-
Notifications
You must be signed in to change notification settings - Fork 0
/
mnist_model.py
153 lines (107 loc) · 4.8 KB
/
mnist_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
147
148
149
150
151
152
153
import torch
from torch import nn
from torch import optim
import torch.nn.init as I
# ^ source code: https://pytorch.org/docs/stable/_modules/torch/nn/init.html
import torch.nn.functional as F
# ^ source code: https://pytorch.org/docs/stable/_modules/torch/nn/functional.html
from torchvision import datasets, transforms
import math
import tqdm
import utils
device = "cuda:0" if torch.cuda.is_available() else "cpu:0"
torch.device(device)
from IPython.display import display, Markdown
class MnistModel(nn.Module):
def __init__(self):
super().__init__()
# setup the layers we want, so... `nn.Linear` is what FCN layers are called
self.x_ = nn.Linear(784, 512)
self.h1 = nn.Linear(512, 256)
self.h2 = nn.Linear(256, 128)
self.o_ = nn.Linear(128, 10)
# initialize the weights, this is generally a good habit, but not necessarily required
self.apply(MnistModel.__init_weights)
# setup dataloaders
self.loader_train = None
self.loader_test = None
# prepare for optimization and loss_fn
self.optimizer = None
self.criterion = None
# track loss and accuracy metrics
self.total_loss = []
self.total_accy = []
def forward(self, x):
# complete the forward pass
x = torch.sigmoid(self.x_(x))
x = torch.sigmoid(self.h1(x))
x = torch.sigmoid(self.h2(x))
x = F.log_softmax(self.o_(x), dim=1)
return x
def validate(self):
test_loss = 0
accuracy = 0
for images, labels in self.loader_test:
images.resize_(images.shape[0], 784)
output = self.forward(images)
test_loss = self.criterion(output, labels).item()
preds = torch.exp(output)
equality = (labels.data == preds.max(dim=1)[1])
accuracy += equality.type(torch.FloatTensor).mean()
return test_loss, accuracy
@staticmethod
def __init_weights(m):
if type(m) == nn.Linear:
I.xavier_uniform_(m.weight)
m.bias.data.fill_(0.01)
def fit(self, lr=0.001, n_epochs=10):
self.apply(MnistModel.__init_weights)
self.optimizer = optim.Adam(self.parameters(), lr=lr)
self.criterion = nn.NLLLoss()
self.total_loss = []
self.total_accy = []
for epoch in tqdm.tnrange(n_epochs):
torch.device(device)
self.train()
acc_epoch_loss = 0
steps = 0
for images, labels in self.loader_train:
steps += 1
images.resize_(images.size()[0], 784)
self.optimizer.zero_grad()
output = self.forward(images)
loss = self.criterion(output, labels)
loss.backward()
self.optimizer.step()
acc_epoch_loss += loss.item()
self.eval()
with torch.no_grad():
_, accuracy = self.validate()
self.total_loss.append(acc_epoch_loss / steps)
self.total_accy.append(accuracy / len(self.loader_test))
print(f"Epoch: {(epoch + 1):2d}, Average Loss: "
f"{self.total_loss[-1]:.3f}, Accuracy: "
f"{self.total_accy[-1]:.3f}")
def plot(self):
utils.plot_loss_acc(self.total_loss, self.total_accy)
def test(self):
self.eval()
images, labels = next(iter(self.loader_test))
image = images[0].view(1, 784)
with torch.no_grad():
output = self.forward(image)
preds = torch.exp(output)
utils.view_classify(image.view(1, 28, 28), preds, version="MNIST")
def prepare(self, batch_size=256, shuffle=True):
transform = transforms.Compose([transforms.ToTensor(),
transforms.Normalize((0.5,), (0.5,))])
data_train = datasets.MNIST("data/MNIST", train=True,
transform=transform, download=True)
self.loader_train = torch.utils.data.DataLoader(data_train,
batch_size=batch_size,
shuffle=shuffle)
data_test = datasets.MNIST("data/MNIST", train=False,
transform=transform, download=True)
self.loader_test = torch.utils.data.DataLoader(data_test,
batch_size=batch_size,
shuffle=shuffle)