Install fastai from https://github.com/fastai/fastai
Initial up-scaling of images for UNet and downscaling for submission currently requires skimage which breaks the fastai environment. Not the best solution but I have a separate environment to use skimage.
todo ~
- config setup
- experiment with SGD
- Set up SSH access to local machine
- Rewrite ipynb notebook as a more organized Python package.
- Reflection padding or other ways to deal with image size
- cross fold validation
- early stopping
- Combat overfitting
- Dropout
- Additional data augmentation
- Shearing, zooming, distortions, etc.
- Define hyperparameters and write scripts to tune em
- Dilated conv
- Train model from scratch?
- Implement additional models (see below)
- Ensemble
- *** Architecture search
Resources:
- TernausNet: https://arxiv.org/abs/1801.05746
- Linknet: https://arxiv.org/pdf/1707.03718.pdf
- Pyramid Scene Parsing Network: https://arxiv.org/pdf/1612.01105.pdf
- Fully Convolutional Networks for Semantic Segmentation: https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf
- Segnet: https://arxiv.org/pdf/1511.00561.pdf
- DeepUNet: https://arxiv.org/pdf/1709.00201.pdf
- Dilated Convs for small objects [?]: https://arxiv.org/pdf/1709.00179.pdf
- CRFs: https://arxiv.org/pdf/1711.04483.pdf
- As post-processing?
- DeepLab: https://arxiv.org/pdf/1606.00915.pdf
https://www.kaggle.com/sudalairajkumar/winning-solutions-of-kaggle-competitions