How do I export a TensorFlow model as a .tflite fi

Question

Background information:

I have written a TensorFlow model very similar to the premade iris classification model provided by TensorFlow. The differences are relatively minor:

• I am classifying football exercises, not iris species.
• I have 10 features and one label, not 4 features and one label.
• I have 5 different exercises, as opposed to 3 iris species.
• My trainData contains around 3500 rows, not only 120.
• My testData contains around 330 rows, not only 30.
• I am using a DNN classifier with n_classes=6, not 3.

I now want to export the model as a .tflite file. But according to the TensorFlow Developer Guide, I need to first export the model to a tf.GraphDef file, then freeze it and only then will I be able to convert it. However, the tutorial provided by TensorFlow to create a .pb file from a custom model only seems to be optimized for image classification models.

Question:

So how do I convert a model like the iris classification example model into a .tflite file? 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? An example based on the iris classification code or a link to a more explicit tutorial would be very useful!

---

Other information:

• OS: macOS 10.13.4 High Sierra
• TensorFlow Version: 1.8.0
• Python Version: 3.6.4
• Using PyCharm Community 2018.1.3

Code:

The iris classification code can be cloned by entering the following command:

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

But in case you don't want to download the whole package, here it is:

This is the classifier file called premade_estimator.py:

    #  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)

And this is the data file called iris_data.py:

    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

** UPDATE **

Ok so I have found a seemingly very useful piece of code 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)

This little guy directly converts a simple model to a TensorFlow Lite Model. Now all I have to do is find a way to adapt this to the iris classification model. Any suggestions?

Answer 1

Here is a more standard way of doing this instead of using toco_convert. Thanks to Pannag Sanketi for the toco based example above which is the basis of this code.

Note that the output layer is logits because we are using a classification NN. If we had a regression NN, it will be different. The classifier is the NN model that you built.

    def export_tflite(classifier):
        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
            input_tensor = sess.graph.get_tensor_by_name("dnn/input_from_feature_columns/input_layer/concat:0")
            out_tensor = sess.graph.get_tensor_by_name("dnn/logits/BiasAdd:0")

            # here the code differs from the toco example above
            sess.run(tf.global_variables_initializer())
            converter = tf.lite.TFLiteConverter.from_session(sess, [input_tensor], [out_tensor])
            tflite_model = converter.convert()
            open("converted_model.tflite", "wb").write(tflite_model)

Answer 2

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?

Yes, as you pointed out in the updated question, it is possible to freeze the graph and use toco_convert in python api directly. It needs the graph to be frozen and the input and output shapes to be determined. In your question, there is no freeze graph step since there are no variables. If you have variables and run toco without converting those to constants first, toco will complain!

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

This one is slightly trickier and needs more work. Basically, you need to load the graph and figure out the input and output tensor names and then freeze the graph and call toco_convert. For finding the input and output tensor names in this case (where you have not defined the graph), you have to poke around the graph generated and determine them based on input shapes, names, etc. Here is the code that you can append at the end of your main function in premade_estimator.py to generate the tflite graph in this case.

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)

Note: I am assuming the logits from the final layer (before the Softmax is applied) as the output, corresponding to the node dnn/logits/BiasAdd. If you want the probabilities, I believe it is dnn/head/predictions/probabilities.