前言

虽然TensorFlow为我们提供了很方便的API接口去实现全连接层，卷积层等等功能，但有时候，我们需要实现自己的层逻辑和网络模型逻辑。这时候，我们可以选择继承TensorFlow中的方法，实现init和call方法，完成自己的模型设计，具体代码如下，不懂的欢迎留言~

提示：以下是本篇文章正文内容，下面案例可供参考

import  tensorflow as tf
from    tensorflow.keras import datasets, layers, optimizers, Sequential, metrics
from     tensorflow import keras

def preprocess(x, y):
    """
    x is a simple image, not a batch
    """
    x = tf.cast(x, dtype=tf.float32) / 255.
    x = tf.reshape(x, [28*28])
    y = tf.cast(y, dtype=tf.int32)
    y = tf.one_hot(y, depth=10)
    return x,y

batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())

db = tf.data.Dataset.from_tensor_slices((x,y))
db = db.map(preprocess).shuffle(60000).batch(batchsz)
ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz) 

sample = next(iter(db))
print(sample[0].shape, sample[1].shape)

class MyDense(layers.Layer):

    def __init__(self, inp_dim, outp_dim):
        super(MyDense, self).__init__()

        self.kernel = self.add_variable('w', [inp_dim, outp_dim])
        self.bias = self.add_variable('b', [outp_dim])

    def call(self, inputs, training=None):

        out = inputs @ self.kernel + self.bias

        return out 

class MyModel(keras.Model):

    def __init__(self):
        super(MyModel, self).__init__()

        self.fc1 = MyDense(28*28, 256)
        self.fc2 = MyDense(256, 128)
        self.fc3 = MyDense(128, 64)
        self.fc4 = MyDense(64, 32)
        self.fc5 = MyDense(32, 10)

    def call(self, inputs, training=None):

        x = self.fc1(inputs)
        x = tf.nn.relu(x)
        x = self.fc2(x)
        x = tf.nn.relu(x)
        x = self.fc3(x)
        x = tf.nn.relu(x)
        x = self.fc4(x)
        x = tf.nn.relu(x)
        x = self.fc5(x) 

        return x

network = MyModel()

network.compile(optimizer=optimizers.Adam(lr=0.01),
        loss=tf.losses.CategoricalCrossentropy(from_logits=True),
        metrics=['accuracy']
    )

network.fit(db, epochs=5, validation_data=ds_val,
              validation_freq=2)
 
network.evaluate(ds_val)

sample = next(iter(ds_val))
x = sample[0]
y = sample[1] # one-hot
pred = network.predict(x) # [b, 10]
# convert back to number 
y = tf.argmax(y, axis=1)
pred = tf.argmax(pred, axis=1)

print(pred)
print(y)