-
Notifications
You must be signed in to change notification settings - Fork 12
/
train_numpy_model.py
146 lines (126 loc) · 4.63 KB
/
train_numpy_model.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
import torchvision
import argparse
import time
from cnn.transforms import PIL2numpy, Normalize, OneHot
from cnn.numpy_model import (
Conv2d, ReLU, Sigmoid, Softmax, Maxpool2d, Flatten, Linear,
CrossEntropyLoss
)
class CnnFromScratch:
def __init__(self):
self.conv1 = Conv2d(1, 2, 3, 1)
self.conv2 = Conv2d(2, 5, 2, 2, padding=1)
self.max_pool = Maxpool2d(2, 2, padding=1)
self.fc1 = Linear(320, 1000)
self.fc2 = Linear(1000, 10)
self.flatten = Flatten()
self.relu = ReLU()
self.sigmoid1 = Sigmoid()
self.sigmoid2 = Sigmoid()
self.softmax = Softmax()
def load_weights(self, load_path):
self.conv1.load_weights('conv_w_1', 'conv_b_1', load_path)
self.conv2.load_weights('conv_w_2', 'conv_b_2', load_path)
self.fc1.load_weights('fc_w_1', 'fc_b_1', load_path)
self.fc2.load_weights('fc_w_2', 'fc_b_2', load_path)
def __call__(self, x):
x = self.conv1(x)
x = self.relu(x)
x = self.max_pool(x)
x = self.conv2(x)
x = self.sigmoid1(x)
x = self.flatten.matrices2vector(x)
x = self.fc1(x)
x = self.sigmoid2(x)
x = self.fc2(x)
x = self.softmax(x)
return x
def backprop(self, x, lr=0.01):
x = self.softmax.backprop(x)
x = self.fc2.backprop(x, lr)
x = self.sigmoid2.backprop(x)
x = self.fc1.backprop(x, lr)
x = self.flatten.vector2matrices(x)
x = self.sigmoid1.backprop(x)
x = self.conv2.backprop(x, lr)
x = self.max_pool.backprop(x)
x = self.relu.backprop(x)
x = self.conv1.backprop(x, lr)
def get_train_dataset():
transforms = torchvision.transforms.Compose([
PIL2numpy(),
Normalize(),
])
target_transform = torchvision.transforms.Compose([
OneHot()
])
train_dataset = torchvision.datasets.MNIST(
root='/workdir/data',
train=True,
download=True,
transform=transforms,
target_transform=target_transform
)
test_dataset = torchvision.datasets.MNIST(
root='/workdir/data',
train=False,
download=True,
transform=transforms,
target_transform=target_transform
)
return train_dataset, test_dataset
def train_loop(dataset, model, criterion, print_log_freq, lr):
loss_log = []
acc_log = []
start_time = time.time()
for idx, (image, target) in enumerate(dataset):
pred = model([image])
loss = criterion(target, pred)
x = criterion.backprop(target, pred)
model.backprop(x, lr=lr)
loss_log.append(loss.sum())
acc_log.append(pred.argmax() == target.argmax())
if idx % print_log_freq == 0:
loss_avg = sum(loss_log[-print_log_freq:])/print_log_freq
acc_avg = sum(acc_log[-print_log_freq:])/print_log_freq
loop_time = time.time() - start_time
start_time = time.time()
print(f'Train step {idx}, Loss: {loss_avg:.5f}, '
f'Acc: {acc_avg:.4f}, time: {loop_time:.1f}')
def val_loop(dataset, model, criterion):
loss_log = []
acc_log = []
start_time = time.time()
for idx, (image, target) in enumerate(dataset):
pred = model([image])
loss = criterion(target, pred)
loss_log.append(loss.sum())
acc_log.append(pred.argmax() == target.argmax())
loss_avg = sum(loss_log)/len(loss_log)
acc_avg = sum(acc_log)/len(acc_log)
loop_time = time.time() - start_time
print(f'Val step, Loss: {loss_avg:.5f}, '
f'Acc: {acc_avg:.4f}, time: {loop_time:.1f}')
def main(args):
train_dataset, test_dataset = get_train_dataset()
model = CnnFromScratch()
if args.load_path:
model.load_weights(args.load_path)
print('Numpy model weights were loaded')
criterion = CrossEntropyLoss()
for epoch in range(args.num_epochs):
train_loop(train_dataset, model, criterion,
args.print_log_freq, args.lr)
val_loop(test_dataset, model, criterion)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--print_log_freq', type=int, default=1000,
help='Frequency of printing of training logs')
parser.add_argument('--load_path', type=str, default='',
help='Path to model weights to start training with')
parser.add_argument('--num_epochs', type=int, default=10,
help='Total number of epochs')
parser.add_argument('--lr', type=float, default=0.01,
help='learning rate')
args = parser.parse_args()
main(args)