How do I train this model on a custom dataset?

[qqaadir]

Hi @soo89,

Wonderful work.

Could you please help understand how can I train this semi-supervised model on my own dataset. I only have images and xml files. I have not worked with VOC/COCO datasets before. Could you please give some suggestions about where should I start? Do I need to split my dataset into labelled and unlabelled datasets to perform semi-supervised learning?

I appreciate your response.

[soo89]

Thanks for your interest.

I suggest you try to train the basic CSD model with PASCAL VOC first.

And, please check the "voc07_consistency.py" file in the data folder.

If you understand how it works, you can apply this algorithm to your dataset.

[qqaadir]

@soo89
Thank you for the prompt response. I will check this "voc07_consistency.py" out. I have trained your model on VOC2007 and VOC2012 datasets. But I am having difficulty in understanding how semi-supervised learning is working. For semi-supervised learning, one should be having a set of labelled and unlabelled images as shown in Fig. 1 (d) of your paper. But VOC2007 and VOC2012 are fully labelled. How did you use partial labels in your code to perform semi-supervised learning? Could you please explain this part to me? Thank you.

[soo89]

please check 'voc07_consistency.py' in detail.

I did not use VOC2012 labeled data.

In other words, it is used the VOC07 images and labeled data and only VOC12 images.

    if (img_id[0][(len(img_id[0]) - 7):] == 'VOC2007'):
        target = ET.parse(self._annopath % img_id).getroot()
        img = cv2.imread(self._imgpath % img_id)
        semi = np.array([1])
    elif (img_id[0][(len(img_id[0]) - 7):] == 'VOC2012'):
        img = cv2.imread(self._imgpath % img_id)
        target = np.zeros([1, 5])
        semi = np.array([0])

[qqaadir]

@soo89 Got it. Thanks. But I am still not sure how to train on my own dataset. I guess, the easiest method that comes to my mind is to replace XML and JPG files in their respective folders with my own files and create trainval.txt files (filenames) for both VOC2007 and VOC2012. Also, I have to change my filenames to 000001.xml (and .jpg). What you think, am I missing something?

[soo89]

I also think that it seems the easiest way.
There seems to be no problem. Good luck.

[qqaadir]

@soo89 One last question before you close this thread. After training this model using train_csd.py and evaluating using eval.py, what will be the output? Will I get label files (.xml) in terms of bounding boxes for unlabelled data? Also, do I must perform training using train_ssd.py and train_csd.py and then run eval.py Or I can just run train_csd.py and then eval.py for semi-supervised?

[soo89]

it does not need a pre-trained supervised object detection model.
It means that you can train only with 'train_csd.py'.

The output of 'eval.py' is the mAP score for the VOC07 test.
To get a box and score, you need to find the code or make the code.

[qqaadir]

Thanks for all your answers to my questions. I will contact you again if i have any other questions. You can close this thread.