[FR] Enable YOLOv4 engine

[bafu]

Concept If we provide YOLOv4 engine, YOLO user can have a faster and more accurate detection engine to use.

Suggested Implementation (if any)

1. Study inference performance
   • TFLite (by tensorflow-yolov4-tflite) - @bafu
   • Darknet (by AlexeyAB's version) - @grandpaul
2. Implement v4 engine and service.

[bafu]

TFLite model

yolov4-tiny-416.zip

Steps to create the TFLite model

1. Convert raw model to TF checkpoints

   $ python3 save_model.py --weights ./data/yolov4-tiny.weights --output ./checkpoints/yolov4-tiny-416-tflite --input_size 416 --model yolov4 --tiny --framework tflite
   

2. Convert checkpoints to TFLite model

   $ python3 convert_tflite.py --weights ./checkpoints/yolov4-tiny-416-tflite --output ./checkpoints    /yolov4-tiny-416.tflite
   

3. Run a testing inference by using the TFLite model

   $ python3 detect.py --weights ./checkpoints/yolov4-tiny-416.tflite --size 416 --model yolov4 --image ./data/kite.jpg --framework tflite
   

Benchmark

detect.py.patch.txt

CPU: Intel i7-10710U (12-core)

Inference takes 0.11111974716186523 s

Environment

• tensorflow-yolov4-tflite: commit dafb42f
• tensorflow-cpu: 2.2.0

[bafu]

Darknet benchmark

CPU: Intel Pentium Gold G5400

• v4
  • Standard: 26s
  • Tiny: 2.1s
• v3
  • Standard: 28s
  • Tiny: 2.1s

[bafu]

Failed to run the TFLite model on RPi 4:

pi@raspberrypi:~/codes/tensorflow-yolov4-tflite $ python3 detect.py --weights ./checkpoints/yolov4-tiny-416.tflite --size 416 --model yolov4 --image ./data/kite.jpg --framework tflite
Traceback (most recent call last):
  File "detect.py", line 90, in <module>
    app.run(main)
  File "/usr/local/lib/python3.7/dist-packages/absl/app.py", line 300, in run
    _run_main(main, args)
  File "/usr/local/lib/python3.7/dist-packages/absl/app.py", line 251, in _run_main
    sys.exit(main(argv))
  File "detect.py", line 49, in main
    interpreter = tf.lite.Interpreter(model_path=FLAGS.weights)
  File "/usr/local/lib/python3.7/dist-packages/tensorflow/lite/python/interpreter.py", line 77, in __init__
    model_path))
ValueError: Didn't find op for builtin opcode 'RESIZE_BILINEAR' version '3'
Registration failed.

TF Wheel is from PINTO0309

pi@raspberrypi:~/codes/tensorflow-yolov4-tflite $ python3 -m pip freeze | grep tensorflow
tensorflow==1.14.0
tensorflow-estimator==1.14.0

Solution

Upgrade TF from 1.14.0 to 2.2.0 can fix this issue.

[bafu]

YOLOv4 Tiny (TFLite) inference time on RPi 4

pi@raspberrypi:~/codes/tensorflow-yolov4-tflite $ python3 detect.py --weights ./checkpoints/yolov4-tiny-416.tflite --size 416 --model yolov4 --image ./data/kite.jpg --framework tflite                   
[{'name': 'input_1', 'index': 0, 'shape': array([  1, 416, 416,   3]), 'shape_signature': array([  1, 416, 416,   3]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
[{'name': 'Identity', 'index': 224, 'shape': array([   1, 2535,    4]), 'shape_signature': array([   1, 2535,    4]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'Identity_1', 'index': 206, 'shape': array([   1, 2535,   80]), 'shape_signature': array([   1, 2535,   80]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
Inference takes 0.8501577377319336 s

[bafu]

YOLOv4 Tiny TFLite model and label files: https://github.com/DT42/BerryNet/pull/49#issuecomment-674473590