Support all the ONNX ops - Githubissues

ShigekiKarita / grain

autograd mir and CUDA library for dynamic neural networks in D.

https://shigekikarita.github.io/grain

Boost Software License 1.0

66 stars 7 forks source link

Support all the ONNX ops #1

Open ShigekiKarita opened 6 years ago

ShigekiKarita commented 6 years ago

Still no plan to impl onnx conversion but list here

https://github.com/onnx/onnx/blob/master/docs/Operators.md

Unary

Binary (NOTE: some ops are partially supported via CUDNN opTensor)

[x] Add
[ ] And
[x] Conv (NOTE CPU supports arbitrary shapes but CUDNN supports only 4/5 dim and limited shapes)
[ ] ConvTranspose
[x] Div
[ ] Equal
[x] Gemm
[ ] Greater
[ ] Less
[x] MatMul (NOTE: no batch matmul support yet)
[ ] Max
[ ] Min
[x] Mul
[ ] Not
[ ] Or
[x] Sub
[ ] Xor

Random
- [ ] Dropout
- [ ] RandomNormal
- [ ] RandomNormalLike
- [ ] RandomUniform
- [ ] RandomUniformLike

Reduction (NOTE: some ops are partially supported via CUDNN opTensor)

Topology

[ ] Concat
[ ] Flatten
[ ] Pad
[x] Reshape
[x] Shape
[x] Size
[ ] Slice
[ ] Split
[x] Squeeze
[ ] Tile
[ ] Transpose
[x] Unsqueeze

Learnable/Parametric

[ ] BatchNormalization
[ ] RNN
[ ] GRU
[ ] LSTM
[ ] PRelu (NOTE already wrapped in cudnn.activationForward/Backward)

what?

non onnx ops