Got nan in loss function

[ijalalfrz]

Hi , im using this as my object detection but get nan value for classification and regression loss, im using pytorch 1.3.1

[sudarshan-kamath]

Hi, I had the same problem. This was because of the following reason:

The zeroth class is the background and I was having a dataset for detecting only a single object. I had forgotten to change the class label to 1 for the class I was detecting. Perhaps, you are having the same issue.

[thimabru1010]

Hi, I had the same problem. This was because of the following reason:

The zeroth class is the background and I was having a dataset for detecting only a single object. I had forgotten to change the class label to 1 for the class I was detecting. Perhaps, you are having the same issue.

I am having the same issue. I am training in one class only and all my loss results are nan. I printed labels and it seems my unique class label is 1 already. Where did u exactly find in ur code the class label wrong?

When I print the class_dict in _read_data on open_images.py class (actually I modified this to fit in my dataset) I have the following output:

{'BACKGROUND': 0, 'Face': 1}

[zinercodes]

@thimabru1010 could you resolve the issue?

[thimabru1010]

I remember I managed to found a solution, but this was a long time ago. Seeing my code maybe the problem was on my box labels. I was forgetting to divide bboxes by width and height. But quite not sure. If you want I can share my code with you.

[zinercodes]

@thimabru1010 thanks for quick reply. Yes, if you share your code, it will be really helpful.

[thimabru1010]

Hey @zinercodes,

Later than never!

I put the code in my repository: https://github.com/thimabru1010/ssd-thermal-face-detection

I was trying to train the SSD with mobilenet on a custom dataset. To do this I created an organize.py to generate put data (boxes and class) on the right structure. We only need the image_id, bboxes coordinates, and class name. I think the problem I was facing was due to the normalization of bounding boxes. I wasn't dividing by width and height. Not sure if this is your problem, but hope it helps.

Sorry for the late reply.

Here is a picture of the old commit and the change I made. Hope this guides you.