27、CIFAR100实战

CIFAR100和cifar10的数据量一样，10是将数据集分为10类，而100是将每一类再分为10类
1、Pipeline
　　Load datasets
　　Build Network
　　Train
　　Test
2、layers 13层网络

3、代码

  1 import  tensorflow as tf
  2 from    tensorflow.keras import layers, optimizers, datasets, Sequential
  3 import  os
  4 
  5 os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
  6 tf.random.set_seed(2345)
  7 
  8 gpus = tf.config.experimental.list_physical_devices(device_type='GPU')
  9 for gpu in gpus:
 10     tf.config.experimental.set_memory_growth(gpu, True)
 11 
 12 
 13 conv_layers = [ # 5 units of conv + max pooling
 14     # unit 1  每个unit包含两个卷积层和一个最大化池化层
 15     layers.Conv2D(64, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),  #64个3x3 kernel，对图片进行扩充
 16     layers.Conv2D(64, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 17     layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),   #池化尺寸为2x2.步长为2
 18 
 19     # unit 2
 20     layers.Conv2D(128, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 21     layers.Conv2D(128, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 22     layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),
 23 
 24     # unit 3
 25     layers.Conv2D(256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 26     layers.Conv2D(256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 27     layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),
 28 
 29     # unit 4
 30     layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 31     layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 32     layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),
 33 
 34     # unit 5  保持512不变，因为参数量已经很多了
 35     layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 36     layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
 37     layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same')
 38 
 39 ]
 40 
 41 
 42 def preprocess(x, y):
 43     # [0~1] 归一化
 44     x = tf.cast(x, dtype=tf.float32) / 255.
 45     y = tf.cast(y, dtype=tf.int32)
 46     return x,y
 47 
 48 
 49 (x,y), (x_test, y_test) = datasets.cifar100.load_data()
 50 y = tf.squeeze(y, axis=1)
 51 y_test = tf.squeeze(y_test, axis=1)
 52 print(x.shape, y.shape, x_test.shape, y_test.shape)
 53 
 54 
 55 train_db = tf.data.Dataset.from_tensor_slices((x,y))
 56 train_db = train_db.shuffle(1000).map(preprocess).batch(128)
 57 
 58 test_db = tf.data.Dataset.from_tensor_slices((x_test,y_test))
 59 test_db = test_db.map(preprocess).batch(64)
 60 
 61 sample = next(iter(train_db))
 62 print('sample:', sample[0].shape, sample[1].shape,
 63       tf.reduce_min(sample[0]), tf.reduce_max(sample[0]))
 64 
 65 
 66 def main():
 67 
 68     # [b, 32, 32, 3] => [b, 1, 1, 512] 经过卷积网络之后的shape改变
 69     conv_net = Sequential(conv_layers)  #卷积网络
 70 
 71     fc_net = Sequential([
 72         layers.Dense(256, activation=tf.nn.relu),
 73         layers.Dense(128, activation=tf.nn.relu),
 74         layers.Dense(100, activation=None),  #最后一层设置为100类
 75     ])
 76 
 77     #网络的建立
 78     conv_net.build(input_shape=[None, 32, 32, 3])
 79     fc_net.build(input_shape=[None, 512])
 80     optimizer = optimizers.Adam(lr=1e-4)
 81 
 82     # [1, 2] + [3, 4] => [1, 2, 3, 4]
 83     variables = conv_net.trainable_variables + fc_net.trainable_variables  #求导的参数，两个列表相加相当于拼接
 84 
 85     for epoch in range(50):
 86 
 87         for step, (x,y) in enumerate(train_db):
 88 
 89             with tf.GradientTape() as tape:
 90                 # [b, 32, 32, 3] => [b, 1, 1, 512]
 91                 out = conv_net(x)
 92                 # flatten, => [b, 512]
 93                 out = tf.reshape(out, [-1, 512])
 94                 # [b, 512] => [b, 100]
 95                 logits = fc_net(out)
 96                 # [b] => [b, 100]
 97                 y_onehot = tf.one_hot(y, depth=100)
 98                 # compute loss
 99                 loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
100                 loss = tf.reduce_mean(loss)
101 
102             grads = tape.gradient(loss, variables)
103             optimizer.apply_gradients(zip(grads, variables))
104 
105             if step %100 == 0:
106                 print(epoch, step, 'loss:', float(loss))
107 
108 
109 
110         total_num = 0
111         total_correct = 0
112         for x,y in test_db:
113 
114             out = conv_net(x)
115             out = tf.reshape(out, [-1, 512])
116             logits = fc_net(out)
117             prob = tf.nn.softmax(logits, axis=1)
118             pred = tf.argmax(prob, axis=1)
119             pred = tf.cast(pred, dtype=tf.int32)
120 
121             correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)
122             correct = tf.reduce_sum(correct)
123 
124             total_num += x.shape[0]
125             total_correct += int(correct)
126 
127         acc = total_correct / total_num
128         print(epoch, 'acc:', acc)
129 
130 
131 
132 if __name__ == '__main__':
133     main()
posted on 2020-01-03 11:01 Luaser 阅读(3263) 评论(0) 编辑收藏举报