1 files changed, 139 insertions, 0 deletions
diff --git a/model_conditioning_examples/post_training_quantization.py b/model_conditioning_examples/post_training_quantization.py
new file mode 100644
index 0000000..ab535ac
--- /dev/null
+++ b/model_conditioning_examples/post_training_quantization.py
@@ -0,0 +1,139 @@
+#  Copyright (c) 2021 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  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.
+"""
+This script will provide you with an example of how to perform post-training quantization in TensorFlow.
+
+The output from this example will be a TensorFlow Lite model file where weights and activations are quantized to 8bit
+integer values.
+
+Quantization helps reduce the size of your models and is necessary for running models on certain hardware such as Arm
+Ethos NPU.
+
+In addition to quantizing weights, post-training quantization uses a calibration dataset to
+capture the minimum and maximum values of all variable tensors in your model.
+By capturing these ranges it is possible to fully quantize not just the weights of the model but also the activations.
+
+Depending on the model you are quantizing there may be some accuracy loss, but for a lot of models the loss should
+be minimal.
+
+If you are targetting an Arm Ethos-U55 NPU then the output TensorFlow Lite file will also need to be passed through the Vela
+compiler for further optimizations before it can be used.
+
+For more information on using Vela see: https://git.mlplatform.org/ml/ethos-u/ethos-u-vela.git/about/
+For more information on post-training quantization
+see: https://www.tensorflow.org/lite/performance/post_training_integer_quant
+"""
+import pathlib
+
+import numpy as np
+import tensorflow as tf
+
+from training_utils import get_data, create_model
+
+
+def post_training_quantize(keras_model, sample_data):
+    """Quantize Keras model using post-training quantization with some sample data.
+
+    TensorFlow Lite will have fp32 inputs/outputs and the model will handle quantizing/dequantizing.
+
+    Args:
+        keras_model: Keras model to quantize.
+        sample_data: A numpy array of data to use as a representative dataset.
+
+    Returns:
+        Quantized TensorFlow Lite model.
+    """
+
+    converter = tf.lite.TFLiteConverter.from_keras_model(keras_model)
+
+    # We set the following converter options to ensure our model is fully quantized.
+    # An error should get thrown if there is any ops that can't be quantized.
+    converter.optimizations = [tf.lite.Optimize.DEFAULT]
+    converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+
+    # To use post training quantization we must provide some sample data that will be used to
+    # calculate activation ranges for quantization. This data should be representative of the data
+    # we expect to feed the model and must be provided by a generator function.
+    def generate_repr_dataset():
+        for i in range(100):  # 100 samples is all we should need in this example.
+            yield [np.expand_dims(sample_data[i], axis=0)]
+
+    converter.representative_dataset = generate_repr_dataset
+    tflite_model = converter.convert()
+
+    return tflite_model
+
+
+def evaluate_tflite_model(tflite_save_path, x_test, y_test):
+    """Calculate the accuracy of a TensorFlow Lite model using TensorFlow Lite interpreter.
+
+    Args:
+        tflite_save_path: Path to TensorFlow Lite model to test.
+        x_test: numpy array of testing data.
+        y_test: numpy array of testing labels (sparse categorical).
+    """
+
+    interpreter = tf.lite.Interpreter(model_path=str(tflite_save_path))
+
+    interpreter.allocate_tensors()
+    input_details = interpreter.get_input_details()
+    output_details = interpreter.get_output_details()
+
+    accuracy_count = 0
+    num_test_images = len(y_test)
+
+    for i in range(num_test_images):
+        interpreter.set_tensor(input_details[0]['index'], x_test[i][np.newaxis, ...])
+        interpreter.invoke()
+        output_data = interpreter.get_tensor(output_details[0]['index'])
+
+        if np.argmax(output_data) == y_test[i]:
+            accuracy_count += 1
+
+    print(f"Test accuracy quantized: {accuracy_count / num_test_images:.3f}")
+
+
+def main():
+    x_train, y_train, x_test, y_test = get_data()
+    model = create_model()
+
+    # Compile and train the model in fp32 as normal.
+    model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
+                  loss=tf.keras.losses.sparse_categorical_crossentropy,
+                  metrics=['accuracy'])
+
+    model.fit(x=x_train, y=y_train, batch_size=128, epochs=5, verbose=1, shuffle=True)
+
+    # Test the fp32 model accuracy.
+    test_loss, test_acc = model.evaluate(x_test, y_test)
+    print(f"Test accuracy float: {test_acc:.3f}")
+
+    # Quantize and export the resulting TensorFlow Lite model to file.
+    tflite_model = post_training_quantize(model, x_train)
+
+    tflite_models_dir = pathlib.Path('./conditioned_models/')
+    tflite_models_dir.mkdir(exist_ok=True, parents=True)
+
+    quant_model_save_path = tflite_models_dir / 'post_training_quant_model.tflite'
+    with open(quant_model_save_path, 'wb') as f:
+        f.write(tflite_model)
+
+    # Test the quantized model accuracy. Save time by only testing a subset of the whole data.
+    num_test_samples = 1000
+    evaluate_tflite_model(quant_model_save_path, x_test[0:num_test_samples], y_test[0:num_test_samples])
+
+
+if __name__ == "__main__":
+    main()