alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 1 | # Copyright (c) 2021 Arm Limited. All rights reserved. |
| 2 | # SPDX-License-Identifier: Apache-2.0 |
| 3 | # |
| 4 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | # you may not use this file except in compliance with the License. |
| 6 | # You may obtain a copy of the License at |
| 7 | # |
| 8 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | # |
| 10 | # Unless required by applicable law or agreed to in writing, software |
| 11 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | # See the License for the specific language governing permissions and |
| 14 | # limitations under the License. |
| 15 | """ |
| 16 | This script will provide you with a short example of how to perform clustering of weights (weight sharing) in |
| 17 | TensorFlow using the TensorFlow Model Optimization Toolkit. |
| 18 | |
| 19 | The output from this example will be a TensorFlow Lite model file where weights in each layer have been 'clustered' into |
| 20 | 16 clusters during training - quantization has then been applied on top of this. |
| 21 | |
| 22 | By clustering the model we can improve compression of the model file. This can be essential for deploying certain |
| 23 | models on systems with limited resources - such as embedded systems using an Arm Ethos NPU. |
| 24 | |
| 25 | After performing clustering we do post-training quantization to quantize the model and then generate a TensorFlow Lite file. |
| 26 | |
| 27 | If you are targetting an Arm Ethos-U55 NPU then the output TensorFlow Lite file will also need to be passed through the Vela |
| 28 | compiler for further optimizations before it can be used. |
| 29 | |
| 30 | For more information on using Vela see: https://git.mlplatform.org/ml/ethos-u/ethos-u-vela.git/about/ |
| 31 | For more information on clustering see: https://www.tensorflow.org/model_optimization/guide/clustering |
| 32 | """ |
| 33 | import pathlib |
| 34 | |
| 35 | import tensorflow as tf |
| 36 | import tensorflow_model_optimization as tfmot |
| 37 | |
| 38 | from training_utils import get_data, create_model |
| 39 | from post_training_quantization import post_training_quantize, evaluate_tflite_model |
| 40 | |
| 41 | |
| 42 | def prepare_for_clustering(keras_model): |
| 43 | """Prepares a Keras model for clustering.""" |
| 44 | |
| 45 | # Choose the number of clusters to use and how to initialize them. Using more clusters will generally |
| 46 | # reduce accuracy so you will need to find the optimal number for your use-case. |
| 47 | number_of_clusters = 16 |
| 48 | cluster_centroids_init = tfmot.clustering.keras.CentroidInitialization.LINEAR |
| 49 | |
| 50 | # Apply the clustering wrapper to the whole model so weights in every layer will get clustered. You may find that |
| 51 | # to avoid too much accuracy loss only certain non-critical layers in your model should be clustered. |
| 52 | clustering_ready_model = tfmot.clustering.keras.cluster_weights(keras_model, |
| 53 | number_of_clusters=number_of_clusters, |
| 54 | cluster_centroids_init=cluster_centroids_init) |
| 55 | |
| 56 | # We must recompile the model after making it ready for clustering. |
| 57 | clustering_ready_model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), |
| 58 | loss=tf.keras.losses.sparse_categorical_crossentropy, |
| 59 | metrics=['accuracy']) |
| 60 | |
| 61 | return clustering_ready_model |
| 62 | |
| 63 | |
| 64 | def main(): |
| 65 | x_train, y_train, x_test, y_test = get_data() |
| 66 | model = create_model() |
| 67 | |
| 68 | # Compile and train the model first. |
| 69 | # In general it is easier to do clustering as a fine-tuning step after the model is fully trained. |
| 70 | model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), |
| 71 | loss=tf.keras.losses.sparse_categorical_crossentropy, |
| 72 | metrics=['accuracy']) |
| 73 | |
| 74 | model.fit(x=x_train, y=y_train, batch_size=128, epochs=5, verbose=1, shuffle=True) |
| 75 | |
| 76 | # Test the trained model accuracy. |
| 77 | test_loss, test_acc = model.evaluate(x_test, y_test) |
| 78 | print(f"Test accuracy before clustering: {test_acc:.3f}") |
| 79 | |
| 80 | # Prepare the model for clustering. |
| 81 | clustered_model = prepare_for_clustering(model) |
| 82 | |
| 83 | # Continue training the model but now with clustering applied. |
| 84 | clustered_model.fit(x=x_train, y=y_train, batch_size=128, epochs=1, verbose=1, shuffle=True) |
| 85 | test_loss, test_acc = clustered_model.evaluate(x_test, y_test) |
| 86 | print(f"Test accuracy after clustering: {test_acc:.3f}") |
| 87 | |
| 88 | # Remove all variables that clustering only needed in the training phase. |
| 89 | model_for_export = tfmot.clustering.keras.strip_clustering(clustered_model) |
| 90 | |
| 91 | # Apply post-training quantization on top of the clustering and save the resulting TensorFlow Lite model to file. |
| 92 | tflite_model = post_training_quantize(model_for_export, x_train) |
| 93 | |
| 94 | tflite_models_dir = pathlib.Path('./conditioned_models/') |
| 95 | tflite_models_dir.mkdir(exist_ok=True, parents=True) |
| 96 | |
| 97 | clustered_quant_model_save_path = tflite_models_dir / 'clustered_post_training_quant_model.tflite' |
| 98 | with open(clustered_quant_model_save_path, 'wb') as f: |
| 99 | f.write(tflite_model) |
| 100 | |
| 101 | # Test the clustered quantized model accuracy. Save time by only testing a subset of the whole data. |
| 102 | num_test_samples = 1000 |
| 103 | evaluate_tflite_model(clustered_quant_model_save_path, x_test[0:num_test_samples], y_test[0:num_test_samples]) |
| 104 | |
| 105 | |
| 106 | if __name__ == "__main__": |
| 107 | main() |