Implemented retina net on CentOS

environment

• python 3.7.7
• CUDA 10.0.130
• gcc 7.3.0

​

Environment

git clone https://github.com/tensorflow/tpu.git
sudo apt-get install -y python-tk
pip install tensorflow-gpu==1.15
pip install --user Cython matplotlib opencv-python-headless pyyaml Pillow
pip install 'git+https://github.com/cocodataset/cocoapi#egg=pycocotools&subdirectory=PythonAPI'

​

Inference with trained model

1. Download any model https://github.com/tensorflow/tpu/blob/master/models/official/detection/MODEL_ZOO.md

2. Perform inference

• Example of coco_label_map.csv

  1:person
  2:bicycle
  3:car

category_id: In the form of category If you want to change the class, create a csv file according to the above format

• Execution code

  python ~/tpu/models/official/detection/inference.py \
    --model="retinanet" \
    --image_size=640\
    --checkpoint_path="./detection_retinanet_50/model.  ckpt" \
    --label_map_file="./retinanet/tpu/models/official/  detection/datasets/coco_label_map.csv" \
    --image_file_pattern="path/to/input/file" \
    --output_html="path/to/output/file" \
    --max_boxes_to_draw=10 \
    --min_score_threshold=0.05

• Output in html format.
• Official git can only infer with images.

​

3. Learn with original data

1. Download the trained data https://github.com/tensorflow/tpu/blob/master/models/official/detection/MODEL_ZOO.md ​
2. Create input data

• Input is assumed to be in tfrecorfd format.
• Prepare the dataset in coco format (image + * .json)
• Convert to tfrecord format with the following program ​

  #!/bin/bash
​
  TRAIN_IMAGE_DIR="path/to/train/images/dir"
  TRAIN_OBJ_ANNOTATIONS_FILE="path/to/train/file"
  OUTPUT_DIR="path/to/output/dir"
  VAL_IMAGE_DIR="path/to/test/images/dir"
  VAL_OBJ_ANNOTATIONS_FILE="path/to/test/images/dir"
​
  function create_train_dataset(){
    python3 create_coco_tf_record.py \
      --logtostderr \
      --include_masks \
      --image_dir="${TRAIN_IMAGE_DIR}" \
      --object_annotations_file="$  {TRAIN_OBJ_ANNOTATIONS_FILE}" \
      --output_file_prefix="${OUTPUT_DIR}/train" \
      --num_shards=256
  }
  function create_val_dataset() {
    SCRIPT_DIR=$(dirname "$(readlink -f "$0")")
    PYTHONPATH="tf-models:tf-models/research"
    python3 $SCRIPT_DIR/create_coco_tf_record.py \
      --logtostderr \
      --include_masks \
      --image_dir="${VAL_IMAGE_DIR}" \
      --object_annotations_file="$  {VAL_OBJ_ANNOTATIONS_FILE}" \
      --output_file_prefix="${OUTPUT_DIR}/val" \
      --num_shards=32
  }
​
  create_train_dataset
  create_val_dataset

​ 3. Perform learning

  MODEL_DIR="<path to the directory to store model files>"
  TRAIN_FILE_PATTERN="<path to the TFRecord training data>"
  EVAL_FILE_PATTERN="<path to the TFRecord validation data>"
  VAL_JSON_FILE="<path to the validation annotation JSON file>"
  RESNET_CHECKPOINT="<path to trained model>"
  python ~/tpu/models/official/detection/main.py \
    --model="retinanet" \
    --model_dir="${MODEL_DIR?}" \
    --mode=train \
    --eval_after_training=True \
    --use_tpu=False \
    --params_override="{train: { checkpoint: { path: ${RESNET_CHECKPOINT?}, prefix: resnet50/ }, train_file_pattern: ${TRAIN_FILE_PATTERN?} }, eval: { val_json_file: ${VAL_JSON_FILE?}, eval_file_pattern: ${EVAL_FILE_PATTERN?} }}"

• You can grasp the execution time from the log at the time of learning execution

  INFO:tensorflow:examples/sec: 0.622754
  INFO:tensorflow:global_step/sec: 0.078258

​

4. Reasoning

1. Run ​

  python ~/tpu/models/official/detection/inference.py \
      --model="retinanet" \
      --image_size=640\
      --checkpoint_path="path/to/input" \
      --label_map_file="path/to/label" \
      --image_file_pattern="path/to/input/file" \
      --output_html="path/to/output/file" \
      --max_boxes_to_draw=10 \
      --min_score_threshold=0.05

--Model before learning with original data (Inference results will be uploaded) ​ --Model after learning (Inference results will be uploaded)

5. Evaluation

1. Run

• config_file uses params.yaml which is output during training

  python ${RETINA_ROOT}/evaluate_model.py\
    --model="retinanet"\
    --checkpoint_path="path/to/imput/file"\
    --config_file="${CONFIG_PATH}"\
    --params_override="${PARAMS_PATH}"\ 
    --dump_predictions_only = True\
    --predictions_path="path/to/output/file"

​ ​

Bonus: Performance comparison

Since it's a big deal, I compared it with other models built in the past. All were batch size 8 and were trained and evaluated using the original dataset.

Learning speed

Compare the average time spent on 100 iterations.

accuracy

Compare the accuracy from the AP and inference results at 2000 iterations.