1. For cusotm models fix the config.py file with own obj.names where there is coco.names. Issue fix, hunglc007#147 (comment). In this repo I have replaced coco.names with obj.names. It should contain custom class names.
2. Set private static boolean isTiny = true; instead of false to make android code work for default model with repo.
3. Commented out #converter.representative_dataset = representative_data_gen in convert_tflite.py since I don't have coco dataset. Reference, hunglc007#214.
4. For converting yolov4 tiny model to tflite use --framework tflite flag in save_model.py. Reference, hunglc007#234 (comment).
5. For some reason the provided coco model and some custom models detect objects correctly, but bounding box is shown on top left corner. Setting isGPU = false in YoloV4Classifier.java increases inference time, but allows bounding boxes to be drawn correctly. One of my custom yolov4 tiny model worked with GPU, but no longer works. 🎯

Train custom model with, https://github.com/AlexeyAB/darknet and deploy with android example. Set isTiny = false for full yolov4, TF_OD_API_IS_QUANTIZED = true for int8 quantized model and appropriate model path, class name.

An issue with unsupported ops EXP, SPLIT, SPLIT_V is shown when deploying to android. It is solved by using implementation 'org.tensorflow:tensorflow-lite-select-tf-ops:0.0.0-nightly' on android build.gradle and below code to convert tensorflow saved model. It is unquantized FP32 model.

Tiny weights to saved model,

## yolov4-tiny
!python save_model.py \
--weights "yolov4-tiny-obj_best.weights" \
--output "checkpoints/yolov4-tiny-416" \
--input_size 320 --model yolov4 --tiny --framework tflite

Saved model to tflite,

import tensorflow as tf
converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]

converter.target_spec.supported_ops = [tf.lite.OpsSet.SELECT_TF_OPS]
#converter.target_spec.supported_types = [tf.float16]
#converter.target_spec.supported_ops = [tf.lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8]

tflite_quant_model = converter.convert()

tflite_path = '/content/model.tflite'
with open(tflite_path, 'wb') as f:
  f.write(tflite_quant_model)

Classes can also be read from *.txt file by providing path. It can also be added in YoloV4Classifier.java by adding lines like below and commenting out code below.

d.labels.add("car");
d.labels.add("bike");

//        String actualFilename = labelFilename.split("file:///android_asset/")[1];
//        InputStream labelsInput = assetManager.open(actualFilename);
//        BufferedReader br = new BufferedReader(new InputStreamReader(labelsInput));
//        String line;
//        while ((line = br.readLine()) != null) {
//            LOGGER.w(line);
//            d.labels.add(line);
//        }
//        br.close();

The given models detections show on the top left corner for some reason. I remember fixing it, but can no longer find the fix. Yet my custom trained tflite model on 320x320 with FP16 quantization works correctly. Also the provided model with the repo is yolo tiny.

For custom model deployment OUTPUT_WIDTH_TINY or OUTPUT_WIDTH_FULL needs to be filled with proper value. The value can be easily found on the error log. Yolov4 full not tested yet.

java.lang.IllegalArgumentException: Cannot copy from a TensorFlowLite tensor (Identity) with shape [1, 1500, 4] to a Java object with shape [1, 2535, 4].

For example, if current width values are {2535, 2535} then just change it to {1500, 1500}.

In YoloV4Classifier.java change INPUT_SIZE to appropriate size. For example if network height width is 320 then input size is 320. Also change TF_OD_API_INPUT_SIZE to appropriate resolution value in MainActivity.java for single image and DetectorActivity.java for realtime processing.

Set TF_OD_API_IS_QUANTIZED to true if the custom model is FP16 quantized.

This is an error from converted tflite model. I fixed it previously, but no longer working. One probable(need to check) reason is the model was trained in a different resolution say for example 608x608 in training *.cfg, but when converting to saved model it was asked to use 416 or 320.

An important note, hunglc007#128.

YOLOv4, YOLOv4-tiny Implemented in Tensorflow 2.0. Convert YOLO v4, YOLOv3, YOLO tiny .weights to .pb, .tflite and trt format for tensorflow, tensorflow lite, tensorRT.

Download yolov4.weights file: https://drive.google.com/open?id=1cewMfusmPjYWbrnuJRuKhPMwRe_b9PaT

• Tensorflow 2.3.0 on build.gradle and using same version to convert model. Colab can be used.

# Convert darknet weights to tensorflow
## yolov4
python save_model.py --weights ./data/yolov4.weights --output ./checkpoints/yolov4-416 --input_size 416 --model yolov4 

## yolov4-tiny
python save_model.py --weights ./data/yolov4-tiny.weights --output ./checkpoints/yolov4-tiny-416 --input_size 416 --model yolov4 --tiny

# Run demo tensorflow
python detect.py --weights ./checkpoints/yolov4-416 --size 416 --model yolov4 --image ./data/kite.jpg

python detect.py --weights ./checkpoints/yolov4-tiny-416 --size 416 --model yolov4 --image ./data/kite.jpg --tiny

If you want to run yolov3 or yolov3-tiny change --model yolov3 in command

# Save tf model for tflite converting
python save_model.py --weights ./data/yolov4.weights --output ./checkpoints/yolov4-416 --input_size 416 --model yolov4 --framework tflite

# yolov4
python convert_tflite.py --weights ./checkpoints/yolov4-416 --output ./checkpoints/yolov4-416.tflite

# yolov4 quantize float16
python convert_tflite.py --weights ./checkpoints/yolov4-416 --output ./checkpoints/yolov4-416-fp16.tflite --quantize_mode float16

# yolov4 quantize int8
python convert_tflite.py --weights ./checkpoints/yolov4-416 --output ./checkpoints/yolov4-416-int8.tflite --quantize_mode int8 --dataset ./coco_dataset/coco/val207.txt

# Run demo tflite model
python detect.py --weights ./checkpoints/yolov4-416.tflite --size 416 --model yolov4 --image ./data/kite.jpg --framework tflite

Yolov4 and Yolov4-tiny int8 quantization have some issues. I will try to fix that. You can try Yolov3 and Yolov3-tiny int8 quantization

python save_model.py --weights ./data/yolov3.weights --output ./checkpoints/yolov3.tf --input_size 416 --model yolov3
python convert_trt.py --weights ./checkpoints/yolov3.tf --quantize_mode float16 --output ./checkpoints/yolov3-trt-fp16-416

# yolov3-tiny
python save_model.py --weights ./data/yolov3-tiny.weights --output ./checkpoints/yolov3-tiny.tf --input_size 416 --tiny
python convert_trt.py --weights ./checkpoints/yolov3-tiny.tf --quantize_mode float16 --output ./checkpoints/yolov3-tiny-trt-fp16-416

# yolov4
python save_model.py --weights ./data/yolov4.weights --output ./checkpoints/yolov4.tf --input_size 416 --model yolov4
python convert_trt.py --weights ./checkpoints/yolov4.tf --quantize_mode float16 --output ./checkpoints/yolov4-trt-fp16-416

python benchmarks.py --size 416 --model yolov4 --weights ./data/yolov4.weights

Training can be done with Darknet using custom dataset. Both yolov4 full and tiny version is supported. Process for going from darknet weights to android is, darknet weight -> saved model -> tflite/quantied tflite.

• YOLOv4: Optimal Speed and Accuracy of Object Detection YOLOv4.
• darknet

This project is inspired by these previous fantastic YOLOv3 implementations:

• Yolov3 tensorflow
• Yolov3 tf2

This repo is based on:

• https://github.com/hunglc007/tensorflow-yolov4-tflite