-
Notifications
You must be signed in to change notification settings - Fork 263
/
main.py
279 lines (239 loc) · 7.7 KB
/
main.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
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
# pyre-strict
import argparse
import sys
import tempfile
from argparse import Namespace
from typing import List, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Adadelta
from torch.optim.lr_scheduler import StepLR
from torch.utils.data import DataLoader
from torcheval.metrics import MulticlassAccuracy
from torchtnt.framework.auto_unit import AutoUnit, TrainStepResults
from torchtnt.framework.fit import fit
from torchtnt.framework.state import State
from torchtnt.utils import init_from_env, seed, TLRScheduler
from torchtnt.utils.loggers import TensorBoardLogger
from torchvision import datasets, transforms
Batch = Tuple[torch.Tensor, torch.Tensor]
class Net(nn.Module):
def __init__(self) -> None:
super(Net, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 3, 1)
self.conv2 = nn.Conv2d(32, 64, 3, 1)
self.dropout1 = nn.Dropout(0.25)
self.dropout2 = nn.Dropout(0.5)
self.fc1 = nn.Linear(9216, 128)
self.fc2 = nn.Linear(128, 10)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.conv1(x)
x = F.relu(x)
x = self.conv2(x)
x = F.relu(x)
x = F.max_pool2d(x, 2)
x = self.dropout1(x)
x = torch.flatten(x, 1)
x = self.fc1(x)
x = F.relu(x)
x = self.dropout2(x)
x = self.fc2(x)
output = F.log_softmax(x, dim=1)
return output
class MyUnit(AutoUnit[Batch]):
def __init__(
self,
*,
tb_logger: TensorBoardLogger,
train_accuracy: MulticlassAccuracy,
log_every_n_steps: int,
lr: float,
gamma: float,
module: torch.nn.Module,
device: torch.device,
strategy: str,
precision: Optional[str],
gradient_accumulation_steps: int,
detect_anomaly: bool,
clip_grad_norm: float,
) -> None:
super().__init__(
module=module,
device=device,
strategy=strategy,
precision=precision,
gradient_accumulation_steps=gradient_accumulation_steps,
detect_anomaly=detect_anomaly,
clip_grad_norm=clip_grad_norm,
)
self.tb_logger = tb_logger
self.lr = lr
self.gamma = gamma
# create an accuracy Metric to compute the accuracy of training
self.train_accuracy = train_accuracy
self.log_every_n_steps = log_every_n_steps
def configure_optimizers_and_lr_scheduler(
self, module: torch.nn.Module
) -> Tuple[torch.optim.Optimizer, Optional[TLRScheduler]]:
optimizer = Adadelta(module.parameters(), lr=self.lr)
lr_scheduler = StepLR(optimizer, step_size=1, gamma=self.gamma)
return optimizer, lr_scheduler
def compute_loss(
self, state: State, data: Batch
) -> Tuple[torch.Tensor, torch.Tensor]:
inputs, targets = data
outputs = self.module(inputs)
outputs = torch.squeeze(outputs)
loss = torch.nn.functional.nll_loss(outputs, targets)
return loss, outputs
def on_train_step_end(
self,
state: State,
data: Batch,
step: int,
results: TrainStepResults,
) -> None:
loss, outputs = results.loss, results.outputs
_, targets = data
self.train_accuracy.update(outputs, targets)
if step % self.log_every_n_steps == 0:
accuracy = self.train_accuracy.compute()
self.tb_logger.log("accuracy", accuracy, step)
self.tb_logger.log("loss", loss, step)
def on_train_epoch_end(self, state: State) -> None:
super().on_train_epoch_end(state)
# reset the metric every epoch
self.train_accuracy.reset()
def on_eval_step_end(
self,
state: State,
data: Batch,
step: int,
loss: torch.Tensor,
outputs: torch.Tensor,
) -> None:
if step % self.log_every_n_steps == 0:
self.tb_logger.log("evaluation loss", loss, step)
def main(argv: List[str]) -> None:
# parse command line arguments
args = get_args(argv)
# seed the RNG for better reproducibility. see docs https://pytorch.org/docs/stable/notes/randomness.html
seed(args.seed)
# device and process group initialization
device = init_from_env()
# avoid torch autocast exception
if device.type == "mps":
device = torch.device("cpu")
path = tempfile.mkdtemp()
tb_logger = TensorBoardLogger(path)
on_cuda = device.type == "cuda"
transform = transforms.Compose(
[transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]
)
train_dataset = datasets.MNIST(
"../data", train=True, download=True, transform=transform
)
eval_dataset = datasets.MNIST("../data", train=False, transform=transform)
train_dataloader = DataLoader(
train_dataset, batch_size=args.batch_size, pin_memory=on_cuda
)
eval_dataloader = DataLoader(
eval_dataset, batch_size=args.test_batch_size, pin_memory=on_cuda
)
module = Net()
train_accuracy = MulticlassAccuracy(device=device)
my_unit = MyUnit(
tb_logger=tb_logger,
train_accuracy=train_accuracy,
log_every_n_steps=args.log_every_n_steps,
lr=args.lr,
gamma=args.gamma,
module=module,
device=device,
strategy="ddp",
precision=args.precision,
gradient_accumulation_steps=4,
detect_anomaly=True,
clip_grad_norm=1.0,
)
fit(
my_unit,
train_dataloader=train_dataloader,
eval_dataloader=eval_dataloader,
max_epochs=args.max_epochs,
max_train_steps_per_epoch=args.max_train_steps_per_epoch,
)
if args.save_model:
torch.save(module.state_dict(), "mnist_cnn.pt")
def get_args(argv: List[str]) -> Namespace:
"""Parse command line arguments"""
parser = argparse.ArgumentParser()
parser.add_argument(
"--seed", type=int, default=1, metavar="S", help="random seed (default: 1)"
)
parser.add_argument(
"--batch-size",
type=int,
default=64,
metavar="N",
help="input batch size for training (default: 64)",
)
parser.add_argument(
"--test-batch-size",
type=int,
default=1000,
metavar="N",
help="input batch size for testing (default: 1000)",
)
parser.add_argument(
"--max-epochs",
type=int,
default=14,
metavar="N",
help="number of epochs to train (default: 14)",
)
parser.add_argument(
"--lr",
type=float,
default=1.0,
metavar="LR",
help="learning rate (default: 1.0)",
)
parser.add_argument(
"--gamma",
type=float,
default=0.7,
metavar="M",
help="Learning rate step gamma (default: 0.7)",
)
parser.add_argument(
"--save-model",
action="store_true",
default=False,
help="For Saving the current Model",
)
parser.add_argument(
"--log-every-n-steps", type=int, default=10, help="log every n steps"
)
parser.add_argument(
"--precision",
type=str,
default=None,
help="fp16 or bf16",
choices=["fp16", "bf16"],
)
parser.add_argument(
"--max-train-steps-per-epoch",
type=int,
default=20,
help="the max number of steps to run per epoch for training",
)
return parser.parse_args(argv)
if __name__ == "__main__":
main(sys.argv[1:])