An Enhanced feedforward deep nerual network forward/inference library. Protobuf is used for model io, since protobuf is good at flexibility and model compatibility. The system has the same function with Net, Support both float and 8bits quantized model. Refer https://github.com/robin1001/net for performance details.
One main concern is the model size, since we want smaller model size, epecially for embeding device. Here is the comparison of Net/Xnet model size.
| Model | Net(K) | XNet(K) |
|---|---|---|
| Float | 2616.12 | 2616.25 |
| 8bits | 657.14 | 1072.65 |
we can see for float model, the model size is close. However, for 8bits model, XNet is much larger than Net. Because protobuf use varint encoding for integers(8/16/32), it's useful for compression, especially for int32, but not for int8.