2.2. Training for Object Detection¶

This tutorial covers training a neural network model on a GPU server to perform object detection of Human Faces in the Open Images Dataset V4.

2.2.1. Preparation¶

The Open Images V4 dataset is available from the official website, however it is too large for this tutorial. So we provide a reduced dataset.

Blueoil supports 2 formats for object detection.

OpenImagev4 format
DeLTA-Mark format

Note: Please see the details in Prepare training dataset

You can download a subset of Open Images V4 from: our server.

$ wget https://s3-ap-northeast-1.amazonaws.com/leapmind-public-storage/datasets/openimages_face.tgz
$ tar xf openimages_face.tgz

This dataset consists of 2866 Human Face images and 5170 annotation boxes.

2.2.2. Generate a configuration file¶

Generate your model configuration file interactively by running the python blueoil/cmd/main.py init command.

$ PYTHONPATH=.:lmnet:dlk/python/dlk python blueoil/cmd/main.py init

Below is an example of initialization.

#### Generate config ####
your model name ():  objectdetection
choose task type  object_detection
choose network  LMFYoloQuantize
choose dataset format  OpenImagesV4
training dataset path:  {dataset_dir}
set validataion dataset? (if answer no, the dataset will be separated for training and validation by 9:1 ratio.)  no
batch size (integer):  16
image size (integer x integer):  224x224
how many epochs do you run training (integer):  1000
initial learning rate:  0.001
choose learning rate schedule ({epochs} is the number of training epochs you entered before):  '3-step-decay' -> learning rate decrease by 1/10 on {epochs}/3 and {epochs}*2/3 and {epochs}-1
enable data augmentation?  Yes
Please choose augmentors:  done (5 selections)
-> select Brightness, Color, FlipLeftRight, Hue, SSDRandomCrop
apply quantization at the first layer?  no

If configuration finishes, the configuration file is generated in the {Model name}.yml under current directory.

When you wnat to create config yaml in specific filename or directory, you can use -o option.

$ PYTHONPATH=.:lmnet:dlk/python/dlk python blueoil/cmd/main.py init -o ./configs/my_config.yml

2.2.3. Train a network model¶

Train your model by running python blueoil/cmd/main.py train with a model configuration.

$ PYTHONPATH=.:lmnet:dlk/python/dlk python blueoil/cmd/main.py train -c {PATH_TO_CONFIG.yml}

When training has started, the training log and checkpoints are generated under ./saved/{Mode name}_{TIMESTAMP}.

Training runs on the TensorFlow backend. So you can use TensorBoard to visualize your training process.

$ tensorboard --logdir=saved/{Model name}_{TIMESTAMP} --port {Port}

Metrics / Accuracy

Loss, Weight Decay

Images / Final Detect Boxes

2.2.4. Convert training result to FPGA ready format.¶

Convert trained model to executable binary files for x86, ARM, and FPGA. Currently, conversion for FPGA only supports Intel Cyclone® V SoC FPGA.

$ PYTHONPATH=.:lmnet:dlk/python/dlk python blueoil/cmd/main.py convert -e {Model name}

python blueoil/cmd/main.py convert automatically executes some conversion processes.

Converts Tensorflow checkpoint to protocol buffer graph.
Optimizes graph.
Generates source code for executable binary.
Compiles for x86, ARM and FPGA.

If conversion is successful, output files are generated under ./saved/{Mode name}_{TIMESTAMP}/export/save.ckpt-{Checkpoint No.}/{Image size}/output.

output
 ├── fpga (include preloader and FPGA configuration file)
 │   ├── preloader-mkpimage.bin
 │   ├── soc_system.rbf
 │   └── soc_system.dtb
 ├── models
 │   ├── lib (include trained model library)
 │   │   ├── lib_arm.so
 │   │   ├── lib_fpga.so
 │   │   └── lib_x86.so
 │   └── meta.yaml (model configuration)
 ├── python
 │   ├── lmnet (include pre-process/post-process)
 │   ├── README.md
 │   ├── requirements.txt
 │   ├── run.py (inference script)
 │   └── usb_camera_demo.py (demo script for object detection)
 └── README.md

2.2.5. Run inference script on x86 Linux (Ubuntu 16.04)¶

Prepare images for inference (not included in the training dataset)

You can find test imgaes on Creative Commons. Sample
```
  $ wget https://farm4.staticflickr.com/1172/1144309435_eff42ee683_o.jpg
```
Run the inference script.

Explore into the output/python directory, and run run.py and the inference result is saved in ./output/output.json.

Note: If you run the script for the first time, you have to setup a python environment (2.7 or 3.5+) and required python packages.
```
$ cd {output/python directory}
$ sudo pip install -r requirements.txt  # for the first time only
$ python run.py \
      -i {inference image path} \
      -m ../models/lib/lib_x86.so \
      -c ../models/meta.yaml
```
Tips: The default thredhold for object detection is 0.05. If you find too many boxses when running demo, you can edit meta.yml and set threshold to 0.4 or 0.5 as below code.
```
ExcludeLowScoreBox:
    threshold: 0.4
```

Check inference results. An example output file is shown below.

{
    "classes": [
        {
            "id": 0,
            "name": "Humanface"
        }
    ],
    "date": "2019-01-25T13:15:00.928639",
    "results": [
        {
            "file_path": "{inference_image_path}",
            "prediction": [
                {
                    "box": [
                        1151.0905382973808,
                        413.834205695561,
                        836.2847709655762,
                        1097.0762968063354
                    ],
                    "class": {
                        "id": 0,
                        "name": "Humanface"
                    },
                    "score": "0.8219463229179382"
                }
            ]
        }
    ],
    "task": "IMAGE.OBJECT_DETECTION",
    "version": 0.2
}