search

NVIDIA-AI-IOT / redtail

Perception and AI components for autonomous mobile robotics.
BSD 3-Clause "New" or "Revised" License
1.01k stars 346 forks source link

Re-Training Second Head of TrailNet #133

Closed ktnguyen2 closed 4 years ago

ktnguyen2 commented 4 years ago

I was able to achieve 85% accuracy with the IDSIA trails dataset, and further fine-tuned using my own data. When I try to train the second head with my own data, I get very high accuracy (~98%), but when I test a list of images from my training data, the network only predicts one class.

daylanc commented 4 years ago

@ktnguyen2 I am running into the same issue of my model only predicting one class after training both rotation and translation heads. What did you do to solve this issue? I'm not understanding how to "freeze" the rotational parameters after training them. It seems like training both heads are exactly the same except for the pretrained model for the translation head.

ktnguyen2 commented 4 years ago

@daylanc I found that if you get high enough accuracy (at least 85%), then test the model by deploying it in ROS and analyze the results based on test images that you send it through via imagepub node.

It is a very tedious and crude method of evaluation, but it's the best that I found without looking deeper into NVIDIA DIGITS.

soldierofhell commented 4 years ago

@ktnguyen2, @daylanc, Hi guys, I'm preparing Tensorflow2 + TensorRT on Jetson PoC for Ardupilot open-source. Are you interested in sharing your data? Of course just for PoC implementation, not for public sharing. We all know that Caffe is closed and to move this concept forward switching to TF is mandatory (seems like TF is now more advanced with TensorRT then PyTorch), Regards

ktnguyen2 commented 4 years ago

@soldierofhell I commend your efforts! Do you mean data as in my image datasets?

soldierofhell commented 4 years ago

@ktnguyen2, I mean to my current knowledge there's no data available for lateral head so I'm asking if people who collected their own data similar as in the paper would kindly share it to prepare PoC or some reference models in TF2