EfficientDet-tf-custom_dataset

EfficientDet tensorflow object detection implementation with custom dataset

This is based on the official implentation of EfficientDet by google. Have a look at their paper for more theoritical knowledge. Thank you for the great work.

Table of Contents

1. Prerequisites
2. Inference
3. Results
4. Training
   1. Preprocess
   2. Training
5. Common_issues

Prerequisites

• lxml

• Cython

• matplotlib

• pycocotools (pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI)

• Tensorflow

• recommedned : create a conda environment with tensorflow 2 by using following command

  conda create -n tf2 tensorflow-gpu
  conda activate tf2

  Note : Make sure that you are in the above conda environment while processing every operations in this.

Inference

• Open the inference.py file and modify the line with your custom class and mapping id same as mentioned while generating tfrecords.

• Following command is used for inference.

    python model_inspect.py --runmode=infer --model_name=$MODEL \
    --input_image_size=1920x1280 --max_boxes_to_draw=100   --min_score_thresh=0.2 \
    --ckpt_path='trained weight with path' --input_image=img.png --num_classes  1 --output_image_dir='path to output directory'

Results

Will update on this soon

Training

Preprocess

• Tfrecords have to be generated from custom dataset, which will be acceptable by the tensorflow training pipeline.

• Lets make a folder tfrecord

  cd EfficientDet-tf-custom_dataset
  mkdir tfrecord

a) Kitti dataset to tfrecord

1. Kitti dataset can be converted into tfrecord directly by using the following command( script is from tensorflow repo)

   python kiti_convert/create_kitti2tfrecord.py \
       --data_dir='path of kitti_dataset folder'
       --output_path='path to store tfrecords'
       --label_map_path = 'path to label_map.pbtxt'

   Note : install tf object detection api by pip install tensorflow-object-detection-api

• Below is the folder structre of kitti format used by create_kitti2tfrecord.py

  kitti_dataset/ 
        ├─images/
        |       ├─ img1.jpg
        |       ├─ img2.jpg
        ├─labels/ 
                ├─ img1.txt
                ├─ img2.txt 
  

• label_map.pbtxt file format for one class

  item {
  id: 1
  name: 'class_name'
  }
  

  Note : There are also several scripts available to convert from kitti to tfrecords. (eg: cvdata_convert command from cvdata pip package)

2. Kitti cab be converted into Pascal voc by using the following script

   python kitti_convert/kitti2voc.py -ipi 'path to images' -ipl 'path to kitti labels'

   The dataset/create_pascal_tfrecord.py script to generate tfrecords from PascalVOC whihc is mentioned below.

b) PascalVOC to tfrecord

• Open the create_tfrecord.py file in the dataset folder and modify this line with custom classes.

• create train.txt file by using following command

  cd VOCdevkit/VOC2012 && ls -1 JPEGImages | cut -d. -f1 > train.txt && cd -
  head VOCdevkit/VOC2012/train.txt

  The above command will generate a train.txt file under VOCdevkit/VOC2012 , move it into ImageSets/Main directory.

• Run the following command and generate the tfrecords from PascalVOC.

  PYTHONPATH=".:$PYTHONPATH"  python dataset/create_pascal_tfrecord.py  \
      --data_dir=VOCdevkit --year=VOC2012  --output_path=tfrecord/pascal

  Note : The tf record can be alos genreated for train and validation if needed.

Training

• Download backbone using follwoing command into the EfficientDet-tf-custom_dataset folder.

  wget https://storage.googleapis.com/cloud-tpu-checkpoints/efficientnet/ckptsaug/efficientnet-b0.tar.gz
  tar xf efficientnet-b0.tar.gz 

• Start the training using following command

  python main.py --mode=train_and_eval \
      --training_file_pattern=tfrecord/pascal*.tfrecord \
      --validation_file_pattern=tfrecord/pascal*.tfrecord \
      --val_json_file=tfrecord/json_pascal.json \
      --model_name=efficientdet-d0 \
      --model_dir=/tmp/efficientdet-d0-scratch  \
      --backbone_ckpt=efficientnet-b0  \
      --train_batch_size=8 \
      --eval_batch_size=8 --eval_samples=512 \
      --num_examples_per_epoch=5717 --num_epochs=1  \
      --hparams="use_bfloat16=false,num_classes=20,moving_average_decay=0" \
      --use_tpu=False

Common issues

• TypeError: object of type <class 'numpy.float64'> cannot be safely interpreted as an integer.
  solution : pip install numpy==1.17.0 (downgrade numpy from 1.18 to 1.17)

• tensorflow.python.framework.errors_impl.InvalidArgumentError: assertion failed [[{{node parser/Assert/Assert}}]] [[IteratorGetNext]]

  solution : change MAX_NUM_INSTANCES with maximum number of instances in image of the custom dataset

Note : For more features and information , please visit the official EfficientDet-tf github repo