I ran the initial cifar10 network from Caffe-Jacinto repo on the TDA2x, and got an accuracy of 89%.
This is the network:
I tried training the same network (using the same solver.prototxt, train,prototxt, and deploy.prototxt) in original Caffe.
While the network did train in Caffe, it did not work on the TDA2x. The network did import and run on EVE, but the output was the same for all test images.
Does Caffe-Jacinto make the resulting networks more compatible with the TDA2x?
