Skip to content

Latest commit

 

History

History
178 lines (142 loc) · 10.3 KB

README.md

File metadata and controls

178 lines (142 loc) · 10.3 KB
Raw

Object Detection Sample

This sample demonstrates DL model compression capabilities for Object Detection task. The sample consists of basic steps such as DL model initialization, dataset preparation, training loop over epochs and validation steps. The sample receives a configuration file where the training schedule, hyper-parameters, and compression settings are defined.

Features

  • RetinaNet from the official TF repository with minor modifications (custom implementation of upsamling is replaced with equivalent tf.keras.layers.UpSampling2D). YOLOv4 from the keras-YOLOv3-model-set repository.
  • Support TensorFlow Datasets (TFDS) and TFRecords for COCO2017 dataset.
  • Configuration file examples for sparsity, quantization, filter pruning and quantization with sparsity.
  • Export to Frozen Graph or TensorFlow SavedModel that is supported by the OpenVINO™ toolkit.
  • Distributed training on multiple GPUs on one machine is supported using tf.distribute.MirroredStrategy.

Installation

At this point it is assumed that you have already installed nncf. You can find information on downloading nncf here.

To work with the sample you should install the corresponding Python package dependencies:

pip install -r examples/tensorflow/requirements.txt

Quantize Pretrained Model

This scenario demonstrates quantization with fine-tuning of RetinaNet with ResNet-50 backbone on the COCO2017 dataset.

Dataset Preparation

The object detection sample supports TensorFlow Datasets (TFDS) and TFRecords. The dataset type is specified in the configuration file by setting the "dataset_type" parameter to "tfds" or "tfrecords" accordingly. The TFDS format is used by default in the configuration file.

Using TFDS

Please read the following guide for more information on how to use TFDS to download and prepare a dataset. For the COCO2017 dataset, TFDS supports automatic download. All you need to do is to specify the dataset and its type in the configuration file as follows:

    "dataset": "coco/2017",
    "dataset_type": "tfds"

Legacy TFRecords

To download the COCO2017 dataset and convert it to TFRecord format please use download_and_preprocess_coco.sh script from the official TensorFlow TPU repository.

bash <path_to_tensorflow_tpu_repo>/tools/datasets/download_and_preprocess_coco.sh <path_to_coco_data_dir>

This script installs the required libraries and then runs the dataset preprocessing. The output of the script is *.tfrecord files in your local data directory.

The COCO2017 dataset in TFRecords format should be specified in the configuration file as follows:

    "dataset": "coco/2017",
    "dataset_type": "tfrecords"

Run Object Detection Sample

  • If you did not install the package, add the repository root folder to the PYTHONPATH environment variable.
  • Go to the examples/tensorflow/object_detection folder.
  • Download the pre-trained weights in H5 format and provide the path to them using --weights flag. The link to the archive with pre-trained weights can be found in the TensorFlow checkpoint column of the results table. Select the checkpoint corresponding to the None compression algorithm, which includes the pre-trained weights for the FP32 model, without applying any compression algorithms.
  • (Optional) Before compressing a model, it is highly recommended checking the accuracy of the pretrained model, use the following command: 
    python main.py \
--mode=test \
--config=configs/quantization/retinanet_coco_int8.json \
--weights=<path_to_H5_file_with_pretrained_weights>
--data=<path_to_dataset> \
--disable-compression
  • Run the following command to start compression with fine-tuning on all available GPUs on the machine: 
    python main.py \
--mode=train \
--config=configs/quantization/retinanet_coco_int8.json \
--weights=<path_to_H5_file_with_pretrained_weights>
--data=<path_to_dataset> \
--log-dir=../../results/quantization/retinanet_coco_int8
  • Use the --resume flag with the path to the checkpoint to resume training from the defined checkpoint or folder with checkpoints to resume training from the last checkpoint.

