This repo contains some arraymancer demos and benchmarks.
These benchmarks share same model implementation and hyperparameters across frameworks.
This benchmark consists of a classification on 209 RGB 3x64x64 images, classifying them as cats or noncats. The model is a logistic unit. Simple batch gradient descent is used.
| Framework | Backend | Forward+Backward Pass Time |
|---|---|---|
| Arraymancer | OpenMP + MKL | 0.458ms |
| Torch7 | MKL | 0.686ms |
| Numpy | MKL | 0.723ms |
| Framework | Backend | Forward+Backward Pass Time |
|---|---|---|
| Arraymancer | CUDA | WIP |
| Torch7 | CUDA | 0.286ms |
As the above benchmark, consists of a classification on 209 RGB 3x64x64 images, classifying them as cats or noncats. The model is a deep fully connected neural network of layer sizes [209, 16, 8, 4, 1] (3 hidden layers + inputs/outputs layers). The activation function for the hidden layers is ReLU, the layer layer activation function is Sigmoid, and the loss is the binary cross entropy. Adam optimizer is used for batch gradient descent.
| Framework | Backend | Forward+Backward Pass Time |
|---|---|---|
| Arraymancer | OpenMP + MKL | 2.907ms |
| PyTorch | MKL | 6.797ms |
| Framework | Backend | Forward+Backward Pass Time |
|---|---|---|
| Arraymancer | CUDA | WIP |
| PyTorch | CUDA | 4.765ms |
- Intel(R) Core(TM) i7-3770K CPU @ 3.50GHz
- GeForce GTX 1080 Ti
- ArchLinux (kernel 4.12.13-1-ARCH, glibc 2.26)
- GCC 7.2.0
- MKL 2017.17.0.4.4
- OpenBLAS 0.2.20
- CUDA 8.0.61
- Nim 0.18.0 (head)