-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_mnist.py
153 lines (134 loc) · 4.52 KB
/
train_mnist.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
import os
import argparse
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim.lr_scheduler import CosineAnnealingLR
from relu import ReLUAlpha
from tqdm import tqdm
import pandas as pd
from data_utils import get_mnist_loaders
from train import train, test
def boolean_string(s):
if s not in {"False", "True"}:
raise ValueError("Not a valid boolean string")
return s == "True"
def init(alpha=0, regularization="relu"):
if regularization == "relu":
net = torch.torch.nn.Sequential(
nn.Flatten(),
nn.Linear(28 * 28, 2048),
ReLUAlpha(alpha),
nn.Linear(2048, 2048),
ReLUAlpha(alpha),
nn.Linear(2048, 2048),
ReLUAlpha(alpha),
nn.Linear(2048, 10),
).to(device)
elif regularization == "batch_norm":
net = torch.torch.nn.Sequential(
nn.Flatten(),
nn.Linear(28 * 28, 2048),
nn.BatchNorm1d(2048),
ReLUAlpha(alpha),
nn.Linear(2048, 2048),
nn.BatchNorm1d(2048),
ReLUAlpha(alpha),
nn.Linear(2048, 2048),
nn.BatchNorm1d(2048),
ReLUAlpha(alpha),
nn.Linear(2048, 10),
).to(device)
elif regularization == "dropout":
net = torch.torch.nn.Sequential(
nn.Flatten(),
nn.Linear(28 * 28, 2048),
ReLUAlpha(alpha),
nn.Dropout(alpha),
nn.Linear(2048, 2048),
ReLUAlpha(alpha),
nn.Dropout(alpha),
nn.Linear(2048, 2048),
ReLUAlpha(alpha),
nn.Dropout(alpha),
nn.Linear(2048, 10),
).to(device)
return net
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Pytorch relu experiment impact of different values for ReLU'(0)"
)
parser.add_argument("--alpha", type=float, default="0")
parser.add_argument("--epochs", type=int, default=3, help="nb epochs")
parser.add_argument("--regularization", type=str, default="batch_norm")
parser.add_argument(
"--nb_experiment",
type=int,
default=30,
help="number of experiment independant run for each configuration",
)
parser.add_argument(
"--learning_rate",
type=float,
default=0.01,
help="learning_rate",
)
parser.add_argument("--dropout_rate", type=float, default=0, help="dropout_rate")
args = vars(parser.parse_args())
outdir = f"./results/mnist_sgd"
alpha = args["alpha"]
regularization = args["regularization"]
if not os.path.exists(outdir):
os.mkdir(outdir)
if regularization == "relu":
file_name = f"relu_{alpha}.pkl"
elif regularization == "batch_norm":
file_name = f"batch_norm_relu_{alpha}.pkl"
elif regularization == "dropout":
file_name = f'dropout_{args["dropout_rate"]}_relu_{alpha}.pkl'
n_epochs = args["epochs"]
nb_experiment = args["nb_experiment"]
print(f"Running MNIST with ReLU'(0)={alpha} for {n_epochs} epochs")
print(f"OUTDIR: {outdir}")
print(f"File: {file_name}")
device = "cuda" if torch.cuda.is_available() else "cpu"
# Data
trainloader, testloader = get_mnist_loaders()
results_df = pd.DataFrame(
columns=[
"run_id",
"epoch",
"train_loss",
"train_accuracy",
"test_loss",
"test_accuracy",
"relu",
]
)
for k in tqdm(range(nb_experiment), desc="run_loop", leave=False):
net = init(alpha, regularization)
optimizer = optim.SGD(
net.parameters(),
lr=args["learning_rate"],
momentum=0.9,
weight_decay=5e-4,
)
scheduler = CosineAnnealingLR(optimizer, T_max=n_epochs)
for epoch in tqdm(range(n_epochs), desc="epoch_loop", leave=False):
train_loss, train_acc = train(net, optimizer, trainloader)
test_loss, test_acc = test(net, testloader)
results_df = results_df.append(
{
"run_id": k,
"epoch": epoch,
"test_loss": test_loss,
"train_loss": train_loss,
"test_accuracy": test_acc,
"train_accuracy": train_acc,
"relu": alpha,
},
ignore_index=True,
)
scheduler.step()
path = os.path.join(outdir, file_name)
results_df.to_pickle(path)