SqueezeDet on Keras

SqueezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real-Time Object Detection for Autonomous Driving

By Bichen Wu, Alvin Wan, Forrest Iandola, Peter H. Jin, Kurt Keutzer (UC Berkeley & DeepScale)

This repository contains a Keras implementation of SqueezeDet, a convolutional neural network based object detector described in this paper: https://arxiv.org/abs/1612.01051. If you find this work useful for your research, please consider citing:

@inproceedings{squeezedet,
    Author = {Bichen Wu and Forrest Iandola and Peter H. Jin and Kurt Keutzer},
    Title = {SqueezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real-Time Object Detection for Autonomous Driving},
    Journal = {arXiv:1612.01051},
    Year = {2016}
}

Installation

Please have a look at our Installation Guide

How do I run it?

I will show an example on the KITTI dataset. If you have any doubts, most scripts run with the -h flag give you the arguments you can pass

• Download the KITTI training example from here and here

• Unzip them

  unzip data_object_image_2.zip

  unzip data_object_label_2.zip

  You should get a folder called training.

• Inside the repository folder create a folder for the experiment. If you don't mind or dont want to type .. all the time you can do it in the scripts folder

  cd path/to/squeezeDet

  mkdir experiments

  mkdir experiments/kitti

  cd experiments/kitti

• SqueezeDet takes a list of images with full paths to the images and the same for labels. It's the same for training and evaluation. Create a list of full path names of images and labels:

  find /path/to/training/image_2/ -name "*png" | sort > images.txt

  find /path/to/training/label_2/ -name "*txt" | sort > labels.txt

• Create a training test split

  python ../../main/utils/train_val_split.py

  You should get img_train.txt, gt_train.txt, img_val.txt gt_val.txt, img_test.txt, gt_test.txt . Testing set is empty by default.

• Create a config file

  python ../../main/config/create_config.py

  Depending on the GPU change the batch size inside squeeze.config and other parameters like learning rate.

• Run training, this starts with pre-trained weights from imagenet

  python ../../scripts/train.py --init ../../main/model/imagenet.h5

• In another shell, to run evaluation

  • If you have no second GPU or none at all:

    python ../../scripts/eval.py --gpu ""

  • Otherwise:

    python ../../scripts/eval.py

    This will run evaluation in parallel on the second GPU.

• To run training on multiple GPUS:

  python ../../scripts/train.py --gpus 2 --init ../../main/model/imagenet.h5

  To run on the first 2 GPUS. Then you have to run evaluation on the third or CPU, if you have it.

• scripts/scheduler.py allows you to run multiple trainings after another. Check out the dummy scripts/schedule.config for an example. Run this with

  python ../../scripts/scheduler.py --schedule ../../scripts/schedule.config --train ../../scripts/train.py --eval ../../scripts/eval.py

Tensorboard visualization

For tensoboard visualization you can can run:

tensorboard --logdir log

Open in your brower localhost:6006 or the IP where you ran the training. On the first page you can see the losses, sublosses and metrics like mean average precision and f1 scores.

On the second page, you find visualizations of a couple of validation images with their ground truth bounding boxes and how the predictions change over the course of the training.

The third page gives you a nice view over the network graph.