I’m trying to use a neural network to drive a car: the NN takes as input LIDAR data (an array of distances like 2.43564 m) and outputs a steering command (an integer from 1 to 5).
By using Tensorflow, I built and trained a DNNClassifier on my workstation. Afterwards, in order to evaluate the model performances, I took one single LIDAR scan and its corresponding steering command. I loaded the trained network from the last checkpoint and fed in these data: the accuracy resulted to be 100%.
Then, I literally copy-pasted all the files to the Jeston TX1 (where I had previously installed Tensorflow by means of a wheel file) and executed the exact same script I used on my workstation to evaluate the model performances: I obtained an accuracy of 0%.
I also tried the same on another workstation and I still obtained an accuracy of 100%.
How is this possible? Does it have something to do with the different architectures of the workstation and the TX1? (I don’t know if it is relevant but I used pickle to process LIDAR data).