Previously we were able to get the Nano and set it up, however it wouldn’t boot up anymore, so we did some trouble shooting, searching websites, and trying to get the Nano back on to move forward. So in the meantime we decided to simply order another one to try to connect the backend and ensure we can capture frames to actually process them. This came in at approximately 5:30pm today and is currently being set up. The plan is to get it to a point tonight where we are able to record, perform a mundane operation, and save a video using it in combination with our camera and SD card. More progress is hoped for tomorrow but these initial operations are hoped to get done tonight so that we have a good jumping off point for work/progress starting tomorrow.
In regards to the actual CNN, the code and data to train the model are ready to go, but we are waiting for amazon to fulfill our request of 8 vCPU’s before we can launch our instance and begin training. I have run some small batches on my local machine and the code works fine and produces results with a reasonable loss value, but there are 80,000 iterations and with a larger batch size it would take about 1-2 weeks to train on my machine, so aws is a large bottleneck in the process of actually training the model. Regardless of the being able to train or not, the author of the paper has shared with me the retrained weights from their results, so if aws does fulfill the request in time we should still be able to have a working model.
0 Comments