python - 如何将 TensorFlow 模型导出为 .tflite 文件？

Question

背景信息:

我编写了一个与 premade iris classification model 非常相似的 TensorFlow 模型由 TensorFlow 提供。差异相对较小:

• 我正在对足球运动进行分类，而不是鸢尾属物种。
• 我有 10 个特征和一个标签，而不是 4 个特征和一个标签。
• 我有 5 种不同的练习，而不是 3 种鸢尾花。
• 我的 trainData 包含大约 3500 行，而不仅仅是 120 行。
• 我的测试数据包含大约 330 行，而不仅仅是 30 行。
• 我正在使用 n_classes=6 而不是 3 的 DNN 分类器。

我现在想将模型导出为 .tflite 文件。但是根据TensorFlow Developer Guide ，我需要先将模型导出到 tf.GraphDef 文件，然后卡住它，然后才能转换它。然而，tutorial TensorFlow 提供的用于从自定义模型创建 .pb 文件的文件似乎仅针对图像分类模型进行了优化。

问题:

那么如何将像鸢尾花分类示例模型这样的模型转换成 .tflite 文件呢？有没有更简单、更直接的方法，无需将其导出到 .pb 文件，然后卡住它等等？基于鸢尾花分类代码的示例或指向更明确教程的链接将非常有用!

---

其他信息:

• 操作系统:macOS 10.13.4 High Sierra
• TensorFlow 版本:1.8.0
• Python 版本:3.6.4
• 使用 PyCharm 社区 2018.1.3

代码:

可以通过输入以下命令克隆鸢尾花分类代码:

git clone https://github.com/tensorflow/models

但如果你不想下载整个包，这里是:

这是名为 premade_estimator.py 的分类器文件:

    #  Copyright 2016 The TensorFlow Authors. All Rights Reserved.
    #
    #  Licensed under the Apache License, Version 2.0 (the "License");
    #  you may not use this file except in compliance with the License.
    #  You may obtain a copy of the License at
    #
    #  http://www.apache.org/licenses/LICENSE-2.0
    #
    #  Unless required by applicable law or agreed to in writing,                         software
    #  distributed under the License is distributed on an "AS IS" BASIS,
    #  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    #  See the License for the specific language governing permissions and
    #  limitations under the License.
    """An Example of a DNNClassifier for the Iris dataset."""
    from __future__ import absolute_import
    from __future__ import division
    from __future__ import print_function

    import argparse
    import tensorflow as tf

    import iris_data

    parser = argparse.ArgumentParser()
    parser.add_argument('--batch_size', default=100, type=int, help='batch size')
    parser.add_argument('--train_steps', default=1000, type=int,
                help='number of training steps')


    def main(argv):
        args = parser.parse_args(argv[1:])

        # Fetch the data
        (train_x, train_y), (test_x, test_y) = iris_data.load_data()

        # Feature columns describe how to use the input.
        my_feature_columns = []
        for key in train_x.keys():
                    my_feature_columns.append(tf.feature_column.numeric_column(key=key))

        # Build 2 hidden layer DNN with 10, 10 units respectively.
        classifier = tf.estimator.DNNClassifier(
            feature_columns=my_feature_columns,
            # Two hidden layers of 10 nodes each.
            hidden_units=[10, 10],
            # The model must choose between 3 classes.
            n_classes=3)

        # Train the Model.
        classifier.train(
            input_fn=lambda: iris_data.train_input_fn(train_x, train_y,
                                              args.batch_size),
            steps=args.train_steps)

        # Evaluate the model.
        eval_result = classifier.evaluate(
            input_fn=lambda: iris_data.eval_input_fn(test_x, test_y,
                                             args.batch_size))

        print('\nTest set accuracy:         {accuracy:0.3f}\n'.format(**eval_result))

        # Generate predictions from the model
        expected = ['Setosa', 'Versicolor', 'Virginica']
        predict_x = {
            'SepalLength': [5.1, 5.9, 6.9],
            'SepalWidth': [3.3, 3.0, 3.1],
            'PetalLength': [1.7, 4.2, 5.4],
            'PetalWidth': [0.5, 1.5, 2.1],
        }

        predictions = classifier.predict(
            input_fn=lambda: iris_data.eval_input_fn(predict_x,
                                                     labels=None,
                                                     batch_size=args.batch_size))

        template = '\nPrediction is "{}" ({:.1f}%), expected "{}"'

        for pred_dict, expec in zip(predictions, expected):
            class_id = pred_dict['class_ids'][0]
            probability = pred_dict['probabilities'][class_id]

            print(template.format(iris_data.SPECIES[class_id],
                          100 * probability, expec))


    if __name__ == '__main__':
        # tf.logging.set_verbosity(tf.logging.INFO)
        tf.app.run(main)

这是名为 iris_data.py 的数据文件:

    import pandas as pd
    import tensorflow as tf

    TRAIN_URL = "http://download.tensorflow.org/data/iris_training.csv"
    TEST_URL = "http://download.tensorflow.org/data/iris_test.csv"

