4
$\begingroup$

  1. How to train darkflow for my custom object really really fast during debugging in quad core PC and without GPU? (Can I train with about 10 images and test with only those images, just to check if all convolutions are working as expected. And with 20 epoch?). There is only once class and all license plates are similar in pattern with varying in angle and its digits in plate.

  2. I am using tiny yolo config and weight. So, what all parameters I should tune in yolo based .cfg file to do it? I feel if TV like object can be detected with same training weight and config for tiny yolo then license plate too.

Overview: I am training for object, license plate, with darkflow. I tried with about 100 custom datasets I had created. This is only for initial POC, actual implementation will have more number of images. And upon testing with test images with trained graph, object is not highlighted rather highlighting squares are shown at random location within image and random in number, starting with 2-3 square boxes to lot many number. But non of those were highlighting to the actual license plate object. It took me 20 hour of training time to verify it. I used training images for testing as well, and also a plane black screen images to test what's going on. But highlighting squares are still random in number and location even on blank screen image.

| improve this question | |
$\endgroup$
2
$\begingroup$

  1. Assuming that you are using DarkFlow, you can train your model with --train flag and don't use --gpu flag since you don't want to use the GPU memory. And yes it's possible to train with just 10 images (basically over-fitting) just to confirm the working of your model. Place those 10 images in train/images directory and the annotations of these images in train/annotations directory (in PASCAL VOC format). For faster performance, use pre-trained weights file for tuning your model so that you don't need to train for initial same layers (and since you just want to detect a single object - license plates - use tiny-yolo-voc.weights. Further more, use --epoch flag to reduce the number of epochs.

    flow --model cfg/tiny-yolo-voc-new.cfg --load bin/tiny-yolo-voc.weights --train --annotation train/Annotations --dataset train/Images --epoch 20

  2. Change the number of classes and the number of filters as defined in the readme first. Then you can change the height and width of the input image the model will be trained on. Higher resolution for more accuracy while lower resolution for higher performance and speed. You can also change the activation functions for the layers, but I recommend leaving them unchanged.

| improve this answer | |
$\endgroup$
  • $\begingroup$ I had used exactly same command and there had been the persistent issue as in rundown above. Still working on RCA intermittently. $\endgroup$ – Abhishek Dwivedi Oct 25 '17 at 7:02
  • $\begingroup$ I worked with 90 images and the model started to roughly predict after 300 epochs. Try with 300 or more epochs, then check. $\endgroup$ – Syed Rafay Oct 25 '17 at 7:13
  • $\begingroup$ I would love to hear the results from 10 images after 300+ epochs, please notify if you plan doing it. @AbhishekDwivedi $\endgroup$ – Syed Rafay Oct 27 '17 at 5:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.