5 ms·
Does that mean you're training a lower bit-rate(?) network or you are training a full network to 'know' it will eventually be running under quantization? I'd i
by UncleEntity 2y ago
Does that mean you're training a lower bit-rate(?) network or you are training a full network to 'know' it will eventually be running under quantization?
I'd imagine there's differences in the two approaches?
- cpldcpu 2y agoThe latter one. The network is trained in full precision (this is required for the gradient calculation), but the weights are nudged towards the quantized values.