Keras Sequential API 在构建模型后用抽象类“ModuleWrapper”替换每一层答案

【问题标题】：Keras Sequential API is replacing every layer with abstract class 'ModuleWrapper' after building modelKeras Sequential API 在构建模型后用抽象类“ModuleWrapper”替换每一层
【发布时间】：2021-05-28 18:12:13
【问题描述】：

我正在尝试使用 Tensorflow 的 (2.5) Keras API 创建顺序模型。
在训练我的模型后，我发现我无法保存我的模型，因为层 ModuleWrapper 的配置没有实现，这给我带来了很多困惑，因为我不是使用任何名为“ModuleWrapper”的层。我也没有使用任何自制图层。

经过大量测试，我发现 Keras Sequential API 不知何故无法识别它自己的层，并用 Abstract Class(?) ModuleWrapper 替换它们。

任何关于为什么会发生这种情况的帮助将不胜感激！

进口

import tensorflow as tf  # version 2.5

from tensorflow import keras
from keras.layers.advanced_activations import LeakyReLU, Softmax
from keras.layers.convolutional import Conv2D, MaxPooling2D, SeparableConv2D
from keras.layers.core import Dense, Flatten, Dropout, Reshape, Activation
from keras.layers.normalization import BatchNormalization
from keras.layers.recurrent import LSTM

模型

def create_model():

  input_shape = (180, 18, 1)

  data_format = 'channels_last'
  batch_norm_axis = -1  # must be 1 if data_format = 'channels_first'
  conv_activation = 'relu'
  padding = 'same'

  model = keras.Sequential(name="CPDP_4h_1dim")

  model.add(BatchNormalization(name="batch0"))

  model.add(Conv2D(name="Conv1", filters=64, input_shape=input_shape, kernel_size=(6, 6), padding=padding, activation=conv_activation, data_format=data_format))
  model.add(BatchNormalization(name="batch1", axis=batch_norm_axis))
  model.add(MaxPooling2D(name="pool1", pool_size=(2, 2), strides=(1,1)))
  model.add(Dropout(name="dropout1", rate=0.35))

  model.add(Conv2D(name="Conv2", filters=128, kernel_size=(6, 6), padding=padding, activation=conv_activation, data_format=data_format))
  model.add(BatchNormalization(name="batch2", axis=batch_norm_axis))
  model.add(MaxPooling2D(name="pool2", pool_size=(2, 2), strides=(1,1)))
  model.add(Dropout(name="dropout2", rate=0.35))

  model.add(Conv2D(name="Conv3", filters=128, kernel_size=(3, 3), padding=padding, activation=conv_activation, data_format=data_format))
  model.add(BatchNormalization(name="batch3", axis=batch_norm_axis))
  model.add(MaxPooling2D(name="pool3", pool_size=(2, 2), strides=(1,1)))
  model.add(Dropout(name="dropout3", rate=0.15))

  model.add(Conv2D(name="Conv4", filters=256, kernel_size=(3, 3), padding=padding, activation=conv_activation, data_format=data_format))
  model.add(BatchNormalization(name="batch4", axis=batch_norm_axis))
  model.add(MaxPooling2D(name="pool4", pool_size=(2, 2), strides=(1,1)))
  model.add(Dropout(name="dropout4", rate=0.25))

  model.add(Conv2D(name="Conv5", filters=256, kernel_size=(3, 3), padding=padding, activation=conv_activation, data_format=data_format))
  model.add(BatchNormalization(name="batch5", axis=batch_norm_axis))
  model.add(MaxPooling2D(name="pool5", pool_size=(2, 2), strides=(1,1)))
  model.add(Dropout(name="dropout5", rate=0.25))


  # [batch, width, height, features]
  # width are timesteps
  # LSTM expectationms: [batch, timesteps, feature] 
  # --> transform to [batch, width, (height,features)]
  model.add(Reshape((175, 13*256), input_shape=(None, 175, 13, 256),  name="reshape_for_lstm"))
  model.add(LSTM(name="lstm1", units=512, return_sequences=True, dropout=0.25))
  model.add(LSTM(name="lstm2", units=256, return_sequences=False, dropout=0.15))

  model.add(Flatten(name="flatten1"))

  model.add(Dense(name="dense1", units=256))
  model.add(Activation('relu'))
  model.add(Dropout(name="dropout5", rate=0.15))

  model.add(Dense(name="dense15", units=256))
  model.add(Activation('relu'))
  model.add(Dropout(name="dropout51", rate=0.15))
  
  model.add(Dense(name="dense2", units=128))
  model.add(Activation('relu'))
  model.add(Dropout(name="dropout6", rate=0.15))

  model.add(Dense(name="dense3", units=64))
  model.add(Activation('relu'))
  model.add(Dropout(name="dropout7", rate=0.15))

  model.add(Dense(name="dense4", units=3))
  model.add(Activation('softmax'))
  return model

model = create_model()
model.build(input_shape=(None, 180, 18, 1))

