Quantization Aware Training using the SAT method
Context
A more sophisticated method for performing the QAT is the SAT framework. In this method, the weights are quantized using the DoReFa nodes, and the activations are quantized using the CG-PACT node. DoReFa requires a TanhClamp node to work. Also a ScaleAdjust node ensure that the flowing gradient are properly scaled.
Key steps
-
test the SAT operators (cpu) -
test the SAT operators (cuda) -
implement the quantizers insertion and init routines -
benchmark the SAT QAT on the MNIST (cpu) -
benchmark the SAT QAT on the CIFAR-10 (cuda) -
benchmark the SAT QAT on the ImageNet (cuda)
RESULTS
TODO
Edited by Benjamin Halimi