Итак, вот модель со стандартной функцией потерь.
target = Input(shape=(1, ), dtype='int32')
w_inputs = Input(shape=(1, ), dtype='int32')
w_emb = Embedding(V, dim, embeddings_initializer='glorot_uniform',name='word_emb')(w_inputs)
w_flat= Flatten()(w_emb)
# context
w1= Dense(input_dim=dim, units=V, activation='softmax') # because I want to use predicition on valid set)
w= w1(w_flat)
model = Model(inputs=[w_inputs], outputs=[w])
model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd',metrics=['accuracy'])
Работает нормально.Учитывая, что потери NCE недоступны в кератах, я записал пользовательские потери.
def model_loss(layer,labels, inputs, num_sampled, num_classes, num_true):
weights= K.transpose( layer.get_weights()[0])
biases = layer.get_weights()[1]
def loss(y_true, y_pred):
if K.learning_phase() == 1:
compute_loss = tf.nn.nce_loss(weights, biases, labels, inputs, num_sampled, num_classes, num_true,
partition_strategy="div")
else:
logits = tf.matmul(K.squeeze(inputs,axis=0), K.transpose(weights))
logits = tf.nn.bias_add(logits, biases)
labels_one_hot = tf.one_hot(labels, num_classes)
loss = tf.nn.sigmoid_cross_entropy_with_logits(
labels=labels_one_hot[:][0][:],
logits=logits)
compute_loss = tf.reduce_sum(loss, axis=1)
return compute_loss
return loss
И изменил последнюю строку на:
model.compile(loss=model_loss(w1,target, w_emb, num_sampled, num_classes, num_true), optimizer='sgd',metrics=['accuracy'])
Это компилируется между прочим.
И при исполнении умирает.
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-68-d3b3ef93b81b> in <module>
3 epochs=epochs, steps_per_epoch = seq_len,
4
----> 5 verbose=1, max_queue_size=15)
/opt/conda/lib/python3.6/site-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
/opt/conda/lib/python3.6/site-packages/keras/engine/training.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)
1416 use_multiprocessing=use_multiprocessing,
1417 shuffle=shuffle,
-> 1418 initial_epoch=initial_epoch)
1419
1420 @interfaces.legacy_generator_methods_support
/opt/conda/lib/python3.6/site-packages/keras/engine/training_generator.py in fit_generator(model, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)
38
39 do_validation = bool(validation_data)
---> 40 model._make_train_function()
41 if do_validation:
42 model._make_test_function()
/opt/conda/lib/python3.6/site-packages/keras/engine/training.py in _make_train_function(self)
507 training_updates = self.optimizer.get_updates(
508 params=self._collected_trainable_weights,
--> 509 loss=self.total_loss)
510 updates = (self.updates +
511 training_updates +
/opt/conda/lib/python3.6/site-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
/opt/conda/lib/python3.6/site-packages/keras/optimizers.py in get_updates(self, loss, params)
182 @interfaces.legacy_get_updates_support
183 def get_updates(self, loss, params):
--> 184 grads = self.get_gradients(loss, params)
185 self.updates = [K.update_add(self.iterations, 1)]
186
/opt/conda/lib/python3.6/site-packages/keras/optimizers.py in get_gradients(self, loss, params)
89 grads = K.gradients(loss, params)
90 if None in grads:
---> 91 raise ValueError('An operation has `None` for gradient. '
92 'Please make sure that all of your ops have a '
93 'gradient defined (i.e. are differentiable). '
ValueError: An operation has `None` for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.
Проблема, конечно, в том, что веса не обновляются в слое, следовательно, не градиент.Как я могу сделать это без создания собственного слоя?Я попробовал этот подход, но я отказался от измерения таких вещей, как val_acc, используя слой.