Speedtest Chainer vs. Pytorch implementation from fb research

[wkentaro]

Speed test (vs. PyTorch implementation)

Configuration:

• GTX 1080Ti (used also on monitor)

• CUDA 8.0.61

• CUDNN 5.1.10

• PyTorch 1.0.0.dev20181024 (with conda install pytorch-nightly cuda80 -c pytorch)

• Chainer 5.0.0, Cupy 5.0.0 (with pip install chainer cupy-cuda80)

• CPU -> GPU communication of input.

• BBox prediction and suppression.

• Mask prediction for remaining bboxes (nms thresh: 0.5, score thresh: 0.7).

• GPU -> CPU communication of output.

# Chainer implementation (this repo)
% pwd
/home/wkentaro/chainer-mask-rcnn/examples/coco
% ./speedtest.py --gpu 0 --times 10
==> Benchmark: gpu=0, times=10
==> Image file: https://raw.githubusercontent.com/facebookresearch/Detectron/master/demo/33823288584_1d21cf0a26_k.jpg
==> Testing Mask R-CNN RestNet50-C4 with Chainer
Elapsed time: 3.09 [s / 10 evals]
Hz: 3.24 [hz]

# PyTorch implementation (https://github.com/facebookresearch/maskrcnn-benchmark)
% git clone https://github.com/wkentaro/maskrcnn-benchmark.git -b speedtest_r50_c4  # then install it
% pwd
/home/wkentaro/maskrcnn-benchmark/demo
% ./speedtest.py --gpu 0 --times 10
==> Benchmark: gpu=0, times=10
==> Image file: https://raw.githubusercontent.com/facebookresearch/Detectron/master/demo/33823288584_1d21cf0a26_k.jpg
==> Testing Mask R-CNN ResNet-C4 with PyTorch
Elapsed time: 3.44 [s / 10 evals]
Hz: 2.91 [hz]