Spaces:
Runtime error
Runtime error
Upload model.py
Browse files
model.py
ADDED
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import time
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
import torchvision.datasets as datasets
|
7 |
+
import torchvision.transforms as transforms
|
8 |
+
from torch.utils.data import DataLoader
|
9 |
+
import torch.optim as optim
|
10 |
+
|
11 |
+
from tqdm import tqdm
|
12 |
+
|
13 |
+
|
14 |
+
def get_mean_std(loader):
|
15 |
+
'''
|
16 |
+
Calculates mean and std of input images.
|
17 |
+
|
18 |
+
Args:
|
19 |
+
loader (torch.DataLoader): Loader with images
|
20 |
+
|
21 |
+
Returns:
|
22 |
+
mean (torch.Tensor): Mean of images in loader
|
23 |
+
std (torch.Tensor): Standard deviation of images in loader
|
24 |
+
'''
|
25 |
+
channels_sum, channels_squared_sum, num_batches = 0, 0, 0
|
26 |
+
|
27 |
+
for data, _ in loader:
|
28 |
+
channels_sum += torch.mean(data, dim=[0,2,3]) # mean across [no. of examples, height, width]
|
29 |
+
channels_squared_sum += torch.mean(data**2, dim=[0,2,3]) # squared mean across [no. of examples, height, width]
|
30 |
+
num_batches += 1
|
31 |
+
|
32 |
+
mean = channels_sum/num_batches
|
33 |
+
std = (channels_squared_sum/(num_batches-mean**2))**0.5
|
34 |
+
|
35 |
+
return mean, std
|
36 |
+
|
37 |
+
|
38 |
+
class Net(nn.Module):
|
39 |
+
'''
|
40 |
+
model definition
|
41 |
+
'''
|
42 |
+
|
43 |
+
def __init__(self):
|
44 |
+
super(Net, self).__init__()
|
45 |
+
|
46 |
+
self.layer1 = nn.Sequential(
|
47 |
+
nn.Conv2d(1, 32, kernel_size=5),
|
48 |
+
nn.ReLU(),
|
49 |
+
)
|
50 |
+
self.layer2 = nn.Sequential(
|
51 |
+
nn.Conv2d(32, 32, kernel_size=5, bias=False),
|
52 |
+
nn.BatchNorm2d(32),
|
53 |
+
nn.ReLU(),
|
54 |
+
nn.MaxPool2d((2, 2)),
|
55 |
+
nn.Dropout2d(0.25),
|
56 |
+
)
|
57 |
+
self.layer3 = nn.Sequential(
|
58 |
+
nn.Conv2d(32, 64, kernel_size=3),
|
59 |
+
nn.ReLU(),
|
60 |
+
)
|
61 |
+
self.layer4 = nn.Sequential(
|
62 |
+
nn.Conv2d(64, 64, kernel_size=3, bias=False),
|
63 |
+
nn.BatchNorm2d(64),
|
64 |
+
nn.ReLU(),
|
65 |
+
nn.MaxPool2d((2, 2)),
|
66 |
+
nn.Dropout2d(0.25),
|
67 |
+
nn.Flatten(),
|
68 |
+
)
|
69 |
+
self.layer5 = nn.Sequential(
|
70 |
+
nn.Linear(576, 256, bias=False),
|
71 |
+
nn.BatchNorm1d(256),
|
72 |
+
nn.ReLU(),
|
73 |
+
)
|
74 |
+
self.layer6 = nn.Sequential(
|
75 |
+
nn.Linear(256, 128, bias=False),
|
76 |
+
nn.BatchNorm1d(128),
|
77 |
+
nn.ReLU(),
|
78 |
+
)
|
79 |
+
self.layer7 = nn.Sequential(
|
80 |
+
nn.Linear(128, 84, bias=False),
|
81 |
+
nn.BatchNorm1d(84),
|
82 |
+
nn.ReLU(),
|
83 |
+
nn.Dropout(0.25),
|
84 |
+
)
|
85 |
+
self.layer8 = nn.Sequential(
|
86 |
+
nn.Linear(84, 10),
|
87 |
+
nn.LogSoftmax(dim=1),
|
88 |
+
)
|
89 |
+
|
90 |
+
def forward(self, x):
|
91 |
+
x = transforms.Normalize(mean, std)(x)
|
92 |
+
x = self.layer1(x)
|
93 |
+
x = self.layer2(x)
|
94 |
+
x = self.layer3(x)
|
95 |
+
x = self.layer4(x)
|
96 |
+
x = self.layer5(x)
|
97 |
+
x = self.layer6(x)
|
98 |
+
x = self.layer7(x)
|
99 |
+
x = self.layer8(x)
|
100 |
+
|
101 |
+
return x
|
102 |
+
|
103 |
+
|
104 |
+
# downloads and loads MNIST train set
|
105 |
+
transform = transforms.Compose([transforms.ToTensor(), transforms.RandomAffine(degrees=10, translate=(0.1,0.1))])
|
106 |
+
train_data = datasets.MNIST(root='./data', train=True, download=True, transform=transform)
