mnistでの畳み込みネットワークが動かない

前提・実現したいこと

畳み込みネットワークを導入したいのですが、mnistで行うとエラーが出ます

発生している問題・エラーメッセージ

Invalid operation is performed in: Convolution2DFunction (Forward)

Expect: in_types[0].ndim == 4
Actual: 2 != 4
と出てしまいます

該当のソースコード

python
1from chainer.datasets import mnist
2import matplotlib.pyplot as plt
3from chainer.datasets import split_dataset_random
4from chainer import iterators
5from chainer import optimizers
6from chainer import serializers
7from chainer import training
8import random
9import numpy
10import chainer
11import chainer.links as L
12import chainer.functions as F
13
14
15def reset_seed(seed=0):
16	random.seed(seed)
17	numpy.random.seed(seed)
18	if chainer.cuda.available:
19		chainer.cuda.cupy.random.seed(seed)
20		
21reset_seed(0)
22
23train_val, test = mnist.get_mnist()
24train, valid = split_dataset_random(train_val, 50000, seed=0)
25
26batchsize = 128
27
28train_iter = iterators.SerialIterator(train, batchsize)
29valid_iter = iterators.SerialIterator(valid, batchsize, False, False)
30test_iter = iterators.SerialIterator(test, batchsize, False, False)
31
32
33class MyNet(chainer.Chain):
34
35    def __init__(self, n_out):
36        super(MyNet, self).__init__()
37        with self.init_scope():
38            self.conv1 = L.Convolution2D(None, 28, 3, 3, 1)
39            self.conv2 = L.Convolution2D(28,28, 3, 3, 1)
40            self.conv3 = L.Convolution2D(28,28, 3, 3, 1)
41            self.fc4 = L.Linear(None, 1000)
42            self.fc5 = L.Linear(1000, n_out)
43
44    def __call__(self, x):
45        h = F.relu(self.conv1(x))
46        h = F.relu(self.conv2(h))
47        h = F.relu(self.conv3(h))
48        h = F.relu(self.fc4(h))
49        h = self.fc5(h)
50        return h
51
52
53gpu_id = 0  # CPUを用いたい場合は、-1を指定してください
54
55lr_decay=None
56
57net = MyNet(10)
58
59if gpu_id >= 0:
60    net.to_gpu(gpu_id)
61
62net = L.Classifier(net)
63
64optimizer = optimizers.SGD(lr=0.01).setup(net)
65
66updater = training.StandardUpdater(train_iter, optimizer, device=gpu_id)
67
68max_epoch = 10
69
70trainer = training.Trainer(
71    updater, (max_epoch, 'epoch'), out='mnist_MyNet_result')
72
73from chainer.training import extensions
74
75trainer.extend(extensions.LogReport())
76trainer.extend(extensions.snapshot(filename='snapshot_epoch-{.updater.epoch}'))
77trainer.extend(extensions.Evaluator(valid_iter, net, device=gpu_id), name='val')
78trainer.extend(extensions.PrintReport(['epoch', 'main/loss', 'main/accuracy', 'val/main/loss', 'val/main/accuracy', 'elapsed_time','lr']))
79trainer.extend(extensions.PlotReport(['main/loss', 'val/main/loss'], x_key='epoch', file_name='loss.png'))
80trainer.extend(extensions.PlotReport(['main/accuracy', 'val/main/accuracy'], x_key='epoch', file_name='accuracy.png'))
81if lr_decay is not None:
82	trainer.extend(extensions.ExponentialShift('lr', 0.1), trigger=lr_decay)
83trainer.run()
84del trainer
85
86trainer.run()
87
88test_evaluator = extensions.Evaluator(test_iter, net, device=gpu_id)
89results = test_evaluator()
90print('Test accuracy:', results['main/accuracy'])
91
92#以下予測用のコード
93
94reset_seed(0)
95
96infer_net = MyNet(10)
97serializers.load_npz(
98    'mnist_MyNet_result/snapshot_epoch-10',
99    infer_net, path='updater/model:main/predictor/')
100
101gpu_id=0
102
103if gpu_id >= 0:
104    infer_net.to_gpu(gpu_id)
105
106x, t = test[0]
107plt.imshow(x.reshape(28, 28), cmap='gray')
108plt.show()
109
110x = infer_net.xp.asarray(x[None, ...])
111with chainer.using_config('train', False), chainer.using_config('enable_backprop', False):
112    y = infer_net(x)
113y = y.array
114
115print('予測ラベル:', y.argmax(axis=1)[0])