如何在`call`方法中调用加载的Tensorflow模型并传递自定义参数？

huangapple
go评论100阅读模式
英文:

How to call loaded Tensorflow model with custom params in call method?

问题

以下是要翻译的代码部分：

  1. I've been following the [Tensorflow text generation tutorial][1] which includes two models, "MyModel" and "OneStep". "MyModel" is an RNN operating on vectorized strings; "OneStep" essentially wraps "MyModel" and operates on strings directly.
  3. The tutorial saves and loads "OneStep", and I followed this successfully, but I now want to save and reload "MyModel". This is not done in the tutorial, and when I try to call the reloaded model with `return_state=True`, I get an error:
  5. # TODO: Loaded model gives an error
  6. for input_example_batch, target_example_batch in train_ds.take(1):
  7.     example_batch_predictions, example_states = loaded_model(input_example_batch, False, None, return_state=True)
  8.     print(example_batch_predictions.shape, "# (batch_size, sequence_length, vocab_size)")
  9.     print(example_states.shape, "   # (batch_size, rnn_units)")
  11. /opt/conda/lib/python3.7/site-packages/tensorflow/python/saved_model/load.py in _call_attribute(instance, *args, **kwargs)
  12.     662 
  14. def custom_model(inputs, custom_param=False):
  15.     return dense(inputs)
  17. model = CustomModel()
  19. sample_inputs = tf.zeros((16, 30))
  20. print('Sample inputs:', sample_inputs)
  22. sample_outputs = model(sample_inputs)
  23. print('Sample outputs:', sample_outputs)
  25. model.save('saved_model')
  26. loaded_model = tf.keras.models.load_model('saved_model')
  28. sample_outputs_2 = loaded_model(sample_inputs, custom_param=True)
  29. print('Sample outputs 2:', sample_outputs_2)

如果您有任何其他问题或需要进一步的帮助，请随时提问。

英文:

I've been following the Tensorflow text generation tutorial which includes two models, "MyModel" and "OneStep". "MyModel" is an RNN operating on vectorized strings; "OneStep" essentially wraps "MyModel" and operates on strings directly.

The tutorial saves and loads "OneStep", and I followed this successfully, but I now want to save and reload "MyModel". This is not done in the tutorial, and when I try to call the reloaded model with return_state=True, I get an error:

  1. ---------------------------------------------------------------------------
  2. ValueError                                Traceback (most recent call last)
  3. /tmp/ipykernel_23/2335414736.py in <module>
  4.       1 # TODO: Loaded model gives an error
  5.       2 for input_example_batch, target_example_batch in train_ds.take(1):
  6. ----> 3     example_batch_predictions, example_states = loaded_model(input_example_batch, False, None, return_state=True)
  7.       4     print(example_batch_predictions.shape, "# (batch_size, sequence_length, vocab_size)")
  8.       5     print(example_states.shape, "   # (batch_size, rnn_units)")
  10. /opt/conda/lib/python3.7/site-packages/tensorflow/python/saved_model/load.py in _call_attribute(instance, *args, **kwargs)
  11.     662 
  12.     663 def _call_attribute(instance, *args, **kwargs):
  13. --> 664   return instance.__call__(*args, **kwargs)
  14.     665 
  15.     666 
  17. /opt/conda/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in __call__(self, *args, **kwds)
  18.     883 
  19.     884       with OptionalXlaContext(self._jit_compile):
  20. --> 885         result = self._call(*args, **kwds)
  21.     886 
  22.     887       new_tracing_count = self.experimental_get_tracing_count()
  24. /opt/conda/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in _call(self, *args, **kwds)
  25.     931       # This is the first call of __call__, so we have to initialize.
  26.     932       initializers = []
  27. --> 933       self._initialize(args, kwds, add_initializers_to=initializers)
  28.     934     finally:
  29.     935       # At this point we know that the initialization is complete (or less
  31. /opt/conda/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in _initialize(self, args, kwds, add_initializers_to)
  32.     758     self._concrete_stateful_fn = (
  33.     759         self._stateful_fn._get_concrete_function_internal_garbage_collected(  # pylint: disable=protected-access
  34. --> 760             *args, **kwds))
  35.     761 
  36.     762     def invalid_creator_scope(*unused_args, **unused_kwds):
  38. /opt/conda/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)
  39.    3064       args, kwargs = None, None
  40.    3065     with self._lock:
  41. -> 3066       graph_function, _ = self._maybe_define_function(args, kwargs)
  42.    3067     return graph_function
  43.    3068 
  45. /opt/conda/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _maybe_define_function(self, args, kwargs)
  46.    3461 
  47.    3462           self._function_cache.missed.add(call_context_key)
  48. -> 3463           graph_function = self._create_graph_function(args, kwargs)
  49.    3464           self._function_cache.primary[cache_key] = graph_function
  50.    3465 
  52. /opt/conda/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)
  53.    3306             arg_names=arg_names,
  54.    3307             override_flat_arg_shapes=override_flat_arg_shapes,
  55. -> 3308             capture_by_value=self._capture_by_value),
  56.    3309         self._function_attributes,
  57.    3310         function_spec=self.function_spec,
  59. /opt/conda/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes, acd_record_initial_resource_uses)
  60.    1005         _, original_func = tf_decorator.unwrap(python_func)
  61.    1006 
  62. -> 1007       func_outputs = python_func(*func_args, **func_kwargs)
  63.    1008 
  64.    1009       # invariant: `func_outputs` contains only Tensors, CompositeTensors,
  66. /opt/conda/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in wrapped_fn(*args, **kwds)
  67.     666         # the function a weak reference to itself to avoid a reference cycle.
  68.     667         with OptionalXlaContext(compile_with_xla):
  69. --> 668           out = weak_wrapped_fn().__wrapped__(*args, **kwds)
  70.     669         return out
  71.     670 
  73. /opt/conda/lib/python3.7/site-packages/tensorflow/python/saved_model/function_deserialization.py in restored_function_body(*args, **kwargs)
  74.     292         .format(_pretty_format_positional(args), kwargs,
  75.     293                 len(saved_function.concrete_functions),
  76. --> 294                 "\n\n".join(signature_descriptions)))
  77.     295 
  78.     296   concrete_function_objects = []
  80. ValueError: Could not find matching function to call loaded from the SavedModel. Got:
  81.   Positional arguments (4 total):
  82.     * Tensor("inputs:0", shape=(64, 113), dtype=int64)
  83.     * False
  84.     * None
  85.     * True
  86.   Keyword arguments: {}
  88. Expected these arguments to match one of the following 4 option(s):
  90. Option 1:
  91.   Positional arguments (4 total):
  92.     * TensorSpec(shape=(None, 113), dtype=tf.int64, name='input_1')
  93.     * False
  94.     * None
  95.     * False
  96.   Keyword arguments: {}
  98. Option 2:
  99.   Positional arguments (4 total):
  100.     * TensorSpec(shape=(None, 113), dtype=tf.int64, name='inputs')
  101.     * False
  102.     * None
  103.     * False
  104.   Keyword arguments: {}
  106. Option 3:
  107.   Positional arguments (4 total):
  108.     * TensorSpec(shape=(None, 113), dtype=tf.int64, name='inputs')
  109.     * True
  110.     * None
  111.     * False
  112.   Keyword arguments: {}
  114. Option 4:
  115.   Positional arguments (4 total):
  116.     * TensorSpec(shape=(None, 113), dtype=tf.int64, name='input_1')
  117.     * True
  118.     * None
  119.     * False
  120.   Keyword arguments: {}

