Training Detectnet on COCO Dataset

[benlemond]

I am trying to train a detectnet model on the COCO dataset. I have downloaded the entire dataset, and I have run the coco2kitti.py script to generate the label files from the COCO annotations. I have made a subset of the data about ~6,000 images for training and ~1000 images for validation.

I performed the rest of the steps per a couple of tutorials, yet my training does not seem to be yielding any results at all. In fact, the charts in Digits only show "loss_bbox(train)" and "loss_coverage(train)". There is no mAP line, precision line, or any other line on my charts. Has anyone had these issues or successfully trained detectnet on the COCO dataset? Is there something I am missing in data preparation?

I loaded the dataset up to Digits

• Padding: 640x640
• Minimum box size: 25
• custom classes: dontcare,person
• rest = defaults

I created an object detection model

• epochs: 100
• subtract mean: none
• Solver: adam
• learning rate: 2.5e-05
• Policy: exponential decay
• Custom Network: https://github.com/dusty-nv/jetson-inference/blob/master/data/networks/detectnet.prototxt
• pretrained: bvlc_googlenet.caffemodel

Prototxt: detectnet_prototxt.txt

[yeyaguang2018]

hi,That's the same with me. Will you solve it now?

[liuchang138929]

Did you fix it?
I follow the guide to train the DetectNet in https://github.com/dusty-nv/jetson-inference same problem,,,

[jogiji]

Same issue here tooo

[jogiji]

Update : I tried with larger & different dataset and different learning rate than what was mentioned in tutorial and my model was successful in detecting the objects. I wish detect net could be trained in tensorflow

[rainylt]

Update : I tried with larger & different dataset and different learning rate than what was mentioned in tutorial and my model was successful in detecting the objects. I wish detect net could be trained in tensorflow

Do you get mAP curve finally? Does it mean that when you use a small dataset you can't get the mAP curve?