model quantization experiments about clair

[aquaskyline]

Hi Ram, it's a nice idea but we haven't tried out ourselves. We tried using FP16 instead of FP32 but our model suffered from either non-converging or reduced accuracy problem. Usually, quantization or reduced precision work when the capacity of the network is not fully utilized. From our experiments, the network architecture of Clair has little redundancy. But without a fully interpretable model, we are not 100% sure on that. Please do let us know any exciting results from your side.

[ramcn]

Thanks for the insights. I will keep you posted about my experiments.