如何用使用估计器实现的官方resnet模型进行预测

Question

我使用以下代码：https://github.com/tensorflow/models/blob/master/official/resnet/imagenet_main.py通过修改类的数量来进行二进制分类。该模型经过培训，没有任何问题，并提供了良好的准确性。

在下一步中，我想恢复经过训练的模型进行预测。我按照TensorFlow的“保存和恢复”教程进行操作。但是，我必须以标准的SavedModel格式（不是自动保存的tf.estimator.Estimator模型）导出我的模型。我在代码中添加了这个serving_input_reciever_fn：

def serving_input_receiver_fn():

    serialized_tf_example = tf.placeholder(dtype=tf.string, shape=[None], name='input_exapmle_tensor')
    receiver_tensors = {"predictor_inputs": serialized_tf_example}

    feature_spec = {"image": tf.FixedLenFeature((), tf.string)}

    features = tf.parse_example(serialized_tf_example, feature_spec, example_names='input')

    return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)

在model_fn我添加了这个来声明输出输出：

predict_output = {
  'pred_output_class': tf.argmax(logits, axis=1),
  'pred_output_prob': tf.nn.softmax(logits, name='softmax_tensor')
}

export_output = {'predict_output': tf.estimator.export.PredictOutput(predict_output)}

if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions, export_outputs=export_output)

在主要功能中，我在列车和验证步骤之后添加了这一行：

  resnet_classifier.export_savedmodel(FLAGS.export_dir, serving_input_receiver_fn)

经过培训和验证，我得到了这个错误：

ValueError: Shape must be rank 1 but is rank 0 for 'ParseExample/ParseExample' (op: 'ParseExample') with input shapes: [?], [], [], [0].

当然，预期的标准模型尚未出口。我想serving_input_receiver_fn的任何东西都是错的。也许输入类型与model_fn的输入类型不匹配。我该如何定义这个功能？

---

更新：我尝试使用'tf.estimator.export.build_raw_serving_input_receiver_fn'为模型提供预处理的原始数据。主要功能中的代码：

feature_spec = {"input_image": tf.placeholder(dtype=tf.string, shape=[None, 224, 224, 3], name='input')}

input_receiver_fn = tf.estimator.export.build_raw_serving_input_receiver_fn(feature_spec)
resnet_classifier.export_savedmodel(export_dir_base=FLAGS.export_dir,serving_input_receiver_fn=input_receiver_fn, as_text=True)

然后我收到了这个错误：

Traceback (most recent call last):
  File "classification_main.py", line 306, in <module>
    tf.app.run(argv=[sys.argv[0]] + unparsed)
  File "/home/ding/.virtualenvs/cv-py2/local/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "classification_main.py", line 301, in main
    resnet_classifier.export_savedmodel(export_dir_base=FLAGS.export_dir,serving_input_receiver_fn=input_receiver_fn, as_text=True)
  File "/home/ding/.virtualenvs/cv-py2/local/lib/python2.7/site-packages/tensorflow/python/estimator/estimator.py", line 511, in export_savedmodel
    config=self.config)
  File "/home/ding/.virtualenvs/cv-py2/local/lib/python2.7/site-packages/tensorflow/python/estimator/estimator.py", line 694, in _call_model_fn
    model_fn_results = self._model_fn(features=features, **kwargs)
  File "classification_main.py", line 184, in resnet_model_fn
    inputs=features, is_training=(mode == tf.estimator.ModeKeys.TRAIN))
  File "/home/ding/projektpraktikum/tensorflow_ws/classification/resnet_model.py", line 249, in model
    inputs = tf.transpose(inputs, [0, 3, 1, 2])
  File "/home/ding/.virtualenvs/cv-py2/local/lib/python2.7/site-packages/tensorflow/python/ops/array_ops.py", line 1336, in transpose
    ret = gen_array_ops.transpose(a, perm, name=name)
  File "/home/ding/.virtualenvs/cv-py2/local/lib/python2.7/site-packages/tensorflow/python/ops/gen_array_ops.py", line 5694, in transpose
    "Transpose", x=x, perm=perm, name=name)
  File "/home/ding/.virtualenvs/cv-py2/local/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.py", line 513, in _apply_op_helper
    raise err
TypeError: Failed to convert object of type <type 'dict'> to Tensor. Contents: {'input_image': <tf.Tensor 'input:0' shape=(?, 224, 224, 3) dtype=string>}. Consider casting elements to a supported type.

Answer 1

在你的第一个例子中，输入形状被设置为[None]，意思是任何向量（serialized_tf_example = tf.placeholder(dtype=tf.string, shape=[None], name='input_exapmle_tensor')，它听起来像你将一个字符串作为标量传递。你可以将输入包装在一个列表中（给你一个单元素向量） ，或将形状更改为[]以获得标量。

在第二个例子中，dtype仍然是string，但听起来你想要float32或其他数字类型直接喂食图像。

Answer 2

我终于得到了答案：

1. 在训练过程中，您需要指定export_dir参数以在模型训练后保存模型。保存的文件夹（graph_pb_path）有两部分（一个是名为variables的文件夹，一个是saved_model.pb）
2. 运行以下代码以恢复模型并进行预测

with tf.Session(graph=tf.Graph()) as sess:
        tf.saved_model.loader.load(sess,["serve"], graph_pb_path)
        graph = tf.get_default_graph()
        inputs = graph.get_tensor_by_name('input_tensor:0')
        model = graph.get_tensor_by_name('resnet_model/final_dense:0')
        res = sess.run(model, {inputs:img})

注意：要测试N个图像，需要在以下文件中更改参数batch_size：tensorflow models export