TensorFlow实现的SqueezeDet卷积网路目标检测

SqueezeDet安装与使用

最新推荐文章于 2023-03-09 23:00:41 发布

转载最新推荐文章于 2023-03-09 23:00:41 发布 · 2.5k 阅读

TensorFlow 专栏收录该内容

86 篇文章

订阅专栏

本文介绍SqueezeDet实时目标检测器的安装步骤与演示流程，包括克隆仓库、设置虚拟环境、安装依赖包等内容，并提供了训练及评估模型的具体指南。

Installation:

The following instructions are written for Linux-based distros.

Clone the SqueezeDet repository:
```
git clone https://github.com/BichenWuUCB/squeezeDet.git
```
Let's call the top level directory of SqueezeDet $SQDT_ROOT.

https://github.com/BichenWuUCB/squeezeDet

(Optional) Setup your own virtual environment.
1. The following assumes python is the Python2.7 executable. Navigate to your user home directory, and create the virtual environment there.
```
cd ~
virtualenv env --python=python
```
1. Launch the virtual environment.
```
source env/bin/activate
```
Use pip to install required Python packages:
```
pip install -r requirements.txt
```

Demo:

Download SqueezeDet model parameters from here, untar it, and put it under $SQDT_ROOT/data/ If you are using command line, type:

cd $SQDT_ROOT/data/
wget https://www.dropbox.com/s/a6t3er8f03gdl4z/model_checkpoints.tgz
tar -xzvf model_checkpoints.tgz
rm model_checkpoints.tgz

Now we can run the demo. To detect the sample image $SQDT_ROOT/data/sample.png,
```
cd $SQDT_ROOT/
python ./src/demo.py
```
If the installation is correct, the detector should generate this image:

To detect other image(s), use the flag --input_path=./data/*.png to point to input image(s). Input image(s) will be scaled to the resolution of 1242x375 (KITTI image resolution), so it works best when original resolution is close to that.
SqueezeDet is a real-time object detector, which can be used to detect videos. The video demo will be released later.

Training/Validation:

Download KITTI object detection dataset: images and labels. Put them under $SQDT_ROOT/data/KITTI/. Unzip them, then you will get two directories: $SQDT_ROOT/data/KITTI/training/ and $SQDT_ROOT/data/KITTI/testing/.

Now we need to split the training data into a training set and a vlidation set.

cd $SQDT_ROOT/data/KITTI/
mkdir ImageSets
cd ./ImageSets
ls ../training/image_2/ | grep ".png" | sed s/.png// > trainval.txt

trainval.txt contains indices to all the images in the training data. In our experiments, we randomly split half of indices in trainval.txt into train.txt to form a training set and rest of them into val.txt to form a validation set. For your convenience, we provide a script to split the train-val set automatically. Simply run

cd $SQDT_ROOT/data/
python random_split_train_val.py

then you should get the train.txt and val.txt under $SQDT_ROOT/data/KITTI/ImageSets.

When above two steps are finished, the structure of $SQDT_ROOT/data/KITTI/ should at least contain:

$SQDT_ROOT/data/KITTI/
                  |->training/
                  |     |-> image_2/00****.png
                  |     L-> label_2/00****.txt
                  |->testing/
                  |     L-> image_2/00****.png
                  L->ImageSets/
                        |-> trainval.txt
                        |-> train.txt
                        L-> val.txt

Next, download the CNN model pretrained for ImageNet classification:

cd $SQDT_ROOT/data/
# SqueezeNet
wget https://www.dropbox.com/s/fzvtkc42hu3xw47/SqueezeNet.tgz
tar -xzvf SqueezeNet.tgz
# ResNet50 
wget https://www.dropbox.com/s/p65lktictdq011t/ResNet.tgz
tar -xzvf ResNet.tgz
# VGG16
wget https://www.dropbox.com/s/zxd72nj012lzrlf/VGG16.tgz
tar -xzvf VGG16.tgz

Now we can start training. Training script can be found in $SQDT_ROOT/scripts/train.sh, which contains commands to train 4 models: SqueezeDet, SqueezeDet+, VGG16+ConvDet, ResNet50+ConvDet. Un-comment the model you want to train, and then, type the following to train using only the CPU:
```
cd $SQDT_ROOT/
./scripts/train.sh squeezeDet
```
To train using the GPU, add the -gpu flag.
```
./scripts/train.sh squeezeDet -gpu
```
Training logs are saved to the directory specified by --train_dir.
At the same time, you can launch evaluation by
```
cd $SQDT_ROOT/
./scripts/eval_train.sh
./scripts/eval_val.sh
```
If you've changed the --train_dir in the training script, make sure to also change --checkpoint_dir in the evaluation script to the same as --train_dir so evaluation script knows where to find the checkpoint. The evaluation logs will be dumped into the directory specified by --eval_dir. It's recommended to put --train_dir and --eval_dir under the same $LOG_DIR such that tensorboard can load both training and evaluation logs.

The two scripts simultaneously evaluate the model on training and validation set. The training script keeps dumping checkpoint (model parameters) to the training directory once every 1000 steps (step size can be changed). Once a new checkpoint is saved, evaluation threads load the new checkpoint file and evaluate them on training and validation set.
Finally, to monitor training and evaluation process, you can use tensorboard by
```
tensorboard --logdir=$LOG_DIR
```
Here, $LOG_DIR is the directory where your training and evaluation threads dump log events. As we mentioned it before, your training directory is specified by the flag --train_dir and your evaluation directory is specified by --eval_dir. Then $LOG_DIR should be the upper directory to --train_dir and --eval_dir. From tensorboard, you should be able to see a lot of information including loss, average precision, error analysis, example detections, model visualization, etc.