yolov4-tiny-3l issue resolved

[aditya-5842]

Made the required changes to make the repo work for yolov4-tiny-3l.

[ciklista]

@aditya-5842 could you explain how you generate the weights file for your tiny3L? Can I use any of the pre trained darknet ones (https://github.com/AlexeyAB/darknet/releases) ?

[aditya-5842]

@aditya-5842 could you explain how you generate the weights file for your tiny3L? Can I use any of the pre trained darknet ones (https://github.com/AlexeyAB/darknet/releases) ?

Yes, you can use the pre-trained YOLO-Tiny model (https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.conv.29) to initialize the weights.

[ciklista]

@aditya-5842 could you explain how you generate the weights file for your tiny3L? Can I use any of the pre trained darknet ones (https://github.com/AlexeyAB/darknet/releases) ?

Yes, you can use the pre-trained YOLO-Tiny model (https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.conv.29) to initialize the weights.
getting the following error with the above weights file.
Command was

python save_model.py --weights ./data/yolov4-tiny-conv.29  --output ./checkpoints/yolov4-tiny-3l --input_size 416 --model yolov4-tiny-3l --tiny

(sorry for the screenshot)

[aditya-5842]

@aditya-5842 could you explain how you generate the weights file for your tiny3L? Can I use any of the pre trained darknet ones (https://github.com/AlexeyAB/darknet/releases) ?

Yes, you can use the pre-trained YOLO-Tiny model (https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.conv.29) to initialize the weights.
getting the following error with the above weights file.
Command was

python save_model.py --weights ./data/yolov4-tiny-conv.29  --output ./checkpoints/yolov4-tiny-3l --input_size 416 --model yolov4-tiny-3l --tiny

(sorry for the screenshot)

Sorry, I could not give a detailed answer.

• Like I mentioned above that use YOLOV4-Tiny pretrained to initialize the model. Then train the model on your data. You can use this notebook to train your own model.
• Once training is over, you can use your saved weights to do the conversion. Please note that ./data/yolov4-tiny-conv.29 is not the full Tiny-3l model. It contains only starting layers of the YOLOV4 (tiny) model.

You can one of my trained model. This is for detecting the license-plate/s in image.

[ciklista]

thanks for the clarification! Very helpful.

[ciklista]

@aditya-5842 any reason why you changed the __C.YOLO.ANCHORS_TINY in the config?

[aditya-5842]

@ciklista
Look at line 17. It is anchor-boxes. YOLOV4-Tiny (i.e. by default Tiny-2l) has only 2 Yolo layers at the end and each layer YOLO layer needs 3 different anchor boxes. So total we need 6 boxes (3 *2=6). These anchors are going to be reshaped into 2 x 3 x 2 array. Each continuous pair will make a different anchor box and all anchor boxes should be different.
If you notice anchor 81,82 is repeating in line 17, so it should be corrected.

Now I'm not sure I got the first two values of line 18. May be I did some google search and found these values. But it worked for me.

NOTE: Initially, with line 17 as anchors, my YOLOV4-Tiny was not working.

tensorflow-yolov4-tflite/core/config.py

Lines 17 to 18 in 13e6adc

	# __C.YOLO.ANCHORS_TINY = [23,27, 37,58, 81,82, 81,82, 135,169, 344,319]
	__C.YOLO.ANCHORS_TINY = [10,14, 23,27, 37,58, 81,82, 135,169, 344,319]

[ciklista]

you are right. The original anchors where drawing bounding boxes a little too large. Looking better with yours, thanks for the fix :)

[AntoAndGar]

It's a good job!

[rafcy]

Thank you for the contribution, just tested and it's working!

[KuoEuran]

@aditya-5842 Thank you for your great work, I have a question about the depthwise separable convolution convert from weights to tflite, do you have any idea about this?