zldrobit/yolov5

YOLOv5 in PyTorch > ONNX > CoreML > iOS

Info

This branch provides detection and Android code complement to branch tf-only-export. Since the release of YOLOv5 v6.0, TFLite models can be exported by export.py in ultralytics' master branch. Using models/tf.py to export models is deprecated, and this repo is mainly for Anrdroid demo app. models/tf.py uses TF2 API to construct a tf.Keras model according to *.yaml config files and reads weights from *.pt, without using ONNX.

Because this branch persistently rebases to master branch of ultralytics/yolov5, use git pull --rebase or git pull -f instead of git pull.

Usage

1. Git clone Ultralytics yolov5

git clone https://github.com/ultralytics/yolov5.git
cd yolov5

2. Convert and verify

• Convert weights to fp16 TFLite model, and verify it with

python export.py --weights yolov5s.pt --include tflite --img 320
python detect.py --weights yolov5s-fp16.tflite --img 320

or

• Convert weights to int8 TFLite model, and verify it with

python export.py --weights yolov5s.pt --include tflite --int8 --img 320 --data data/coco128.yaml
python detect.py --weights yolov5s-int8.tflite --img 320

Note that:

• int8 quantization needs dataset images to calibrate weights and activations, and the default COCO128 dataset is downloaded automatically.
• Change --img to the input resolution of your model, if it isn't 320.

3. Clone this repo (tf-android branch) for Android app

git clone https://github.com/zldrobit/yolov5.git yolov5-android

4. Put TFLite models in assets folder of Android project, and change

• inputSize to --img
• output_width according to new/old inputSize ratio
• anchors to m.anchor_grid as ultralytics#1127 (comment) in android/app/src/main/java/org/tensorflow/lite/examples/detection/tflite/DetectorFactory.java
• labelFilename according to the classes of the model in

  yolov5/android/app/src/main/java/org/tensorflow/lite/examples/detection/tflite/DetectorFactory.java

  Lines 19 to 48 in 522d65e

  	if (modelFilename.equals("yolov5s.tflite")) {
  	labelFilename = "file:///android_asset/coco.txt";
  	isQuantized = false;
  	inputSize = 640;
  	output_width = new int[]{80, 40, 20};
  	masks = new int[][]{{0, 1, 2}, {3, 4, 5}, {6, 7, 8}};
  	anchors = new int[]{
  	10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
  	};
  	}
  	else if (modelFilename.equals("yolov5s-fp16.tflite")) {
  	labelFilename = "file:///android_asset/coco.txt";
  	isQuantized = false;
  	inputSize = 320;
  	output_width = new int[]{40, 20, 10};
  	masks = new int[][]{{0, 1, 2}, {3, 4, 5}, {6, 7, 8}};
  	anchors = new int[]{
  	10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
  	};
  	}
  	else if (modelFilename.equals("yolov5s-int8.tflite")) {
  	labelFilename = "file:///android_asset/coco.txt";
  	isQuantized = true;
  	inputSize = 320;
  	output_width = new int[]{40, 20, 10};
  	masks = new int[][]{{0, 1, 2}, {3, 4, 5}, {6, 7, 8}};
  	anchors = new int[]{
  	10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
  	};
  	}

  .

Then run the program in Android Studio.

TODO:

• Add NNAPI support

EDIT:

• Update according YOLOv5 v6.0 release

If you have further question, plz ask in ultralytics#1127

Reference:

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