Validate Your Model Checkpoint

To estimate the test scores of your trained model checkpoint, use the following command:

python main.py \
--mode=test \
--config=configs/quantization/retinanet_coco_int8.json \
--data=<path_to_dataset> \
--resume=<path_to_trained_model_checkpoint>

Export Compressed Model

To export trained model to the Frozen Graph, use the following command:

python main.py \
--mode=export \
--config=configs/quantization/retinanet_coco_int8.json \
--resume=<path_to_trained_model_checkpoint> \
--to-frozen-graph=../../results/retinanet_coco_int8.pb

To export trained model to the SavedModel, use the following command:

python main.py \
--mode=export \
--config=configs/quantization/retinanet_coco_int8.json \
--resume=<path_to_trained_model_checkpoint> \
--to-saved-model=../../results/saved_model

To export trained model to the Keras H5, use the following command:

python main.py \
--mode=export \
--config=configs/quantization/retinanet_coco_int8.json \
--resume=<path_to_trained_model_checkpoint> \
--to-h5=../../results/retinanet_coco_int8.h5

Save Checkpoint without Optimizer

To reduce memory footprint (if no further training is scheduled) it is useful to save the checkpoint without optimizer. Use the following command:

python ../common/prepare_checkpoint.py \
--config=configs/quantization/retinanet_coco_int8.json \
--resume=<path_to_trained_model_checkpoint> \
--checkpoint-save-dir=<path_to_save_optimized_model_checkpoint>

Export to OpenVINO™ Intermediate Representation (IR)

To export a model to the OpenVINO IR and run it using the Intel® Deep Learning Deployment Toolkit, refer to this tutorial.

Train RetinaNet from scratch

  • Download pre-trained ResNet-50 checkpoint from here.
  • If you did not install the package, add the repository root folder to the PYTHONPATH environment variable.
  • Go to the examples/tensorflow/object_detection folder.
  • Run the following command to start training RetinaNet from scratch on all available GPUs on the machine: 
    python main.py \
--mode=train \
--config=configs/retinanet_coco.json \
--data=<path_to_dataset> \
--log-dir=../../results/quantization/retinanet_coco_baseline \
--backbone-checkpoint=<path_to_resnet50-2018-02-07_folder>
  • Export trained model to the Keras H5 format.

Results

Model Compression algorithm Dataset mAP (drop) % NNCF config file TensorFlow checkpoint
RetinaNet None COCO2017 33.44 retinanet_coco.json Link
RetinaNet INT8 (per-tensor, symmetric for weights; per-tensor, symmetric for activations) COCO2017 33.18 (0.26) retinanet_coco_int8.json Link
RetinaNet Sparsity 50% (Magnitude) COCO2017 33.13 (0.31) retinanet_coco_magnitude_sparsity.json Link
YOLOv4 None COCO2017 47.04 yolo_v4_coco.json Link
YOLOv4 INT8 (per-channel, symmetric for weights; per-tensor, asymmetric for activations) COCO2017 46.30 (0.74) yolo_v4_coco_int8.json Link
YOLOv4 Sparsity 50% (Magnitude) COCO2017 46.54 (0.50) yolo_v4_coco_magnitude_sparsity.json Link

Results for filter pruning

Model Compression algorithm Dataset mAP (drop) % GFLOPS MParams NNCF config file TensorFlow checkpoint
RetinaNet None COCO2017 33.44 194.1 (100%) 60.8 (100%) retinanet_coco.json Link
RetinaNet Filter Pruning 40%, geometric_median criterion COCO2017 32.7 (0.74) 107.7 (55.49%) 34.7 (57.07%) retinanet_coco_pruning.json Link
RetinaNet Filter Pruning 40%, geometric_median criterion + INT8 (per-tensor, symmetric for weights; per-tensor, symmetric for activations) COCO2017 32.68 (0.76) 102.6 (52.86%) 33.6 (55.26%) retinanet_coco_pruning_int8.json Link