如何在TF2中动态更新批处理规范动量？

Question

我找到了一个PyTorch实现，它将批处理规范momentum参数从第一个时期的0.1衰减到最后一个时期的0.001。对于如何在momentum中使用批处理规范TF2参数，有什么建议吗？(例如，从0.9开始，以0.999结束)，这就是PyTorch代码中所做的：

# in training script
momentum = initial_momentum * np.exp(-epoch/args.epochs * np.log(initial_momentum/final_momentum))
model_pos_train.set_bn_momentum(momentum)

# model class function
def set_bn_momentum(self, momentum):
    self.expand_bn.momentum = momentum
    for bn in self.layers_bn:
        bn.momentum = momentum

解决方案：

下面所选的答案提供了使用tf.keras.Model.fit() API时的可行解决方案。然而，我使用的是一个定制的训练循环。以下是我所做的：

在每一个时代之后：

mi = 1 - initial_momentum  # i.e., inital_momentum = 0.9, mi = 0.1
mf = 1 - final_momentum  # i.e., final_momentum = 0.999, mf = 0.001
momentum = 1 - mi * np.exp(-epoch / epochs * np.log(mi / mf))
model = set_bn_momentum(model, momentum)

set_bn_momentum函数(归功于这篇文章)：

def set_bn_momentum(model, momentum):
    for layer in model.layers:
        if hasattr(layer, 'momentum'):
            print(layer.name, layer.momentum)
            setattr(layer, 'momentum', momentum)

    # When we change the layers attributes, the change only happens in the model config file
    model_json = model.to_json()

    # Save the weights before reloading the model.
    tmp_weights_path = os.path.join(tempfile.gettempdir(), 'tmp_weights.h5')
    model.save_weights(tmp_weights_path)

    # load the model from the config
    model = tf.keras.models.model_from_json(model_json)

    # Reload the model weights
    model.load_weights(tmp_weights_path, by_name=True)
    return model

这种方法没有给训练例程增加很大的开销。

Answer 1

您可以在每个批处理的开始/结束中设置一个操作，这样您就可以在这个时期控制任何参数。

以下是回调的选项：

class CustomCallback(keras.callbacks.Callback):
    def on_epoch_begin(self, epoch, logs=None):
        keys = list(logs.keys())
        print("Start epoch {} of training; got log keys: {}".format(epoch, keys))

    def on_epoch_end(self, epoch, logs=None):
        keys = list(logs.keys())
        print("End epoch {} of training; got log keys: {}".format(epoch, keys))

    def on_train_batch_begin(self, batch, logs=None):
        keys = list(logs.keys())
        print("...Training: start of batch {}; got log keys: {}".format(batch, keys))

    def on_train_batch_end(self, batch, logs=None):
        keys = list(logs.keys())
        print("...Training: end of batch {}; got log keys: {}".format(batch, keys))

    def on_test_batch_begin(self, batch, logs=None):
        keys = list(logs.keys())
        print("...Evaluating: start of batch {}; got log keys: {}".format(batch, keys))

    def on_test_batch_end(self, batch, logs=None):
        keys = list(logs.keys())
        print("...Evaluating: end of batch {}; got log keys: {}".format(batch, keys))

你可以获得动量

batch = tf.keras.layers.BatchNormalization()
batch.momentum = 0.001

在模型中，必须指定正确的层。

model.layers[1].momentum = 0.001

您可以在回调找到更多的信息和示例。