Custom model & dataset

[larrywal-express]

Thanks for your good work. I have a customized model at two detection scale trained with four class.. How do I modify the code to fit the application? Moreso, how to set up --tfds_root for quantize.py?

[lp6m]

Q1: Does two detection scale means that scale of width and height is different?

If so, you will need to modify the resize part of the input image and the scale calculation of the output bounding box image.
The number of classes is also different, so the postprocess part also needs to be modified.

https://github.com/lp6m/yolov5s_android/blob/master/app/tflite_yolov5_test/app/src/main/cpp/postprocess.cpp#L18

Q2: tfds_root in quantize.py
In general, tensorflow quantization process needs sample input data.
Instead of preparing a sample image in tfds format, you can modify representative_dataset_gen to return a custom training data image.
https://github.com/lp6m/yolov5s_android/blob/master/convert_model/quantize.py#L16-L25

[lp6m]

There are no plans to support models with different height and width sizes in this repository, but the forked project seems to be making an attempt to do so. (I have not tested this.)
rjean@a0eab90

[larrywal-express]

@Ip6m I mean two detection scale for anchors. I have just 6 anchors instead of the 9 anchors.

[lp6m]

If the number of anchors is different from original model, you need to modify postprocess.cpp in this repository.

https://github.com/lp6m/yolov5s_android/blob/master/app/tflite_yolov5_test/app/src/main/cpp/postprocess.cpp

As described also in app/README.md, The algorithm of postprocess.cpp is equivalent to Detect layer of original yolov5 repositor. However, some variables are embedded as constants. It is possible to deal with this by making some modifications.
https://github.com/ultralytics/yolov5/blob/3e7c59ad3bf5414d4b2a26e018f397e27a51c6f1/models/yolo.py#L48-L68

[apanand14]

Hello @lp6m @larrywal-express ,

I have trained yolov5s on custom dataset (classes 2) and converted model to tflite-16fp. I also verified inference using detect.py and loading tflite model. it works fine locally. But when i use that model in your android app then, it doesn't give any error but also it doesn't draw any bounding box or do any detection. Can you please let me know what exactly I'm doing wrong in running custom tflite model in your app?? Your help will be appreciated.

Thank you in advance!!

[lp6m]

@apanand14 Check the output of tfliteInterPreter and postprocess() using break point function of Android Studio.
If the number of classes is different, some parts of the code must be modified.

https://github.com/lp6m/yolov5s_android/blob/master/app/tflite_yolov5_test/app/src/main/cpp/postprocess.cpp#L18
https://github.com/lp6m/yolov5s_android/blob/master/app/tflite_yolov5_test/app/src/main/java/com/example/tflite_yolov5_test/TfliteRunner.java#L178-L184
https://github.com/lp6m/yolov5s_android/blob/master/app/tflite_yolov5_test/app/src/main/java/com/example/tflite_yolov5_test/TfliteRunner.java#L191

[lp6m]

I write the tutorial for custom model & dataset. please check.
#14