|
107 |
+
train_loader = DataLoader(dataset=train_data, batch_size=64, shuffle=True, pin_memory=True)
|
108 |
+
|
109 |
+
# downloads and loads MNIST test set
|
110 |
+
val_data = datasets.MNIST(root='./data', train=False, download=True, transform=transforms.ToTensor())
|
111 |
+
val_loader = DataLoader(dataset=train_data, batch_size=64, shuffle=False, pin_memory=True)
|
112 |
+
|
113 |
+
# uses GPU if available
|
114 |
+
if torch.cuda.is_available():
|
115 |
+
dev = "cuda:0"
|
116 |
+
else:
|
117 |
+
dev = "cpu"
|
118 |
+
|
119 |
+
device = torch.device(dev)
|
120 |
+
|
121 |
+
# gets mean and std of dataset
|
122 |
+
mean, std = get_mean_std(train_loader)
|
123 |
+
|
124 |
+
|
125 |
+
def run_model():
|
126 |
+
# defines parameters
|
127 |
+
model = Net().to(device=device)
|
128 |
+
optimizer = optim.Adam(model.parameters(), lr=0.1)
|
129 |
+
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=0.2, patience=2)
|
130 |
+
criterion = nn.NLLLoss()
|
131 |
+
|
132 |
+
# iterates through epochs
|
133 |
+
for epoch in range(30):
|
134 |
+
print(f"\nEpoch {epoch+1}/{30}.")
|
135 |
+
|
136 |
+
# train loop
|
137 |
+
model.train()
|
138 |
+
|
139 |
+
total_train_loss = 0
|
140 |
+
total_correct = 0
|
141 |
+
|
142 |
+
for i, (images, labels) in enumerate(tqdm(train_loader)):
|
143 |
+
images = images.to(device)
|
144 |
+
labels = labels.to(device)
|
145 |
+
|
146 |
+
optimizer.zero_grad()
|
147 |
+
|
148 |
+
outputs = model(images)
|
149 |
+
|
150 |
+
loss = criterion(outputs, labels)
|
151 |
+
total_train_loss += loss.item()
|
152 |
+
|
153 |
+
loss.backward()
|
154 |
+
optimizer.step()
|
155 |
+
|
156 |
+
# Calculates train accuracy
|
157 |
+
outputs_probs = nn.functional.softmax(
|
158 |
+
outputs, dim=1) # gets probabilities
|
159 |
+
for idx, preds in enumerate(outputs_probs):
|
160 |
+
# if label with max probability matches true label
|
161 |
+
if labels[idx] == torch.argmax(preds.data):
|
162 |
+
total_correct += 1
|
163 |
+
|
164 |
+
train_loss = total_train_loss/(i+1)
|
165 |
+
train_accuracy = total_correct/len(train_data)
|
166 |
+
|
167 |
+
print(f"Train set:- Loss: {train_loss}, Accuracy: {train_accuracy}.")
|
168 |
+
|
169 |
+
# saves model state
|
170 |
+
if not os.path.exists("./saved_models"):
|
171 |
+
os.mkdir("./saved_models")
|
172 |
+
torch.save(model.state_dict(), f"./saved_models/mnist-cnn-{time.time()}.pt")
|
173 |
+
|
174 |
+
# val loop
|
175 |
+
model.eval()
|
176 |
+
|
177 |
+
total_val_loss = 0
|
178 |
+
total_correct = 0
|
179 |
+
|
180 |
+
with torch.no_grad():
|
181 |
+
for i, (images, labels) in enumerate(tqdm(val_loader)):
|
182 |
+
images = images.to(device)
|
183 |
+
labels = labels.to(device)
|
184 |
+
|
185 |
+
outputs = model(images)
|
186 |
+
|
187 |
+
loss = criterion(outputs, labels)
|
188 |
+
total_val_loss += loss.item()
|
189 |
+
|
190 |
+
outputs_probs = nn.functional.softmax(outputs, dim=1)
|
191 |
+
for idx, preds in enumerate(outputs_probs):
|
192 |
+
if labels[idx] == torch.argmax(preds.data):
|
193 |
+
total_correct += 1
|
194 |
+
|
195 |
+
val_loss = total_val_loss/(i+1)
|
196 |
+
val_accuracy = total_correct/len(val_data)
|
197 |
+
|
198 |
+
print(f"Val set:- Loss: {val_loss}, Accuracy: {val_accuracy}.")
|
199 |
+
|
200 |
+
# adjusts lr
|
201 |
+
scheduler.step(val_loss)
|