High Granularity Quantization
HGQ is an gradient-based automatic bitwidth optimization and quantization-aware training algorithm for neural networks to be deployed on FPGAs, By laveraging gradients, it allows for bitwidth optimization at arbitrary granularity, up to per-weight and per-activation level.
Compare to the other heterogeneous quantization approach, like the QKeras counterpart, HGQ provides the following advantages:
High Granularity: HGQ supports per-weight and per-activation bitwidth optimization, or any other lower granularity.
Automatic Quantization: By setting a resource regularization term, HGQ could automatically optimize the bitwidth of all parameters during training. Pruning is performed naturally when a bitwidth is reduced to 0.
Bit-accurate conversion to hls4ml: You get exactly what you get from Keras models from hls4ml models. HGQ provides a bit-accurate conversion interface, proxy models, for bit-accurate conversion to hls4ml models. - still subject to machine float precision limitation.
Accurate Resource Estimation: BOPs estimated by HGQ is roughly #LUTs + 55#DSPs for actual (post place & route) FPGA resource consumption. This metric is available during training, and one can estimate the resource consumption of the final model in a very early stage.
Depending on the specific application, HGQ could achieve up to 20x resource reduction compared to the AutoQkeras approach, while maintaining the same accuracy. For some more challenging tasks, where the model is already under-fitted, HGQ could still improve the performance under the same on-board resource consumption. For more details, please refer to our paper here.