_gym/state_transition_dataset/2021-11-15-csmith.tar.bz2",
"URL of the dataset to use.",
)
flags.DEFINE_string(
"db_sha256",
"0b101a17fdbb1851f38ca46cc089b0026eb740e4055a4fe06b4c899ca87256a2",
"SHA256 checksum of the dataset database.",
)
flags.DEFINE_string(
"vocab_db",
"https://dl.fbaipublicfiles.com/compiler_gym/state_transition_dataset/2021-11-15-vocab.tar.bz2",
"URL of the vocabulary database to use.",
)
flags.DEFINE_string(
"vocab_db_sha256",
"af7781f57e6ef430c561afb045fc03693783e668b21826b32234e9c45bd1882c",
"SHA256 checksum of the vocabulary database.",
)
flags.DEFINE_string(
"device", "cuda:0" if torch.cuda.is_available() else "cpu", "The device to run on."
)
FLAGS = flags.FLAGS
logger = logging.getLogger(__name__)
_DB_DOWNLOAD_LOCK = Lock()
def dataset_looper(epoch_num, data_loader, model, device, optimizer=None, train=True):
times = collections.defaultdict(float)
losses = []
unscaled_mse = []
epoch_grad_clip = []
t1 = time.time()
for data in data_loader:
if data is None:
continue
graph, labels = data
times["get_data"] += time.time() - t1
t1 = time.time()
labels = torch.Tensor(labels).to(device)
graph = graph.to(device)
loss, unscaled, _ = model.get_loss(graph, labels)
losses.append(loss.cpu().data.numpy())
unscaled_mse.append(unscaled.cpu().data.numpy())
times["model_forward"] += time.time() - t1
t1 = time.time()
if train:
optimizer.zero_grad()
loss.backward()
grad_clip = torch.nn.utils.clip_grad_norm_(
model.parameters(), max_norm=400.0
)
epoch_grad_clip.append(grad_clip.cpu().data.numpy())
optimizer.step()
times["model_backward"] += time.time() - t1
t1 = time.time()
avg_loss, avg_unscaled = (
np.mean(losses),
np.mean(unscaled_mse),
)
avg_grad_clip = None
if train:
avg_grad_clip = np.mean(epoch_grad_clip)
times = ", ".join(f"{k}: {humanize_duration(v)}" for k, v in times.items())
print(
f" Epoch {epoch_num + 1} {'training' if train else 'validation'} took: "
f"{{ {times} }}, loss: {avg_loss}, unscaled: {avg_unscaled}, "
f"grad_clip {avg_grad_clip}"
)
return avg_loss, avg_unscaled, avg_grad_clip
def train(dataset, data_loader, model, num_epoch, device):
optimizer = torch.optim.Adam(model.parameters())
for epoch in range(num_epoch):
with Timer(f"Epoch {epoch + 1} of {num_epoch} ({(epoch + 1) / num_epoch:.1%})"):
dataset.set_distribution_type("train")
dataset_looper(epoch, data_loader, model, device, optimizer)
dataset.set_distribution_type("dev")
dataset_looper(epoch, data_loader, model, device, train=False)
def download_and_unpack_database(db: str, sha256: str) -> Path:
"""Download the given database, unpack it to the local filesystem, and
return the path.
"""
local_dir = cache_path(f"state_transition_dataset/{sha256}")
with _DB_DOWNLOAD_LOCK, InterProcessLock(
transient_cache_path(".state_transition_database_download.LOCK")
):
if not (local_dir / ".installed").is_file():
tar_data = io.BytesIO(download(db, sha256))
local_dir.mkdir(parents=True, exist_ok=True)
logger.info("Unpacking database to %s ...", local_dir)
with tarfile.open(fileobj=tar_data, mode="r:bz2") as arc:
extract_tar(arc, str(local_dir))
(local_dir / ".installed").touch()
unpacked = [f for f in local_dir.iterdir() if f.name != ".installed"]
if len(unpacked) != 1:
print(
f"fatal: Archive {db} expected to contain one file, contains: {len(unpacked)}",
file=sys.stderr,
)
return unpacked[0]
def main(argv):
"""Main entry point."""
del argv # unused
node_vocab_pth = download_and_unpack_database(
db=FLAGS.vocab_db, sha256=FLAGS.vocab_db_sha256
)
root_pth = download_and_unpack_database(db=FLAGS.db, sha256=FLAGS.db_sha256)
with open(node_vocab_pth, "rb") as f:
vocab = pickle.load(f)
model = GNNEncoder(
# Add one to the vocab size to accomodate for the out-of-vocab element.
node_vocab_size=len(vocab) + 1,
node_hidden_size=64,
)
# This is required to get the vocab into the right state
# as the vocab is over all nodes of the graph
vocab = {"text": vocab}
model.to(FLAGS.device)
print(model)
dataset = CompilerGymDataset(root_pth, vocab=vocab, dataset_size=FLAGS.dataset_size)
dataset_loader = DataLoader(
dataset,
batch_size=FLAGS.batch_size,
shuffle=True,
num_workers=FLAGS.nproc,
collate_fn=dataset.collate_fn,
)
train(dataset, dataset_loader, model, FLAGS.num_epoch, FLAGS.device)
if __name__ == "__main__":
app.run(main)