使用load_model时，keras内核初始化程序被错误地调用 [英] keras kernel initializers are called incorrectly when using load_model

问题描述

Keras 2.2.4版， tensorflow 1.13.1版， 我正在使用colab笔记本

Keras version 2.2.4, tensorflow version 1.13.1, I'm using colab notebooks

我试图制作一个自定义的初始化程序，并使用model.save()保存模型，但是当我再次加载模型时，出现以下错误:

I'm trying to make a custom initializer and save the model using model.save() but when I load the model again I get the following error:

TypeError:myInit()缺少1个必需的位置参数:"input_shape"

TypeError: myInit() missing 1 required positional argument: 'input_shape'

我有以下代码:

import numpy as np

import tensorflow as tf
import keras

from google.colab import drive

from keras.models import Sequential, load_model
from keras.layers import Dense, Dropout, Flatten, Lambda, Reshape, Activation
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras import backend as K
K.set_image_data_format('channels_first')
K.backend()
# the output should be 'tensorflow'

'tensorflow'

'tensorflow'

def myInit( input_shape, dtype=None):
  weights = np.full( input_shape, 2019 )
  return K.variable( weights, dtype=dtype )

此初始值设定项被赋予input_shape并返回keras张量，如文档中所示: https://keras.io/初始化程序/

This initializer is given an input_shape and returns a keras tensor like in the docs: https://keras.io/initializers/

model = Sequential()
model.add( 
  Dense( 40, input_shape=(784,) )
)
model.add(
  Dense( 30, kernel_initializer=myInit )
)
model.add(
  Dense( 5 )
)
model.build()

权重正确初始化，因为当我调用model.layers[1].get_weights()时，得到的数组充满了2019. 我使用model.save保存模型:

The weights are initialized correctly because when I call model.layers[1].get_weights() I get an array full of 2019. I save the model using model.save:

model.save(somepath)

然后在另一个笔记本中致电

In a different notebook I then call

model = load_model(somepath, 
           custom_objects={
               'tf' : tf,
               'myInit' : myInit
           }
          )

在此笔记本myInit中，还定义了所有导入. 当我呼叫load_model时，出现以下错误:

In this notebook myInit and all the imports are defined as well. When I call load_model I get the following error:

TypeError:myInit()缺少1个必需的位置参数:"input_shape"

TypeError: myInit() missing 1 required positional argument: 'input_shape'

因此，似乎在加载模型时，input_shape没有传递给myInit.有人知道吗?

So it seems when the model is loaded, the input_shape is not passed to myInit. Does anyone have any idea?

完整跟踪:

TypeError                                 Traceback (most recent call last)

<ipython-input-25-544d137de03f> in <module>()
      2            custom_objects={
      3                'tf' : tf,
----> 4                'myInit' : myInit
      5            }
      6           )

/usr/local/lib/python3.6/dist-packages/keras/engine/saving.py in load_model(filepath, custom_objects, compile)
    417     f = h5dict(filepath, 'r')
    418     try:
--> 419         model = _deserialize_model(f, custom_objects, compile)
    420     finally:
    421         if opened_new_file:

/usr/local/lib/python3.6/dist-packages/keras/engine/saving.py in _deserialize_model(f, custom_objects, compile)
    223         raise ValueError('No model found in config.')
    224     model_config = json.loads(model_config.decode('utf-8'))
--> 225     model = model_from_config(model_config, custom_objects=custom_objects)
    226     model_weights_group = f['model_weights']
    227 

/usr/local/lib/python3.6/dist-packages/keras/engine/saving.py in model_from_config(config, custom_objects)
    456                         '`Sequential.from_config(config)`?')
    457     from ..layers import deserialize
--> 458     return deserialize(config, custom_objects=custom_objects)
    459 
    460 

/usr/local/lib/python3.6/dist-packages/keras/layers/__init__.py in deserialize(config, custom_objects)
     53                                     module_objects=globs,
     54                                     custom_objects=custom_objects,
---> 55                                     printable_module_name='layer')

/usr/local/lib/python3.6/dist-packages/keras/utils/generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)
    143                     config['config'],
    144                     custom_objects=dict(list(_GLOBAL_CUSTOM_OBJECTS.items()) +
--> 145                                         list(custom_objects.items())))
    146             with CustomObjectScope(custom_objects):
    147                 return cls.from_config(config['config'])

/usr/local/lib/python3.6/dist-packages/keras/engine/sequential.py in from_config(cls, config, custom_objects)
    298         for conf in layer_configs:
    299             layer = layer_module.deserialize(conf,
--> 300                                              custom_objects=custom_objects)
    301             model.add(layer)
    302         if not model.inputs and build_input_shape:

/usr/local/lib/python3.6/dist-packages/keras/layers/__init__.py in deserialize(config, custom_objects)
     53                                     module_objects=globs,
     54                                     custom_objects=custom_objects,
---> 55                                     printable_module_name='layer')

/usr/local/lib/python3.6/dist-packages/keras/utils/generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)
    145                                         list(custom_objects.items())))
    146             with CustomObjectScope(custom_objects):
--> 147                 return cls.from_config(config['config'])
    148         else:
    149             # Then `cls` may be a function returning a class.