    CSV_COLUMN_NAMES = ['SepalLength', 'SepalWidth',
                        'PetalLength', 'PetalWidth', 'Species']
    SPECIES = ['Setosa', 'Versicolor', 'Virginica']


    def maybe_download():
        train_path = tf.keras.utils.get_file(TRAIN_URL.split('/')[-1], TRAIN_URL)
        test_path = tf.keras.utils.get_file(TEST_URL.split('/')[-1], TEST_URL)

        return train_path, test_path


    def load_data(y_name='Species'):
        """Returns the iris dataset as (train_x, train_y), (test_x, test_y)."""
        train_path, test_path = maybe_download()

        train = pd.read_csv(train_path, names=CSV_COLUMN_NAMES, header=0)
        train_x, train_y = train, train.pop(y_name)

        test = pd.read_csv(test_path, names=CSV_COLUMN_NAMES, header=0)
        test_x, test_y = test, test.pop(y_name)

        return (train_x, train_y), (test_x, test_y)


    def train_input_fn(features, labels, batch_size):
        """An input function for training"""
        # Convert the inputs to a Dataset.
        dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))

        # Shuffle, repeat, and batch the examples.
        dataset = dataset.shuffle(1000).repeat().batch(batch_size)

        # Return the dataset.
        return dataset


    def eval_input_fn(features, labels, batch_size):
        """An input function for evaluation or prediction"""
        features = dict(features)
        if labels is None:
            # No labels, use only features.
            inputs = features
        else:
            inputs = (features, labels)

        # Convert the inputs to a Dataset.
        dataset = tf.data.Dataset.from_tensor_slices(inputs)

        # Batch the examples
        assert batch_size is not None, "batch_size must not be None"
        dataset = dataset.batch(batch_size)

        # Return the dataset.
        return dataset

** 更新 **

好的，所以我找到了一段看似非常有用的代码 on this page :

    import tensorflow as tf

    img = tf.placeholder(name="img", dtype=tf.float32, shape=(1, 64, 64, 3))
    val = img + tf.constant([1., 2., 3.]) + tf.constant([1., 4., 4.])
    out = tf.identity(val, name="out")
    with tf.Session() as sess:
      tflite_model = tf.contrib.lite.toco_convert(sess.graph_def, [img], [out])
      open("test.tflite", "wb").write(tflite_model)

这个小家伙直接将简单模型转换为 TensorFlow Lite 模型。现在我所要做的就是找到一种方法使其适应虹膜分类模型。有什么建议吗？

Answer 1

Is there an easier, more direct way to do it, without having to export it to a .pb file, then freeze it and so on?

是的，正如您在更新的问题中指出的那样，可以 freeze the graph并使用 toco_convert直接在python api中。它需要卡住图形并确定输入和输出形状。在您的问题中，没有卡住图步骤，因为没有变量。如果你有变量并运行 toco 而没有先将它们转换为常量，toco 会报错!

Now all I have to do is find a way to adapt this to the iris classification model. Any suggestions?

这个有点棘手，需要做更多的工作。基本上，您需要加载图形并找出输入和输出张量名称，然后卡住图形并调用 toco_convert。在这种情况下(您尚未定义图形)要查找输入和输出张量名称，您必须查看生成的图形并根据输入形状、名称等确定它们。这是您可以附加的代码premade_estimator.py 中的主要函数结束，在这种情况下生成 tflite 图。

print("\n====== classifier model_dir, latest_checkpoint ===========")
print(classifier.model_dir)
print(classifier.latest_checkpoint())
debug = False

with tf.Session() as sess:
    # First let's load meta graph and restore weights
    latest_checkpoint_path = classifier.latest_checkpoint()
    saver = tf.train.import_meta_graph(latest_checkpoint_path + '.meta')
    saver.restore(sess, latest_checkpoint_path)

    # Get the input and output tensors needed for toco.
    # These were determined based on the debugging info printed / saved below.
    input_tensor = sess.graph.get_tensor_by_name("dnn/input_from_feature_columns/input_layer/concat:0")
    input_tensor.set_shape([1, 4])
    out_tensor = sess.graph.get_tensor_by_name("dnn/logits/BiasAdd:0")
    out_tensor.set_shape([1, 3])

    # Pass the output node name we are interested in.
    # Based on the debugging info printed / saved below, pulled out the
    # name of the node for the logits (before the softmax is applied).
    frozen_graph_def = tf.graph_util.convert_variables_to_constants(
        sess, sess.graph_def, output_node_names=["dnn/logits/BiasAdd"])

    if debug is True:
        print("\nORIGINAL GRAPH DEF Ops ===========================================")
        ops = sess.graph.get_operations()
        for op in ops:
            if "BiasAdd" in op.name or "input_layer" in op.name:
                print([op.name, op.values()])
        # save original graphdef to text file
        with open("estimator_graph.pbtxt", "w") as fp:
            fp.write(str(sess.graph_def))

        print("\nFROZEN GRAPH DEF Nodes ===========================================")
        for node in frozen_graph_def.node:
            print(node.name)
        # save frozen graph def to text file
        with open("estimator_frozen_graph.pbtxt", "w") as fp:
            fp.write(str(frozen_graph_def))

tflite_model = tf.contrib.lite.toco_convert(frozen_graph_def, [input_tensor], [out_tensor])
open("estimator_model.tflite", "wb").write(tflite_model)

注意:我假设最后一层(在应用 Softmax 之前)的 logits 作为输出，对应于节点 dnn/logits/BiasAdd。如果您想要概率，我相信它是 dnn/head/predictions/probabilities。