如何用keras自定义激活函数来训练网络模型

Question

keras 自定义激活函数(失败经验）

*按照网上的经验，使用keras自定义了激活函数，却出现了意想不到的错误，AttributeError: 'Activation' object has no attribute 'name'*，。

问题一

1 首先展示下自己具体代码

### 自定义激活函数my_f
from keras.layers import Activation
from keras import backend as K
from keras.utils.generic_utils import get_custom_objects
def my_f(x, axis=-1):
    ndim = K.ndim(x)
    if ndim==2:
        print('hahhah')
    if ndim>2:
        print('HHH')
    e = math_ops.exp(x - math_ops.reduce_max(x, axis=axis, keepdims=True))
    s = math_ops.reduce_sum(e, axis=axis, keepdims=True)
    return e / s
get_custom_objects().update({'my_f': Activation(my_f)})

### 网络框架
input1= keras.layers.Input(shape=train_data_block5_3.shape[1:])
x1 = keras.layers.GlobalAveragePooling2D()(input1)
x1= keras.layers.Dense(1024, activation='relu')(x1)
x1 = keras.layers.Dropout(0.5)(x1)
out = keras.layers.Dense(17,activation='my_f')(x1)
### 训练
from keras.callbacks import ModelCheckpoint
filepath = 'model.best.h5'
checkpoint = ModelCheckpoint(filepath, monitor='val_accuracy', verbose=1, save_best_only=True,
                            mode='max')
callbacks_list = [checkpoint]
adam = keras.optimizers.Adam(lr=0.001, decay=1e-8)
model = keras.models.Model(inputs=input1, outputs=out)
model.compile(loss='categorical_crossentropy',
              optimizer= adam,
              metrics=['accuracy'])
model.fit([train_data_block5_3, [train_labels_block5_3],
          epochs=50, batch_size=20,
          validation_data=(validation_data_block5_3, validation_labels_block5_3),
          callbacks=callbacks_list)

2 一共50个批次，往往在训练完第一个批次的时候会报错，AttributeError: 'Activation' object has no attribute 'name'

Epoch 00001: val_accuracy improved from -inf to 0.67882, saving model to E:CatV/feature/model.best.h5
Traceback (most recent call last):
  File "C:/Users/anaconda3/envs/tensorflow2/Lib/site-packages/tensorflow_core/python/keras/softmax_tiao.py", line 100, in <module>
    callbacks=callbacks_list)
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\engine\training.py", line 1239, in fit
    validation_freq=validation_freq)
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\engine\training_arrays.py", line 216, in fit_loop
    callbacks.on_epoch_end(epoch, epoch_logs)
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\callbacks\callbacks.py", line 152, in on_epoch_end
    callback.on_epoch_end(epoch, logs)
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\callbacks\callbacks.py", line 719, in on_epoch_end
    self.model.save(filepath, overwrite=True)
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\engine\network.py", line 1152, in save
    save_model(self, filepath, overwrite, include_optimizer)
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\engine\saving.py", line 449, in save_wrapper
    save_function(obj, filepath, overwrite, *args, **kwargs)
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\engine\saving.py", line 541, in save_model
    _serialize_model(model, h5dict, include_optimizer)
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\engine\saving.py", line 129, in _serialize_model
    model_config['config'] = model.get_config()
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\engine\network.py", line 879, in get_config
    layer_config = layer.get_config()
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\layers\core.py", line 925, in get_config
    'activation': activations.serialize(self.activation),
  File "C:\Users\anaconda3\envs\tensorflow2\lib\site-packages\keras\activations.py", line 200, in serialize
    return activation.__name__
AttributeError: 'Activation' object has no attribute '__name__'

Process finished with exit code 1

！

问题二

1. 如果把激活函数改成默认的softmax,能够正常的运行，可是奇怪的是我通过如下命令找到keras源文件的位置，把激活函数的py文件中定义的“softmax”部分注释掉，在jupyter notebook中仍然能够正常运行，在pycharm中则不能。

使用如下命令找到keras源文件位置

keras.__path__

['C:\\Users\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\keras']

在jupyter notebook中依然能正常运行，玄乎至极。

·