/usr/local/lib/python3.6/dist-packages/keras/engine/base_layer.py in from_config(cls, config)
   1107             A layer instance.
   1108         """
-> 1109         return cls(**config)
   1110 
   1111     def count_params(self):

/usr/local/lib/python3.6/dist-packages/keras/legacy/interfaces.py in wrapper(*args, **kwargs)
     89                 warnings.warn('Update your `' + object_name + '` call to the ' +
     90                               'Keras 2 API: ' + signature, stacklevel=2)
---> 91             return func(*args, **kwargs)
     92         wrapper._original_function = func
     93         return wrapper

/usr/local/lib/python3.6/dist-packages/keras/layers/core.py in __init__(self, units, activation, use_bias, kernel_initializer, bias_initializer, kernel_regularizer, bias_regularizer, activity_regularizer, kernel_constraint, bias_constraint, **kwargs)
    846         self.activation = activations.get(activation)
    847         self.use_bias = use_bias
--> 848         self.kernel_initializer = initializers.get(kernel_initializer)
    849         self.bias_initializer = initializers.get(bias_initializer)
    850         self.kernel_regularizer = regularizers.get(kernel_regularizer)

/usr/local/lib/python3.6/dist-packages/keras/initializers.py in get(identifier)
    509     elif isinstance(identifier, six.string_types):
    510         config = {'class_name': str(identifier), 'config': {}}
--> 511         return deserialize(config)
    512     elif callable(identifier):
    513         return identifier

/usr/local/lib/python3.6/dist-packages/keras/initializers.py in deserialize(config, custom_objects)
    501                                     module_objects=globals(),
    502                                     custom_objects=custom_objects,
--> 503                                     printable_module_name='initializer')
    504 
    505 

/usr/local/lib/python3.6/dist-packages/keras/utils/generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)
    152             custom_objects = custom_objects or {}
    153             with CustomObjectScope(custom_objects):
--> 154                 return cls(**config['config'])
    155     elif isinstance(identifier, six.string_types):
    156         function_name = identifier

TypeError: myInit() missing 1 required positional argument: 'input_shape'

请注意，我也将此内容发布在> https://github.com/keras-team /keras/issues/12452 ，但我认为这是个更好的选择.

Note I also posted this on https://github.com/keras-team/keras/issues/12452 but I figured this would be a better place for this.

推荐答案

查看源代码后，我得到了以下工作代码，这应该是定义初始化程序的正确方法(尤其是在使用load_model加载模型时):

After viewing the source code I got the following working code, which should be the proper way to define an initializer (especially when loading a model with load_model):

import numpy as np

import tensorflow as tf
import keras

from google.colab import drive

from keras.models import Sequential, load_model
from keras.layers import Dense
from keras import backend as K
from keras.initializers import Initializer

K.backend()
# the output should be 'tensorflow'

class myInit( Initializer ):
  def __init__(self, myParameter):
    self.myParameter = myParameter

  def __call__(self, shape, dtype=None):

    # array filled entirely with 'myParameter'
    weights = np.full( shape, self.myParameter )
    return K.variable( weights, dtype=dtype )

  def get_config(self):
      return {
          'myParameter' : self.myParameter
      }

建立模型:

model = Sequential()
model.add( 
  Dense( 2, input_shape=(784,) )
)
model.add(
  Dense( 3, kernel_initializer=myInit( 2019 ) )
)
model.add(
  Dense( 5 )
)
model.compile(optimizer='rmsprop',
              loss='binary_crossentropy',
              metrics=['accuracy'])

保存模型:

model.save( somepath )

现在，我们可以将模型加载到其他笔记本中.从其他笔记本导入的内容也应在此处导入，并且myInit也应在此笔记本中定义.

Now we can load the model in a different notebook. The imports from the other notebook should be imported here as well and myInit should also be defined in this notebook.

model = load_model( somepath, 
           custom_objects={
               'tf' : tf,
               'myInit' : myInit
           }
          )

这篇关于使用load_model时，keras内核初始化程序被错误地调用的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！