Hi,
I can train a working YOLOv3 Object Detection model in Darknet that works. I can train an Image Recognition caffe model in DIGITS that works. I cannot for the life of me train an Object Detection model using custom data. I have created all the KITTI labels for the images but precision never goes above zero. I did not see an option in Mechanical Turk for caffe models. Is there a service somewhere? TIA.
Hi,
May I know how do you train your model?
We have a transfer learning toolkit which can help user to retrain a model.
Have you checked it yet?
Thanks.
@AastaLLL I did not know that. I’ll check it out. Thank you.
I just followed the instructions for installing the TLT. I installed on an AWS DL AMI and am running through the detectnet_v2 python notebook. I do not yet see how the TLT makes training an object detection caffe model to run on the Xavier easier. What am I missing?
Hi,
We recommends TLT since you meet some issue on training a model from scratch. (precision doesn’t go up)
You may not train the model correctly or it takes much more data to train a plain model.
Although TLT only support NVIDIA pre-trained model, it can be used by setting database directly.
The training job can be done via running command shown here:
https://developer.nvidia.com/blog/training-custom-pretrained-models-using-tlt/
After that, you can deploy the model with Deepstream SDK on Jetson directly.
Thanks.