Блог

Я внедряю обычай loss для своей модели keras:

 def loss(self, y_true, y_pred):
    self.y_true  = tf.convert_to_tensor(y_true) 
    self.y_pred  = tf.convert_to_tensor(y_pred)
    tf.print("====================")
    tf.print(y_pred)
    tf.print(y_true)
    tf.print("====================")
    s = K.mean(((self.y_true - K.log(self.y_pred))   ((1-self.y_true) - K.log(1-self.y_pred))), axis = -1)
    return s
 

но это приводит к nan :

 [[-nan]
 [-nan]
 [-nan]
 ...
 [-nan]
 [-nan]
 [-nan]]
[[0.132541135]
 [0.102164917]
 [0.0687385723]
 ...
 [0.0518041]
 [0.127239496]
 [0.0189646874]]
 

итак, я решил изменить an epsilon на log часть в функции потерь:

 s = K.mean(((self.y_true - K.log(K.max(self.y_pred, [[0.00000001]])))   ((1-self.y_true) - K.log(K.max(1-self.y_pred, [[0.00000001]])))), axis = -1)
 

затем я столкнулся с этой ошибкой:

 File "/home/m.shah/projects/projects/models/_regression.py", line 52, in train
    result = self.model.fit(train_x, 
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py", line 1100, i
n fit
    tmp_logs = self.train_function(iterator)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 828, in __
call__
    result = self._call(*args, **kwds)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 871, in _c
all
    self._initialize(args, kwds, add_initializers_to=initializers)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 725, in _i
nitialize
    self._stateful_fn._get_concrete_function_internal_garbage_collected(  # pylint: disable=protected-access
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2969, in _get_
concrete_function_internal_garbage_collected
    graph_function, _ = self._maybe_define_function(args, kwargs)
File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2969, in _get_
concrete_function_internal_garbage_collected
    graph_function, _ = self._maybe_define_function(args, kwargs)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 3361, in _maybe_define_function
    graph_function = self._create_graph_function(args, kwargs)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 3196, in _create_graph_function
    func_graph_module.func_graph_from_py_func(
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/framework/func_graph.py", line 990, in func_graph_from_py_func
    func_outputs = python_func(*func_args, **func_kwargs)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 634, in wrapped_fn
    out = weak_wrapped_fn().__wrapped__(*args, **kwds)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/framework/func_graph.py", line 977, in wrapper
    raise e.ag_error_metadata.to_exception(e)
TypeError: in user code:

    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py:805 train_function  *
return step_function(self, iterator)
    /home/m.shah/projects/projects/modules/_loss.py:14 loss  *
        s = K.mean((self.y_true - K.log(K.max(self.y_pred, 0.0000001)))   ((1-self.y_true) - K.log(K.max(1-self.y_pred, 0.0000001))), axis = -1)
    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/util/dispatch.py:201 wrapper  **
        return target(*args, **kwargs)
    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/keras/backend.py:2179 max
        return math_ops.reduce_max(x, axis, keepdims)
    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/util/dispatch.py:201 wrapper
        return target(*args, **kwargs)
    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/ops/math_ops.py:2745 reduce_max
        return reduce_max_with_dims(input_tensor, axis, keepdims, name,
    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/ops/math_ops.py:2757 reduce_max_with_dims
        gen_math_ops._max(input_tensor, dims, keepdims, name=name))
    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/ops/gen_math_ops.py:5626 _max
        _, _, _op, _outputs = _op_def_library._apply_op_helper(
    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/framework/op_def_library.py:628 _apply_op_helper
        _SatisfiesTypeConstraint(base_type,
    /home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/framework/op_def_library.py:59 _SatisfiesTypeConstraint
        raise TypeError(

    TypeError: Value passed to parameter 'reduction_indices' has DataType float32 not in list of allowed values: int32, int64
 

Затем я изменил код на:

 s = K.mean(((self.y_true - K.log(K.max([self.y_pred, [[00000001]]])))   ((1-self.y_true) - K.log(K.max([1-self.y_pred, [[00000001]]])))), axis = -1)
 

and again faced this error:

 Traceback (most recent call last):
  File "experiment5.py", line 80, in <module>
    history = model.train(X_train, y_train, epochs=100)
  File "/home/m.shah/projects/projects/models/_regression.py", line 52, in train
    result = self.model.fit(train_x,
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py", line 1100, in fit
    tmp_logs = self.train_function(iterator)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 828, in __call__
    result = self._call(*args, **kwds)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 888, in _call
    return self._stateless_fn(*args, **kwds)
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2942, in __call__
    return graph_function._call_flat(
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 1918, in _call_flat
    return self._build_call_outputs(self._inference_function.call(
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 555, in call
    outputs = execute.execute(
  File "/home/m.shah/.conda/envs/tensorflow/lib/python3.8/site-packages/tensorflow/python/eager/execute.py", line 59, in quick_execute
    tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
tensorflow.python.framework.errors_impl.InvalidArgumentError:  Shapes of all inputs must match: values[0].shape = [64,1] != values[1].shape = [1,1]
         [[node loss/Max/input (defined at /home/m.shah/projects/projects/modules/_loss.py:14) ]] [Op:__inference_train_function_3127]

Errors may have originated from an input operation.
Input Source operations connected to node loss/Max/input:
 sequential/dense/Relu (defined at /home/m.shah/projects/projects/models/_regression.py:52)

Function call stack:
train_function
 

краткое описание регрессионной модели выглядит следующим образом:

 Layer (type)                 Output Shape              Param #
=================================================================
lstm (LSTM)                  (None, 64)                17152
_________________________________________________________________
dropout (Dropout)            (None, 64)                0
_________________________________________________________________
dense (Dense)                (None, 1)                 65
=================================================================
Total params: 17,217
Trainable params: 17,217
Non-trainable params: 0
_________________________________________________________________
Model Summary:
 

и я добавил tf.print(y_pred.shape) и tf.print(y_true.shape) как журналы, а затем:

 TensorShape([None, 1])
TensorShape([None, 1])
 

Тем временем я изменил функцию на:

 def loss(self, y_true, y_pred):
    config = ConfigProto()
    config.gpu_options.allow_growth = True
    session = InteractiveSession(config=config)
    self.y_true  = tf.convert_to_tensor(y_true) 
    self.y_pred  = tf.convert_to_tensor(y_pred)
    tf.print("=========== shapes =========")
    tf.print(y_pred.shape)
    tf.print(y_true.shape)
    tf.print("====================")
    s = K.mean(((self.y_true - K.log(self.y_pred))   ((1-self.y_true) - K.log(1-self.y_pred))), axis = -1)
 

но опять же это привело к nan тому , что, однако, в нем нет ошибки. Похоже, у меня возникли проблемы с K.max

У кого-нибудь есть какие-нибудь идеи?