Quantized Grads¶

This subproject can be used for emulating training in python with quantized gradients. The quantization is implemented as a fake-quantization.

Python¶

Layers¶

Right now this package supports the following base layers: - [X] Batchnorm2d - [X] Conv1d - [X] Conv2d - [X] Linear - [X] Sigmoid - [X] ReLU

Both quantization schemes can be applied for forward and backward independently. In my opinion the only valid combinations are the following three combinations.

1. Forward Fullresolution / Backward Fullresolution
2. Forward Quantized / Backward Fullresolution
3. Forward Quantized / Backward Quantized

Quantization¶

Right now we support following quantization schemes: - [X] FixedPoint - [X] FixedPointStochastic - [ ] BlockFloatingPoint - [ ] BlockFloatingPointStochastic