我在 google colab 上训练了一个 keras 模型。现在无法在我的系统上本地加载它。

Posted 2023-02-23

【中文标题】我在 google colab 上训练了一个 keras 模型。现在无法在我的系统上本地加载它。

【英文标题】：I trained a keras model on google colab. Now not able to load it locally on my system.

【发布时间】：2019-04-02 17:25:05

【问题描述】：

with open('2model.json','r') as f:
json = f.read()
model = model_from_json(json)
model.load_weights("color_tensorflow_real_mode.h5")

收到此错误：ValueError: Unknown initializer: GlorotUniform

如何解决这个问题？ 每次我在 colab 上制作模型并尝试在本地加载它时，我都无法这样做。 收到此错误消息：

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-17-c3ed162a8277> in <module>()
----> 1 model = model_from_json(json)
      2 model.load_weights("color_tensorflow_real_mode.h5")

~\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\saving.py in model_from_json(json_string, custom_objects)
    349   config = json.loads(json_string)
    350   from tensorflow.python.keras.layers import deserialize  # pylint: disable=g-import-not-at-top
--> 351   return deserialize(config, custom_objects=custom_objects)
    352 
    353 

~\Anaconda3\lib\site-packages\tensorflow\python\keras\layers\serialization.py in deserialize(config, custom_objects)
     62       module_objects=globs,
     63       custom_objects=custom_objects,
---> 64       printable_module_name='layer')

~\Anaconda3\lib\site-packages\tensorflow\python\keras\utils\generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)
    171             custom_objects=dict(
    172                 list(_GLOBAL_CUSTOM_OBJECTS.items()) +
--> 173                 list(custom_objects.items())))
    174       with CustomObjectScope(custom_objects):
    175         return cls.from_config(config['config'])

~\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\network.py in from_config(cls, config, custom_objects)
   1290     # First, we create all layers and enqueue nodes to be processed
   1291     for layer_data in config['layers']:
-> 1292       process_layer(layer_data)
   1293     # Then we process nodes in order of layer depth.
   1294     # Nodes that cannot yet be processed (if the inbound node

~\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\network.py in process_layer(layer_data)
   1276       from tensorflow.python.keras.layers import deserialize as deserialize_layer  # pylint: disable=g-import-not-at-top
   1277 
-> 1278       layer = deserialize_layer(layer_data, custom_objects=custom_objects)
   1279       created_layers[layer_name] = layer
   1280 

~\Anaconda3\lib\site-packages\tensorflow\python\keras\layers\serialization.py in deserialize(config, custom_objects)
     62       module_objects=globs,
     63       custom_objects=custom_objects,
---> 64       printable_module_name='layer')

~\Anaconda3\lib\site-packages\tensorflow\python\keras\utils\generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)
    173                 list(custom_objects.items())))
    174       with CustomObjectScope(custom_objects):
--> 175         return cls.from_config(config['config'])
    176     else:
    177       # Then `cls` may be a function returning a class.

~\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\base_layer.py in from_config(cls, config)
   1615         A layer instance.
   1616     """
-> 1617     return cls(**config)
   1618 
   1619 

~\Anaconda3\lib\site-packages\tensorflow\python\keras\layers\convolutional.py in __init__(self, filters, kernel_size, strides, padding, data_format, dilation_rate, activation, use_bias, kernel_initializer, bias_initializer, kernel_regularizer, bias_regularizer, activity_regularizer, kernel_constraint, bias_constraint, **kwargs)
    464         activation=activations.get(activation),
    465         use_bias=use_bias,
--> 466         kernel_initializer=initializers.get(kernel_initializer),
    467         bias_initializer=initializers.get(bias_initializer),
    468         kernel_regularizer=regularizers.get(kernel_regularizer),

~\Anaconda3\lib\site-packages\tensorflow\python\keras\initializers.py in get(identifier)
    153     return None
    154   if isinstance(identifier, dict):
--> 155     return deserialize(identifier)
    156   elif isinstance(identifier, six.string_types):
    157     config = 'class_name': str(identifier), 'config': 

~\Anaconda3\lib\site-packages\tensorflow\python\keras\initializers.py in deserialize(config, custom_objects)
    145       module_objects=globals(),
    146       custom_objects=custom_objects,
--> 147       printable_module_name='initializer')
    148 
    149 

~\Anaconda3\lib\site-packages\tensorflow\python\keras\utils\generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)
    161       cls = module_objects.get(class_name)
    162       if cls is None:
--> 163         raise ValueError('Unknown ' + printable_module_name + ': ' + class_name)
    164     if hasattr(cls, 'from_config'):
    165       arg_spec = tf_inspect.getfullargspec(cls.from_config)

ValueError: Unknown initializer: GlorotUniform

*** 要求我添加详细信息，而我没有要添加的内容。或者我不确定要添加什么。请帮忙。

【问题讨论】：

您在 Colab 和本地机器中是否有相同的 TensorFlow 版本？

【参考方案1】：

使用加载模型

 from tensorflow.keras.models import load_model

而不是

from keras.models import load_model

我尝试了很多方法，但这是最终奏效的方法！

【讨论】：

非常感谢您的帖子！我使用： from keras.models import load_model 它适用于使用 Google AI 平台“runtimeVersion”：“2.2”和“pythonVersion”：“3.7”生成的模型

我想提供更多关于它是如何工作的细节：模型是在运行时 2.2 - tensorflow 2.2 上生成的。要加载它，我使用以下命令： pip install tensorflow==1.11.0 pip install keras==2.4.0 from keras.models import load_model new_model = load_model(local_model)

【参考方案2】：

当我尝试本地加载在 Colab 上训练的模型时，我遇到了类似的错误（未知层：名称）。我试图更改 keras 版本、tensorflow 版本、conda 版本等，但没有任何帮助。我通过在 Colab 上保存模型的权重、在本地创建相同的模型并将权重加载到该模型来解决了这个问题。

【参考方案3】：

通过运行pip install keras tensorflow 或conda install keras tensorflow，确保您拥有最新版本的Keras 和tensorflow（即2.4.4 和1.11.0）。

如果是 Google Colab 使用已弃用的对象，您可能需要使用自定义对象：

from keras.utils import CustomObjectScope
from keras.initializers import glorot_uniform

with CustomObjectScope('GlorotUniform': glorot_uniform()):
    model = load_model('my_model.h5')

不确定这是否是您的情况。