tensorflow 2.0学习笔记4

2022-01-22

字数统计: 271字 | 阅读时长≈ 1分

tensorflow学习笔记04

有错误还请各位指正！

Eager模式

增加了网络调试的灵活程度，一般情况我们就自主打开了Eager模式

自定义微分

自定义编译过程

loss

调用：tf.keras.losses.

大写的为方法

示例：

1	loss_Func = tf.keras.losses.SparseCategoricalCrossentropy()

optimizer

调用：tf.keras.optimizers.

示例：

1	optimizer = tf.keras.optimizers.Adam(0.0001)

metrics

调用：tf.keras.metrics.

示例：

1 2	train_loss = tf.keras.metrics.Mean('train_loss') train_acc = tf.keras.metrics.SparseCategoricalAccuracy('train_acc')

自定义训练过程

损失计算

1
2
3

def loss(model, x, y):
    y_pre = model(x)
    return loss_Func(y, y_pre)

每一步的训练

def train_step(model, img, lab):
    with tf.GradientTape() as t:
        loss_pre = loss(model, img, lab)
        train_loss(loss_pre)
    grad = t.gradient(loss_pre, model.variables)
    train_acc(lab, model(img))
    optimizer.apply_gradients(zip(grad, model.variables))

整个训练

def train(model,datasets,epochs):
    for epoch in range(epochs):
        for (batch, (img, lab)) in enumerate(datasets):
            train_step(model, img, lab)
        print('Eopch:{} loss is {} , acc is {}'.format(epoch,
                                         train_loss.result(),
                                         train_acc.result()))
        train_loss.reset_states()
        train_acc.reset_states()