INT8 LSTM example¶

This is an example of an 8-bit integer (INT8) quantized TensorFlow Keras model using post-training quantization. In other words, this model can be trained using normal floating-point training, but will be able to run in INT8 mode at inference time. Using post-training quantization requires example representative data to be available. For more information, see the TensorFlow documentation about this subject.

First we define our model and a representative data-set as follows:

import tensorflow as tf

NUM_SAMPLES = 10  # e.g. 10 letters in a word or 10 timestamps
SAMPLE_SIZE = 64  # size of a single sample, e.g. an embedding of size 64

def dataset_example(num_samples: int = 100):
    """Placeholder for a representative data-set. For best quantization
    performance, replace this with a few examples from your own data-set, the
    more the better. This should include any pre-processing needed."""
    for _ in range(num_samples):
        shape = (1, NUM_SAMPLES, SAMPLE_SIZE)
        yield [tf.random.uniform(shape, minval=-1, maxval=1)]

def model() -> tf.keras.Model:
    """Example of a simple single-directional LSTM. Embedding layers are not
    supported and will have to be added as pre-processing steps manually. See the
    official TFLite documentation for more information and constraints:
    https://www.tensorflow.org/lite/models/convert/rnn. See in-line comments below
    for recommendations about tf.keras.layers.LSTM()."""
    in_layer = tf.keras.layers.Input((NUM_SAMPLES, SAMPLE_SIZE), batch_size=1)
    x = in_layer
    # Recommended settings for tf.keras.layers.LSTM:
    # 1) Set unroll=False (the default) for much better speed, RAM, and ROM
    # 2) Set time_major=False (the default) for better speed, RAM, and ROM
    # 3) Set activation='tanh' (the default) for compatibility
    # 4) Set recurrent_activation='sigmoid' (the default) for compatibility
    x = tf.keras.layers.LSTM(units=16, return_sequences=True)(x)
    x = tf.keras.layers.Flatten()(x)
    x = tf.keras.layers.Dense(8)(x)
    return tf.keras.Model(in_layer, x, name="LSTM")

After we have trained the above LSTM example model (not shown in this example), we can convert it to INT8 quantized TFLite format as follows:

from pathlib import Path
from typing import Callable
import tensorflow as tf

def convert_model(model: tf.keras.Model, dataset_gen: Callable) -> bytes:
    converter = tf.lite.TFLiteConverter.from_keras_model(model)
    converter.experimental_new_converter = True
    converter.optimizations = [tf.lite.Optimize.DEFAULT]
    converter.representative_dataset = dataset_gen
    converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
    converter.inference_input_type = tf.int8
    converter.inference_output_type = tf.int8
    return converter.convert()

LSTM_model = model()
# (now train the model or load some weights)
int8_LSTM_model = convert_model(LSTM_model, dataset_example)
Path("LSTM.tflite").write_bytes(int8_LSTM_model)

The resulting LSTM.tflite is now ready to be used with the inference engine.