SharpNet is an Open-source Deep Learning library written in C# 7.0.
- Residual Networks v1, v2 and WideResNet
- Dropout / BatchNorm / Conv / Pooling / Dense / Concatenate / Shortcut layers
- Elu / Relu / Sigmoid / Softmax activations
- SGD & Adam optimizers
- Image Data Augmentation (with Cutout/CutMix/Mixup)
- Ensemble Learning
It can be run both on GPU (using NVIDIA cuDNN) and on the CPU (using MKL Blas).
It is targeted to make a good use of the GPU (even if it is not currently as fast as MxNet) :
- on ResNet18 v1, it is between 1.5x (batch size = 128) and 3x time (batch size = 32) faster then TensorFlow
Add ResNet v2 support=> DONE Add Dense Network support=> DONE Cutout=> DONE Add CutMix=> DONE Add Mixup=> DONE
- Add multi GPU support
- Add RNN / LSTM support
Improve memory efficiency for gradients=> DONE Add Wide ResNet / Wide DenseNet support=> DONE Improve Image Data Augmentation (with rotation)=> DONE