努力让子类损失函数在 Tensorflow (2.2.0) 中工作。
最初尝试了这段代码（我知道它对其他人有用 - 请参阅https://github.com/keras-team/keras/issues/2115#issuecomment-530762739 https://github.com/keras-team/keras/issues/2115#issuecomment-530762739):
import tensorflow.keras.backend as K
from tensorflow.keras.losses import CategoricalCrossentropy
class WeightedCategoricalCrossentropy(CategoricalCrossentropy):
def __init__(self, cost_mat, name='weighted_categorical_crossentropy', **kwargs):
assert(cost_mat.ndim == 2)
assert(cost_mat.shape[0] == cost_mat.shape[1])
super().__init__(name=name, **kwargs)
self.cost_mat = K.cast_to_floatx(cost_mat)
def __call__(self, y_true, y_pred):
return super().__call__(
y_true=y_true,
y_pred=y_pred,
sample_weight=get_sample_weights(y_true, y_pred, self.cost_mat),
)
def get_sample_weights(y_true, y_pred, cost_m):
num_classes = len(cost_m)
y_pred.shape.assert_has_rank(2)
y_pred.shape[1].assert_is_compatible_with(num_classes)
y_pred.shape.assert_is_compatible_with(y_true.shape)
y_pred = K.one_hot(K.argmax(y_pred), num_classes)
y_true_nk1 = K.expand_dims(y_true, 2)
y_pred_n1k = K.expand_dims(y_pred, 1)
cost_m_1kk = K.expand_dims(cost_m, 0)
sample_weights_nkk = cost_m_1kk * y_true_nk1 * y_pred_n1k
sample_weights_n = K.sum(sample_weights_nkk, axis=[1, 2])
return sample_weights_n
使用方法如下：
model.compile(optimizer='adam',
loss={'simple_Class': 'categorical_crossentropy',
'soundClass': 'binary_crossentropy',
'auxiliary_soundClass':'binary_crossentropy',
'auxiliary_class_training': WeightedCategoricalCrossentropy(cost_matrix),
'class_training':WeightedCategoricalCrossentropy(cost_matrix)
},
loss_weights={'simple_Class': 1.0,
'soundClass': 1.0,
'auxiliary_soundClass':0.7,
'auxiliary_class_training': 0.7,
'class_training':0.4})
(where cost_matrix是一个二维 numpy 数组）。训练槽model.fit() with batch_size=512。
但是，这会导致以下错误：
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-21-3428d6d8967a> in <module>()
82 'class_training': class_lables_test}),
83
---> 84 epochs=nb_epoch, batch_size=batch_size, initial_epoch=initial_epoch, verbose=0, shuffle=True, callbacks=[se, tb, cm, mc, es, rs])
85
86 #model.save(save_version_dir,save_format='tf')
10 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py in _method_wrapper(self, *args, **kwargs)
64 def _method_wrapper(self, *args, **kwargs):
65 if not self._in_multi_worker_mode(): # pylint: disable=protected-access
---> 66 return method(self, *args, **kwargs)
67
68 # Running inside `run_distribute_coordinator` already.
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_batch_size, validation_freq, max_queue_size, workers, use_multiprocessing)
846 batch_size=batch_size):
847 callbacks.on_train_batch_begin(step)
--> 848 tmp_logs = train_function(iterator)
849 # Catch OutOfRangeError for Datasets of unknown size.
850 # This blocks until the batch has finished executing.
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/def_function.py in __call__(self, *args, **kwds)
578 xla_context.Exit()
579 else:
--> 580 result = self._call(*args, **kwds)
581
582 if tracing_count == self._get_tracing_count():
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/def_function.py in _call(self, *args, **kwds)
625 # This is the first call of __call__, so we have to initialize.
626 initializers = []
--> 627 self._initialize(args, kwds, add_initializers_to=initializers)
628 finally:
629 # At this point we know that the initialization is complete (or less
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/def_function.py in _initialize(self, args, kwds, add_initializers_to)
504 self._concrete_stateful_fn = (
505 self._stateful_fn._get_concrete_function_internal_garbage_collected( # pylint: disable=protected-access
--> 506 *args, **kwds))
507
508 def invalid_creator_scope(*unused_args, **unused_kwds):
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)
2444 args, kwargs = None, None
2445 with self._lock:
-> 2446 graph_function, _, _ = self._maybe_define_function(args, kwargs)
2447 return graph_function
2448
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/function.py in _maybe_define_function(self, args, kwargs)
2775
2776 self._function_cache.missed.add(call_context_key)
-> 2777 graph_function = self._create_graph_function(args, kwargs)
2778 self._function_cache.primary[cache_key] = graph_function
2779 return graph_function, args, kwargs
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)
2665 arg_names=arg_names,
2666 override_flat_arg_shapes=override_flat_arg_shapes,
-> 2667 capture_by_value=self._capture_by_value),
2668 self._function_attributes,
2669 # Tell the ConcreteFunction to clean up its graph once it goes out of
/usr/local/lib/python3.6/dist-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)
979 _, original_func = tf_decorator.unwrap(python_func)
980
--> 981 func_outputs = python_func(*func_args, **func_kwargs)
982
983 # invariant: `func_outputs` contains only Tensors, CompositeTensors,
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/def_function.py in wrapped_fn(*args, **kwds)
439 # __wrapped__ allows AutoGraph to swap in a converted function. We give
440 # the function a weak reference to itself to avoid a reference cycle.
