Question

我正在使用带有Theano后端的Keras 2.1.3。我有一个名为model的Keras模型，我试图将其保存为ML Engine可接受的格式。谷歌搜索后，我发现了这一点：

from keras.models import Model

import tensorflow as tf
from tensorflow.python.saved_model import builder, tag_constants, signature_constants
from tensorflow.python.saved_model.signature_def_utils_impl import predict_signature_def

tf.keras.backend.clear_session()
sess = tf.Session()
tf.keras.backend.set_session(sess)

# build a copy but without the learning nodes
tf.keras.backend.set_learning_phase(0)  # disables creation of dropout
new = Model.from_config(model.get_config())
new.set_weights(model.get_weights())

# export saved model
model_builder = builder.SavedModelBuilder(export_path)

signature = predict_signature_def(
    inputs={'sequence': new.input},
    outputs={'outputs': new.output})
with tf.keras.backend.get_session() as sess:
    sdefm = {signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature}
    model_builder.add_meta_graph_and_variables(sess=sess,
                                               tags=[tag_constants.SERVING],
                                               signature_def_map=sdefm)
    model_builder.save()

但是我得到了一个奇怪的回溯：

  File "/Users/...
    outputs={'outputs': new.output})
  File "/Users/.../tensorflow/python/saved_model/signature_def_utils_impl.py", line 177, in predict_signature_def
    for key, tensor in inputs.items()}
  File "/Users/.../tensorflow/python/saved_model/signature_def_utils_impl.py", line 177, in <dictcomp>
    for key, tensor in inputs.items()}
  File "/Users/.../tensorflow/python/saved_model/utils_impl.py", line 45, in build_tensor_info
    tensor_shape=tensor.get_shape().as_proto())
AttributeError: 'TensorVariable' object has no attribute 'get_shape'
Exception TypeError: TypeError("'NoneType' object is not callable",) in <bound method Session.__del__ of <tensorflow.python.client.session.Session object at 0x12ff1f150>> ignored

这看起来像我的tensor类型的TensorVariable叫做get_shape()，并且没有。

我做错了什么？我觉得这是一个非常常见的用例，所以不应该太疯狂了吗？

谢谢！

Keras：ML引擎的错误导出模型

0 个答案: