我不知道如何在转换为hub.text_embedding_column的Keras模型中使用Tensorflow Hub嵌入列（tf.Estimator）。
如果不将模型转换为估计量，则可以在Keras模型中使用嵌入。
例如，将一些伪数据定义为：
x_train = ['the quick brown fox', 'jumps over a lazy']
x_eval = ['the quick brown fox', 'jumps over a lazy']
y_train = [0, 1]
y_eval = [0, 1]
然后，我可以使用以下代码来正确无误地训练keras模型
embed = hub.Module('https://tfhub.dev/google/nnlm-en-dim128/1')
def _embed(x):
return embed(tf.squeeze(tf.cast(x, tf.string)))
# workaround for keras
x_train = np.array(x_train, dtype=object)[:, np.newaxis]
x_eval = np.array(x_eval, dtype=object)[:, np.newaxis]
input_text = tf.keras.layers.Input(shape=(1,), dtype=tf.string)
embedding = tf.keras.layers.Lambda(_embed, output_shape=(128,))(input_text)
pred = tf.keras.layers.Dense(1, activation='sigmoid')(dense)
model = tf.keras.Model(inputs=input_text, outputs=pred)
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
with tf.Session() as sess:
sess.run([tf.global_variables_initializer(), tf.tables_initializer()])
model.fit(x_train, y_train, epochs=1, validation_data=(x_eval, y_eval))
但是，如果我尝试使用tf.keras.estimator.model_to_estimator将其转换为估计量，突然之间我将无法再训练模型。
embedding = hub.text_embedding_column('text', 'https://tfhub.dev/google/nnlm-en-dim128/1')
features = {'text': x_train}
labels = y_train[:, np.newaxis]
input_fn = tf.estimator.inputs.numpy_input_fn(features, labels, shuffle=False)
embedding_input = tf.keras.layers.Input(shape=(128,), dtype=tf.float32, name='text')
logits = tf.keras.layers.Dense(1, activation='softmax', name='logits')(embedding_input)
model = tf.keras.Model(inputs=embedding_input, outputs=logits)
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
estimator = tf.keras.estimator.model_to_estimator(model)
estimator.train(input_fn, max_steps=1)
如果我使用tf.estimator.DNNEstimator之类的固定估计量，那么我也可以训练模型而不会出错。
embedding = hub.text_embedding_column('text', 'https://tfhub.dev/google/nnlm-en-dim128/1')
features = {'text': x_train}
labels = y_train[:, np.newaxis]
input_fn = tf.estimator.inputs.numpy_input_fn(features, labels, shuffle=False)
estimator = tf.estimator.DNNClassifier([32], [embedding])
当我尝试将keras模型转换为estimator进行训练时，出现的错误是：
Input 0 of layer logits is incompatible with the layer: : expected min_ndim=2, found ndim=1. Full shape received: [None]
完整的堆栈跟踪如下：
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-15-f1d8a31726e2> in <module>()
22 estimator = tf.keras.estimator.model_to_estimator(model)
23
---> 24 estimator.train(input_fn, max_steps=1)
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/estimator/estimator.pyc in train(self, input_fn, hooks, steps, max_steps, saving_listeners)
374
375 saving_listeners = _check_listeners_type(saving_listeners)
--> 376 loss = self._train_model(input_fn, hooks, saving_listeners)
377 logging.info('Loss for final step: %s.', loss)
378 return self
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/estimator/estimator.pyc in _train_model(self, input_fn, hooks, saving_listeners)
1143 return self._train_model_distributed(input_fn, hooks, saving_listeners)
1144 else:
-> 1145 return self._train_model_default(input_fn, hooks, saving_listeners)
1146
1147 def _train_model_default(self, input_fn, hooks, saving_listeners):
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/estimator/estimator.pyc in _train_model_default(self, input_fn, hooks, saving_listeners)
1168 worker_hooks.extend(input_hooks)
1169 estimator_spec = self._call_model_fn(
-> 1170 features, labels, model_fn_lib.ModeKeys.TRAIN, self.config)
1171 return self._train_with_estimator_spec(estimator_spec, worker_hooks,
1172 hooks, global_step_tensor,
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/estimator/estimator.pyc in _call_model_fn(self, features, labels, mode, config)
1131
1132 logging.info('Calling model_fn.')