使用`model.summary()`

Model: "CPDP_4h_1dim"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
module_wrapper_472 (ModuleWr (None, 180, 18, 1)        4         
_________________________________________________________________
module_wrapper_473 (ModuleWr (None, 180, 18, 64)       2368      
_________________________________________________________________
module_wrapper_474 (ModuleWr (None, 180, 18, 64)       256       
_________________________________________________________________
module_wrapper_475 (ModuleWr (None, 179, 17, 64)       0         
_________________________________________________________________
module_wrapper_476 (ModuleWr (None, 179, 17, 64)       0         
_________________________________________________________________
module_wrapper_477 (ModuleWr (None, 179, 17, 128)      295040    
_________________________________________________________________
module_wrapper_478 (ModuleWr (None, 179, 17, 128)      512       
_________________________________________________________________
module_wrapper_479 (ModuleWr (None, 178, 16, 128)      0         
_________________________________________________________________
module_wrapper_480 (ModuleWr (None, 178, 16, 128)      0         
_________________________________________________________________
module_wrapper_481 (ModuleWr (None, 178, 16, 128)      147584    
_________________________________________________________________
module_wrapper_482 (ModuleWr (None, 178, 16, 128)      512       
_________________________________________________________________
module_wrapper_483 (ModuleWr (None, 177, 15, 128)      0         
_________________________________________________________________
module_wrapper_484 (ModuleWr (None, 177, 15, 128)      0         
_________________________________________________________________
module_wrapper_485 (ModuleWr (None, 177, 15, 256)      295168    
_________________________________________________________________
module_wrapper_486 (ModuleWr (None, 177, 15, 256)      1024      
_________________________________________________________________
module_wrapper_487 (ModuleWr (None, 176, 14, 256)      0         
_________________________________________________________________
module_wrapper_488 (ModuleWr (None, 176, 14, 256)      0         
_________________________________________________________________
module_wrapper_489 (ModuleWr (None, 176, 14, 256)      590080    
_________________________________________________________________
module_wrapper_490 (ModuleWr (None, 176, 14, 256)      1024      
_________________________________________________________________
module_wrapper_491 (ModuleWr (None, 175, 13, 256)      0         
_________________________________________________________________
module_wrapper_492 (ModuleWr (None, 175, 13, 256)      0         
_________________________________________________________________
module_wrapper_493 (ModuleWr (None, 175, 3328)         0         
_________________________________________________________________
module_wrapper_494 (ModuleWr (None, 175, 512)          7866368   
_________________________________________________________________
module_wrapper_495 (ModuleWr (None, 256)               787456    
_________________________________________________________________
module_wrapper_496 (ModuleWr (None, 256)               0         
_________________________________________________________________
module_wrapper_497 (ModuleWr (None, 256)               65792     
_________________________________________________________________
module_wrapper_498 (ModuleWr (None, 256)               0         
_________________________________________________________________
module_wrapper_499 (ModuleWr (None, 256)               0         
_________________________________________________________________
module_wrapper_500 (ModuleWr (None, 256)               65792     
_________________________________________________________________
module_wrapper_501 (ModuleWr (None, 256)               0         
_________________________________________________________________
module_wrapper_502 (ModuleWr (None, 256)               0         
_________________________________________________________________
module_wrapper_503 (ModuleWr (None, 128)               32896     
_________________________________________________________________
module_wrapper_504 (ModuleWr (None, 128)               0         
_________________________________________________________________
module_wrapper_505 (ModuleWr (None, 128)               0         
_________________________________________________________________
module_wrapper_506 (ModuleWr (None, 64)                8256      
_________________________________________________________________
module_wrapper_507 (ModuleWr (None, 64)                0         
_________________________________________________________________
module_wrapper_508 (ModuleWr (None, 64)                0         
_________________________________________________________________
module_wrapper_509 (ModuleWr (None, 3)                 195       
_________________________________________________________________
module_wrapper_510 (ModuleWr (None, 3)                 0         
=================================================================
Total params: 10,160,327
Trainable params: 10,158,661
Non-trainable params: 1,666
_________________________________________________________________

使用`print(model.layers)`

[<tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42845faf90>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840f7f90>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840f2c90>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840f2b90>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840f2490>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42843426d0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840e3710>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840f9c90>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840fd590>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840fb310>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840f9a90>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840ed3d0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840edf90>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840ed290>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840e7a50>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840e73d0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840e4690>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840ddf10>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840c8b10>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f4284097290>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f4284097690>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f4284097950>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840a2050>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840a2ad0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840a2e50>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840aa350>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840aaad0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840aaf10>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840aad50>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840b6710>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840fb990>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840b63d0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840b6a10>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840b69d0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840c1110>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840c1e90>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f42840c12d0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f428404e1d0>,
 <tensorflow.python.keras.engine.functional.ModuleWrapper at 0x7f428404eb10>]

【问题讨论】：

标签： python tensorflow keras

【解决方案1】：

您应该按如下方式导入模块，也不要在同一导入中将 tf 2.x 与旧的独立 keras 混合。

import tensorflow as tf  # version 2.5
from tensorflow import keras
from tensorflow.keras.layers import LeakyReLU, Softmax
from tensorflow.keras.layers import Conv2D, MaxPooling2D, SeparableConv2D
from tensorflow.keras.layers import Dense, Flatten, Dropout, Reshape, Activation
from tensorflow.keras.layers import BatchNormalization
from tensorflow.keras.layers import LSTM

除此之外，模型定义中的所有层名称都应包含唯一名称。但是在你的模型中dropout5 出现了两次，所以考虑一下。

【讨论】：

进口

模型

使用model.summary()

使用print(model.layers)

使用`model.summary()`

使用`print(model.layers)`