PyTorch Implementation of Fully Convolutional Networks.
Project description
# pytorch-fcn
Fully Convolutional Networks implemented with PyTorch.
## TODO
- Support FCN16s and FCN8s.
## Accuracy
**FCN32s**
- `deconv=False`
- `train=SBDClassSeg(split='train')`
- `val=VOC2011(split='seg11val')`
- `batch_size=1`
- `MomentumSGD(lr=1e-10, momentum=0.99, weight_decay=0.0005)`
| epoch | iteration | valid/loss | valid/acc | valid/acc_cls | valid/mean_iu | valid/fwavacc |
|--------:|------------:|-------------:|------------:|----------------:|----------------:|----------------:|
| 9 | 76482 | 59656.847812 | 0.897753 | 0.780288 | 0.628707 | 0.844420 |
<img src="_static/fcn32s_voc2012_best_epoch9.jpg" width="40%" />
<img src="_static/fcn32s_voc2012_visualization_val.gif" width="40%" />
## Speed
It is ~4 times faster than [FCN implemented with Chainer](https://github.com/wkentaro/fcn),
measuring on Titan X Pascal.
```bash
% ./speedtest.py --gpu 0 --times 1000
==> Running on GPU: 0 to evaluate 1000 times
==> Testing FCN32s with Chainer
Elapsed time: 208.34 [s / 1000 evals]
Hz: 4.80 [hz]
==> Testing FCN32s with PyTorch
Elapsed time: 56.30 [s / 1000 evals]
Hz: 17.76 [hz]
```
Fully Convolutional Networks implemented with PyTorch.
## TODO
- Support FCN16s and FCN8s.
## Accuracy
**FCN32s**
- `deconv=False`
- `train=SBDClassSeg(split='train')`
- `val=VOC2011(split='seg11val')`
- `batch_size=1`
- `MomentumSGD(lr=1e-10, momentum=0.99, weight_decay=0.0005)`
| epoch | iteration | valid/loss | valid/acc | valid/acc_cls | valid/mean_iu | valid/fwavacc |
|--------:|------------:|-------------:|------------:|----------------:|----------------:|----------------:|
| 9 | 76482 | 59656.847812 | 0.897753 | 0.780288 | 0.628707 | 0.844420 |
<img src="_static/fcn32s_voc2012_best_epoch9.jpg" width="40%" />
<img src="_static/fcn32s_voc2012_visualization_val.gif" width="40%" />
## Speed
It is ~4 times faster than [FCN implemented with Chainer](https://github.com/wkentaro/fcn),
measuring on Titan X Pascal.
```bash
% ./speedtest.py --gpu 0 --times 1000
==> Running on GPU: 0 to evaluate 1000 times
==> Testing FCN32s with Chainer
Elapsed time: 208.34 [s / 1000 evals]
Hz: 4.80 [hz]
==> Testing FCN32s with PyTorch
Elapsed time: 56.30 [s / 1000 evals]
Hz: 17.76 [hz]
```
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
torchfcn-1.0.tar.gz
(8.6 kB
view hashes)