--> 441 return weak_wrapped_fn().__wrapped__(*args, **kwds)
442 weak_wrapped_fn = weakref.ref(wrapped_fn)
443
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)
966 except Exception as e: # pylint:disable=broad-except
967 if hasattr(e, "ag_error_metadata"):
--> 968 raise e.ag_error_metadata.to_exception(e)
969 else:
970 raise
TypeError: in user code:
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:571 train_function *
outputs = self.distribute_strategy.run(
/usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:951 run **
return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)
/usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2290 call_for_each_replica
return self._call_for_each_replica(fn, args, kwargs)
/usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2649 _call_for_each_replica
return fn(*args, **kwargs)
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:533 train_step **
y, y_pred, sample_weight, regularization_losses=self.losses)
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/compile_utils.py:205 __call__
loss_value = loss_obj(y_t, y_p, sample_weight=sw)
TypeError: __call__() got an unexpected keyword argument 'sample_weight'
当我更换后这个问题就解决了__call__()魔术方法与call()并手动实现一些底层逻辑。这有效，用法相同。这__call__方法改为：
def call(self, y_true, y_pred):
return super().call(y_true, y_pred) * get_sample_weights(y_true, y_pred, self.cost_mat)
即我们计算分类交叉熵损失y_true and y_pred然后直接乘以我们的权重矩阵，而不是传递y_true, y_pred and self-cost_mat到分类交叉熵call方法并使用继承方法自身的逻辑将损失乘以权重。这不是一个大问题，因为代码确实有效 - 但我不明白为什么我无法使用继承类自己的__call__正确实施（按照原始代码）。
我也改变了y_pred.shape[1].assert_is_compatible_with(num_classes) to assert(y_pred.shape[1] == num_classes)- 这是因为y_pred.shape[1]正在返回一个int。我不知道为什么，因为，检查y_pred，当然，它是一个tf.Tensor， 所以，.shape[1]应该返回一个tf.TesnorShape对象，其上.assert_is_compatible_with()可以被召唤。
这是我成功使用的整个类的实现。
注意 - 它包括 from_config 和 get_config 方法，以及对 Keras 损失命名空间（最后一行）的显式分配，以通过以下方式启用整个模型 + 优化器状态保存model.save(save_format='tf')。其中一些功能很难发挥作用：我必须实现对 NumPy 数组的显式转换（请参阅第一行__init__方法）。
class WeightedCategoricalCrossentropy(tensorflow.keras.losses.CategoricalCrossentropy):
def __init__(self, cost_mat, name='weighted_categorical_crossentropy', **kwargs):
cost_mat = np.array(cost_mat)
## when loading from config, self.cost_mat returns as a list, rather than an numpy array.
## Adding the above line fixes this issue, enabling .ndim to call sucessfully.
## However, this is probably not the best implementation
assert(cost_mat.ndim == 2)
assert(cost_mat.shape[0] == cost_mat.shape[1])
super().__init__(name=name, **kwargs)
self.cost_mat = K.cast_to_floatx(cost_mat)
def call(self, y_true, y_pred):
return super().call(y_true, y_pred) * get_sample_weights(y_true, y_pred, self.cost_mat)
def get_config(self):
config = super().get_config().copy()
# Calling .update on the line above, during assignment, causes an error with config becoming None-type.
config.update({'cost_mat': (self.cost_mat)})
return config
@classmethod
def from_config(cls, config):
# something goes wrong here and changes self.cost_mat to a list variable.
# See above for temporary fix
return cls(**config)
def get_sample_weights(y_true, y_pred, cost_m):
num_classes = len(cost_m)
y_pred.shape.assert_has_rank(2)
assert(y_pred.shape[1] == num_classes)
y_pred.shape.assert_is_compatible_with(y_true.shape)
y_pred = K.one_hot(K.argmax(y_pred), num_classes)
y_true_nk1 = K.expand_dims(y_true, 2)
y_pred_n1k = K.expand_dims(y_pred, 1)
cost_m_1kk = K.expand_dims(cost_m, 0)
sample_weights_nkk = cost_m_1kk * y_true_nk1 * y_pred_n1k
sample_weights_n = K.sum(sample_weights_nkk, axis=[1, 2])
return sample_weights_n
tf.keras.losses.WeightedCategoricalCrossentropy = WeightedCategoricalCrossentropy
最后，保存模型的实现如下：
model.save(save_version_dir,save_format='tf')
并按如下方式加载模型：
model = tf.keras.models.load_model(
save_version_dir,
compile=True,
custom_objects={
'WeightedCategoricalCrossentropy': WeightedCategoricalCrossentropy(cost_matrix)
}
)