-> 1133 model_fn_results = self._model_fn(features=features, **kwargs)
1134 logging.info('Done calling model_fn.')
1135
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/estimator/keras.pyc in model_fn(features, labels, mode)
357 """model_fn for keras Estimator."""
358 model = _clone_and_build_model(mode, keras_model, custom_objects, features,
--> 359 labels)
360 model_output_names = []
361 # We need to make sure that the output names of the last layer in the model
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/estimator/keras.pyc in _clone_and_build_model(mode, keras_model, custom_objects, features, labels)
313 model = models.clone_model(keras_model, input_tensors=input_tensors)
314 else:
--> 315 model = models.clone_model(keras_model, input_tensors=input_tensors)
316 else:
317 model = keras_model
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/keras/models.pyc in clone_model(model, input_tensors)
261 return _clone_sequential_model(model, input_tensors=input_tensors)
262 else:
--> 263 return _clone_functional_model(model, input_tensors=input_tensors)
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/keras/models.pyc in _clone_functional_model(model, input_tensors)
154 kwargs['mask'] = computed_mask
155 output_tensors = generic_utils.to_list(layer(computed_tensor,
--> 156 **kwargs))
157 output_masks = generic_utils.to_list(
158 layer.compute_mask(computed_tensor, computed_mask))
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/keras/engine/base_layer.pyc in __call__(self, inputs, *args, **kwargs)
718
719 # Check input assumptions set before layer building, e.g. input rank.
--> 720 self._assert_input_compatibility(inputs)
721 if input_list and self._dtype is None:
722 try:
.../anaconda2/lib/python2.7/site-packages/tensorflow/python/keras/engine/base_layer.pyc in _assert_input_compatibility(self, inputs)
1438 ', found ndim=' + str(ndim) +
1439 '. Full shape received: ' +
-> 1440 str(x.shape.as_list()))
1441 # Check dtype.
1442 if spec.dtype is not None:
ValueError: Input 0 of layer logits is incompatible with the layer: : expected min_ndim=2, found ndim=1. Full shape received: [None]
答案 0 :(得分：0)
我终于设法弄清楚如何将model_to_estimator与TFHub一起使用。您需要在Keras模型外部进行嵌入。您的Keras模型必须将嵌入作为输入，而不是在模型中处理嵌入。但是，您可以将Keras模型用作估算器函数中的函数。
例如，您可以定义一个接受预先计算的嵌入的Keras模型（对于此示例，我想让嵌入返回一个序列而不是单个平均嵌入，因此输入形状具有序列长度）：
import tensorflow as tf
import tensorflow_hub as hub
import numpy as np
import shutil
def create_model(max_seq_len, embedding_size):
model = tf.keras.Sequential()
model.add(tf.keras.layers.Dropout(0.5, input_shape=(max_seq_len, embedding_size)))
model.add(tf.keras.layers.SeparableConv1D(8, 3, padding='same', activation=tf.nn.leaky_relu))
model.add(tf.keras.layers.GlobalAveragePooling1D())
model.add(tf.keras.layers.Dense(2, activation='softmax'))
return model
您将定义一个估计器模型函数，例如：
，而不是先编译此模型，然后再使用model_to_estimator
def model_fn(features, labels, mode, params):
model = create_model(5, 128)
embed = hub.Module(...)
text_seq = pad_seq(features['text'], 5)
embeddings = tf.map_fn(embed, text_seq)
if mode == tf.estimator.ModeKeys.TRAIN:
logits = model(embeddings, training=True)
# some more logic
像这样调用Keras模型，即可从模型中计算出logit。然后，您可以返回tf.estimator.EstimatorSpec创建一个Estimatorm abd，然后从那里进行训练。
您可以参考the Tensorflow MNIST example来了解它们如何将Tensorflow计算围绕Keras模型包装，以创建估算器模型函数，然后创建估算器，即使他们没有使用TFHub的任何东西。