I think this is due to the custom parameter in the call method. Here is a minimal example that reproduces the problem:

  1. import tensorflow as tf
  3. class CustomModel(tf.keras.models.Model):
  4.     def __init__(self):
  5.         super().__init__()
  6.         self.dense = tf.keras.layers.Dense(10)
  7.         
  8.     def call(self, inputs, custom_param=False):
  9.         return self.dense(inputs)
  12. model = CustomModel()
  14. sample_inputs = tf.zeros((16, 30))
  15. print('Sample inputs:', sample_inputs)
  17. sample_outputs = model(sample_inputs)
  18. print('Sample outputs:', sample_outputs)
  20. model.save('saved_model')
  21. loaded_model = tf.keras.models.load_model('saved_model')
  23. sample_outputs_2 = loaded_model(sample_inputs, custom_param=True)
  24. print('Sample outputs 2:', sample_outputs_2)

Calling the reloaded model with custom_param taking any value other than the default always seems to fail.

Is this a bug or by design? How can I modify the model so that it returns just the output sequence while training, but returns the output sequence and state at inference time? This is so I can feed the state back into the model and generate additional characters during inference.

答案1

得分: 0

我找到了解决方法。解决方案在TensorFlow文档中，尽管不太清楚。

使用上述代码，加载的模型类型为keras.saving.saved_model.load.CustomModel，与原始类型不同。要恢复原始类型，需要执行以下操作。

CustomModel类需要get_config和from_config方法。

  1. class CustomModel(tf.keras.models.Model):
  2.     def __init__(self):
  3.         super().__init__()
  4.         self.dense = tf.keras.layers.Dense(10)
  5.         
  6.     def call(self, inputs, custom_param=False):
  7.         return self.dense(inputs)
  9.     def get_config(self):
  10.         return {} # 最初传递给__init__的任何参数都应放在这里。
  12.     @classmethod
  13.     def from_config(cls, config):
  14.         return cls(**config)

加载模型时，需要在custom_objects字典中传递自定义类。

  1. loaded_model = tf.keras.models.load_model('saved_model', custom_objects={'CustomModel': CustomModel})

然后，loaded_model的类型为CustomModel，使用custom_param=True进行调用。

英文:

I figured it out. The solution was in the TensorFlow docs, albeit not very clearly.

With the above code, the loaded model is of type keras.saving.saved_model.load.CustomModel, which is not the same as the original type. To get back the original type, you need to do the following.

The CustomModel class needs the get_config and from_config methods.

  1. class CustomModel(tf.keras.models.Model):
  2.     def __init__(self):
  3.         super().__init__()
  4.         self.dense = tf.keras.layers.Dense(10)
  5.         
  6.     def call(self, inputs, custom_param=False):
  7.         return self.dense(inputs)
  9.     def get_config(self):
  10.         return {} # Any parameters originally passed to __init__ should go here.
  12.     @classmethod
  13.     def from_config(cls, config):
  14.         return cls(**config)

When loading the model, you need to pass the custom class in the custom_objects dictionary.

  1. loaded_model = tf.keras.models.load_model('saved_model', custom_objects={'CustomModel': CustomModel})

Then, loaded_model is of type CustomModel and calling it with custom_param=True works.

huangapple
  • 本文由 发表于 2023年2月14日 06:32:42
  • 转载请务必保留本文链接：https://go.coder-hub.com/75441813.html
  • machine-learning
  • python
  • tensorflow
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定