pytorch-fcn

Fully Convolutional Networks implemented with PyTorch.

Requirements

• pytorch >= 0.2.0
• torchvision >= 0.1.8
• fcn >= 6.1.4
• Pillow
• scipy
• tqdm

Installation

git clone https://github.com/wkentaro/pytorch-fcn.git
cd pytorch-fcn

conda install pytorch cuda80 torchvision -c soumith
pip install .

TODO

• Support FCN16s and FCN8s.

Accuracy

FCN32s

• train=SBDClassSeg(split='train')
• valid=VOC2011(split='seg11val')
• batch_size=1
• MomentumSGD(lr=1e-10, momentum=0.99, weight_decay=0.0005)

Implementation	epoch	iteration	Accuracy	Accuracy Class	Mean IU	FWAV Accuracy
Original	-	-	90.49	76.48	63.63	83.47
Ours	11	100000	90.63	72.78	63.07	83.30

Speed

PyTorch implementation is faster for static inputs and slower for dynamic ones than Chainer one at test time.
(In the previous performance, Chainer one was much slower, but it was fixed via wkentaro/fcn#90.)
Below are tested using a single GPU, Titan X (Pascal).

% ./speedtest.py
==> Benchmark: gpu=1, times=1000, dynamic_input=False
==> Testing FCN32s with Chainer
Elapsed time: 48.98 [s / 1000 evals]
Hz: 20.42 [hz]
==> Testing FCN32s with PyTorch
Elapsed time: 45.15 [s / 1000 evals]
Hz: 22.15 [hz]

% ./speedtest.py --dynamic-input
==> Benchmark: gpu=0, times=1000, dynamic_input=True
==> Testing FCN32s with Chainer
Elapsed time: 48.83 [s / 1000 evals]
Hz: 20.48 [hz]
==> Testing FCN32s with PyTorch
Elapsed time: 57.00 [s / 1000 evals]
Hz: 17.55 [hz]