1 import tensorflow as tf
2 print(tf.__version__)
3
4
5 for gpu in tf.config.experimental.list_physical_devices('GPU'):
6 tf.config.experimental.set_memory_growth(gpu, True)
7
8
9 def nin_block(num_channels, kernel_size, strides, padding):
10 blk = tf.keras.models.Sequential()
11 blk.add(tf.keras.layers.Conv2D(num_channels, kernel_size, strides=strides,
12 padding=padding, activation='relu'))
13 blk.add(tf.keras.layers.Conv2D(num_channels, kernel_size=1, activation='relu'))
14 blk.add(tf.keras.layers.Conv2D(num_channels, kernel_size=1, activation='relu'))
15 return blk
16
17
18 net = tf.keras.models.Sequential()
19 net.add(nin_block(96, kernel_size=11, strides=4, padding='valid'))
20 net.add(tf.keras.layers.MaxPool2D(pool_size=3, strides=2))
21 net.add(nin_block(256, kernel_size=5, strides=1, padding='same'))
22 net.add(tf.keras.layers.MaxPool2D(pool_size=3, strides=2))
23 net.add(nin_block(384, kernel_size=3, strides=1, padding='same'))
24 net.add(tf.keras.layers.MaxPool2D(pool_size=3, strides=2))
25 net.add(tf.keras.layers.Dropout(0.5))
26 net.add(nin_block(10, kernel_size=3, strides=1, padding='same'))
27 net.add(tf.keras.layers.GlobalAveragePooling2D())
28 net.add(tf.keras.layers.Flatten())
29
30
31 X = tf.random.uniform((1, 224, 224, 1))
32 for blk in net.layers:
33 X = blk(X)
34 print(blk.name, 'output shape: \t', X.shape)
35
36
37 import numpy as np
38
39 class DataLoader():
40 def __init__(self):
41 fashion_mnist = tf.keras.datasets.fashion_mnist
42 (self.train_images, self.train_labels), (self.test_images, self.test_labels) = fashion_mnist.load_data()
43 self.train_images = np.expand_dims(self.train_images.astype(np.float32)/255.0, axis=-1)
44 self.test_images = np.expand_dims(self.test_images.astype(np.float32)/255.0, axis=-1)
45 self.train_labels = self.train_labels.astype(np.int32)
46 self.test_labels = self.test_labels.astype(np.int32)
47 self.num_train, self.num_test = self.train_images.shape[0], self.test_images.shape[0]
48
49 def get_batch_train(self, batch_size):
50 index = np.random.randint(0, np.shape(self.train_images)[0], batch_size)
51 #need to resize images to (224, 224)
52 resized_images = tf.image.resize_with_pad(self.train_images[index], 224, 224,)
53 return resized_images.numpy(), self.train_labels[index]
54
55 def get_batch_test(self, batch_size):
56 index = np.random,randint(0, np.shape(self.test_images)[0], batch_size)
57 #need to resize to (224, 224)
58 resized_images = tf.image.resize_with_pad(self.test_images[index], 224, 224,)
59 return resized_images.numpy(), self.test_labels[index]
60
61 batch_size = 128
62 dataLoader = DataLoader()
63 x_batch, y_batch = dataLoader.get_batch_train(batch_size)
64 print('x_batch shape:', x_batch.shape, 'y_batch shape:', y_batch.shape)
65
66
67
68 def train_nin():
69 #net.load_weights('5.8_nin_weights.h5')
70 epoch = 5
71 num_iter = dataLoader.num_train//batch_size
72 for e in range(epoch):
73 for n in range(num_iter):
74 x_batch, y_batch = dataLoader.get_batch_train(batch_size)
75 net.fit(x_batch, y_batch)
76 if n%200 == 0:
77 net.save_weights('5.8_nin_weights.h5')
78
79
80 #optimizer = tf.keras.optimizers.SGD(learning_rate=0.06, momentum=0.3, nesterov=False)
81 optimizer = tf.keras.optimizers.Adam(lr=1e-6)
82 net.compile(optimizer=optimizer,
83 loss='sparse_categorical_crossentropy',
84 metrics=['accuracy'])
85 x_batch, y_batch = dataLoader.get_batch_train(batch_size)
86 net.fit(x_batch, y_batch)
87 train_nin()
