目录

• CIFAR10
• MyDenseLayer

CIFAR10

MyDenseLayer

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

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'


def preprocess(x, y):
    # [0, 255] --> [-1,1]
    x = 2 * tf.cast(x, dtype=tf.float32) / 255. - 1
    y = tf.cast(y, dtype=tf.int32)

    return x, y


batch_size = 128
# x --> [32,32,3], y --> [10k, 1]
(x, y), (x_val, y_val) = datasets.cifar10.load_data()
y = tf.squeeze(y)  # [10k, 1] --> [10k]
y_val = tf.squeeze(y_val)
y = tf.one_hot(y, depth=10)  # [50k, 10]
y_val = tf.one_hot(y_val, depth=10)  # [10k, 10]
print('datasets:', x.shape, y.shape, x_val.shape, y_val.shape, x.min(),
      x.max())

train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.map(preprocess).shuffle(10000).batch(batch_size)
test_db = tf.data.Dataset.from_tensor_slices((x_val, y_val))
test_db = test_db.map(preprocess).batch(batch_size)

sample = next(iter(train_db))
print('batch:', sample[0].shape, sample[1].shape)


class MyDense(layers.Layer):
    # to replace standard layers.Dense()
    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):
        x = inputs @ self.kernel
        return x


class MyNetwork(keras.Model):
    def __init__(self):
        super(MyNetwork, self).__init__()
        self.fc1 = MyDense(32 * 32 * 3, 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):
        """inputs: [b,32,32,32,3]"""
        x = tf.reshape(inputs, [-1, 32 * 32 * 3])
        # [b,32*32*32] --> [b, 256]
        x = self.fc1(x)
        x = tf.nn.relu(x)
        # [b, 256] --> [b,128]
        x = self.fc2(x)
        x = tf.nn.relu(x)
        # [b, 128] --> [b,64]
        x = self.fc3(x)
        x = tf.nn.relu(x)
        # [b, 64] --> [b,32]
        x = self.fc4(x)
        x = tf.nn.relu(x)
        # [b, 32] --> [b,10]
        x = self.fc5(x)

        return x


network = MyNetwork()
network.compile(optimizer=optimizers.Adam(lr=1e-3),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy'])
network.fit(train_db, epochs=5, validation_data=test_db, validation_freq=1)

network.evaluate(test_db)
network.save_weights('weights.ckpt')
del network
print('saved to ckpt/weights.ckpt')

network = MyNetwork()
network.compile(optimizer=optimizers.Adam(lr=1e-3),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metircs=['accuracy'])
network.fit(train_db, epochs=5, validation_data=test_db, validation_freq=1)
network.load_weights('weights.ckpt')
print('loaded weights from file.')

network.evaluate(test_db)

datasets: (50000, 32, 32, 3) (50000, 10) (10000, 32, 32, 3) (10000, 10) 0 255
batch: (128, 32, 32, 3) (128, 10)
Epoch 1/5
391/391 [==============================] - 7s 19ms/step - loss: 1.7276 - accuracy: 0.3358 - val_loss: 1.5801 - val_accuracy: 0.4427
Epoch 2/5
391/391 [==============================] - 7s 18ms/step - loss: 1.5045 - accuracy: 0.4606 - val_loss: 1.4808 - val_accuracy: 0.4812
Epoch 3/5
391/391 [==============================] - 6s 17ms/step - loss: 1.3919 - accuracy: 0.5019 - val_loss: 1.4596 - val_accuracy: 0.4921
Epoch 4/5
391/391 [==============================] - 7s 18ms/step - loss: 1.3039 - accuracy: 0.5364 - val_loss: 1.4651 - val_accuracy: 0.4950
Epoch 5/5
391/391 [==============================] - 6s 16ms/step - loss: 1.2270 - accuracy: 0.5622 - val_loss: 1.4483 - val_accuracy: 0.5030
79/79 [==============================] - 1s 11ms/step - loss: 1.4483 - accuracy: 0.5030
saved to ckpt/weights.ckpt
Epoch 1/5
391/391 [==============================] - 7s 19ms/step - loss: 1.7216 - val_loss: 1.5773
Epoch 2/5
391/391 [==============================] - 10s 26ms/step - loss: 1.5010 - val_loss: 1.5111
Epoch 3/5
391/391 [==============================] - 8s 21ms/step - loss: 1.3868 - val_loss: 1.4657
Epoch 4/5
391/391 [==============================] - 8s 20ms/step - loss: 1.3021 - val_loss: 1.4586
Epoch 5/5
391/391 [==============================] - 7s 17ms/step - loss: 1.2276 - val_loss: 1.4583
loaded weights from file.
79/79 [==============================] - 1s 12ms/step - loss: 1.4483





1.4482733222502697