1
|
|
|
import os |
2
|
|
|
|
3
|
|
|
import torch |
4
|
|
|
import torch.nn as nn |
5
|
|
|
import torch.nn.functional as F |
6
|
|
|
import torch.optim as optim |
7
|
|
|
import torch.utils.data |
8
|
|
|
from torchvision import datasets |
9
|
|
|
from torchvision import transforms |
10
|
|
|
|
11
|
|
|
from hyperactive import Hyperactive |
12
|
|
|
|
13
|
|
|
|
14
|
|
|
""" |
15
|
|
|
derived from optuna example: |
16
|
|
|
https://github.com/optuna/optuna/blob/master/examples/pytorch_simple.py |
17
|
|
|
""" |
18
|
|
|
DEVICE = torch.device("cpu") |
19
|
|
|
BATCHSIZE = 256 |
20
|
|
|
CLASSES = 10 |
21
|
|
|
DIR = os.getcwd() |
22
|
|
|
EPOCHS = 10 |
23
|
|
|
LOG_INTERVAL = 10 |
24
|
|
|
N_TRAIN_EXAMPLES = BATCHSIZE * 30 |
25
|
|
|
N_VALID_EXAMPLES = BATCHSIZE * 10 |
26
|
|
|
|
27
|
|
|
|
28
|
|
|
# Get the MNIST dataset. |
29
|
|
|
train_loader = torch.utils.data.DataLoader( |
30
|
|
|
datasets.MNIST(DIR, train=True, download=True, transform=transforms.ToTensor()), |
31
|
|
|
batch_size=BATCHSIZE, |
32
|
|
|
shuffle=True, |
33
|
|
|
) |
34
|
|
|
valid_loader = torch.utils.data.DataLoader( |
35
|
|
|
datasets.MNIST(DIR, train=False, transform=transforms.ToTensor()), |
36
|
|
|
batch_size=BATCHSIZE, |
37
|
|
|
shuffle=True, |
38
|
|
|
) |
39
|
|
|
|
40
|
|
|
|
41
|
|
|
def pytorch_cnn(params): |
42
|
|
|
linear0 = params["linear.0"] |
43
|
|
|
linear1 = params["linear.1"] |
44
|
|
|
|
45
|
|
|
layers = [] |
46
|
|
|
|
47
|
|
|
in_features = 28 * 28 |
48
|
|
|
|
49
|
|
|
layers.append(nn.Linear(in_features, linear0)) |
50
|
|
|
layers.append(nn.ReLU()) |
51
|
|
|
layers.append(nn.Dropout(0.2)) |
52
|
|
|
|
53
|
|
|
layers.append(nn.Linear(linear0, linear1)) |
54
|
|
|
layers.append(nn.ReLU()) |
55
|
|
|
layers.append(nn.Dropout(0.2)) |
56
|
|
|
|
57
|
|
|
layers.append(nn.Linear(linear1, CLASSES)) |
58
|
|
|
layers.append(nn.LogSoftmax(dim=1)) |
59
|
|
|
|
60
|
|
|
model = nn.Sequential(*layers) |
61
|
|
|
|
62
|
|
|
# model = create_model(params).to(DEVICE) |
63
|
|
|
optimizer = getattr(optim, "Adam")(model.parameters(), lr=0.01) |
64
|
|
|
|
65
|
|
|
# Training of the model. |
66
|
|
|
for epoch in range(EPOCHS): |
67
|
|
|
model.train() |
68
|
|
|
for batch_idx, (data, target) in enumerate(train_loader): |
69
|
|
|
# Limiting training data for faster epochs. |
70
|
|
|
if batch_idx * BATCHSIZE >= N_TRAIN_EXAMPLES: |
71
|
|
|
break |
72
|
|
|
|
73
|
|
|
data, target = data.view(data.size(0), -1).to(DEVICE), target.to(DEVICE) |
74
|
|
|
|
75
|
|
|
optimizer.zero_grad() |
76
|
|
|
output = model(data) |
77
|
|
|
loss = F.nll_loss(output, target) |
78
|
|
|
loss.backward() |
79
|
|
|
optimizer.step() |
80
|
|
|
|
81
|
|
|
# Validation of the model. |
82
|
|
|
model.eval() |
83
|
|
|
correct = 0 |
84
|
|
|
with torch.no_grad(): |
85
|
|
|
for batch_idx, (data, target) in enumerate(valid_loader): |
86
|
|
|
# Limiting validation data. |
87
|
|
|
if batch_idx * BATCHSIZE >= N_VALID_EXAMPLES: |
88
|
|
|
break |
89
|
|
|
data, target = data.view(data.size(0), -1).to(DEVICE), target.to(DEVICE) |
90
|
|
|
output = model(data) |
91
|
|
|
# Get the index of the max log-probability. |
92
|
|
|
pred = output.argmax(dim=1, keepdim=True) |
93
|
|
|
correct += pred.eq(target.view_as(pred)).sum().item() |
94
|
|
|
|
95
|
|
|
accuracy = correct / min(len(valid_loader.dataset), N_VALID_EXAMPLES) |
96
|
|
|
|
97
|
|
|
return accuracy |
|
|
|
|
98
|
|
|
|
99
|
|
|
|
100
|
|
|
search_space = { |
101
|
|
|
"linear.0": list(range(10, 200, 10)), |
102
|
|
|
"linear.1": list(range(10, 200, 10)), |
103
|
|
|
} |
104
|
|
|
|
105
|
|
|
|
106
|
|
|
hyper = Hyperactive() |
107
|
|
|
hyper.add_search(pytorch_cnn, search_space, n_iter=5) |
108
|
|
|
hyper.run() |
109
|
|
|
|