Code for a winning model (3 out of 419) in a Dstl Satellite Imagery Feature Detection challenge
Name: Vladimir Iglovikov
LinkedIn: https://www.linkedin.com/in/iglovikov/
Location: San-Francisco, United States
Name: Sergey Mushinskiy
LinkedIn: https://www.linkedin.com/in/sergeymushinskiy/
Location: Angarsk, Russia
Competition: Dstl Satellite Imagery Feature Detection
If you find this code useful for your publications, please consider citing
@article{DBLP:journals/corr/IglovikovMO17,
author = {Vladimir Iglovikov and
Sergey Mushinskiy and
Vladimir Osin},
title = {Satellite Imagery Feature Detection using Deep Convolutional Neural
Network: A Kaggle Competition},
volume = {abs/1706.06169},
year = {2017},
archivePrefix = {arXiv},
eprint = {1706.06169},
}
To train final models you will need the following:
pip install -r requirements.txt
data / theree_band / *
/ sixteen_band / *
grid_sizes.csv
train_wkt_v4.csv
src / *.py
python get_3_band_shapes.py
cache_train.py
Each class in our solution has separate neural network, so it requires running of several distinct models one by one (or in parallel if there are enough computing resources)
python unet_buidings.py
python unet_structures.py
python unet_road.py
python unet_track.py
python unet_trees.py
python unet_crops.py
For water predictions we used different method and it can be created by running:
python fast_water.py
python slow_water.py
After training finishes (it may require quite a long time depending on hardware used, in our case it was about 7 hours for each stage (50 epochs)) trained weights and model architectures are saved in cache
directory and can be used by prediction scripts (see the next section).
To create predictions run every make_prediction_cropped_*.py file in src
dir. It could take considerable amount of time to generate all predictions as there are a lot of data in test and we use separate models for each class and use test time augmentation and cropping for the best model performance. On Titan X GPU each class took about 5 hours to get predictions.
python make_prediction_cropped_buildings.py
python make_prediction_cropped_structures.py
python make_prediction_cropped_track.py
python make_prediction_cropped_road.py
python make_prediction_cropped_trees.py
python make_prediction_cropped_crops.py
When all predictions are done they should be merged in a single file for submit:
python merge_predictions.py
python merge_predictions.py
The previous step will create file joined.csv
that just merges predictions per class into the unified format.python post_processing.py joined.csv
that will perform some cleaning of the overlapping classes (remove predictions of the slow water from fast water, all other predictions from buildings, etc)
Please, keep in mind that this isn't a production ready code but a very specific solution for the particular competition created in short time frame and with a lot of other constrains (limited training data, scarce computing resources and a small number of attents to check for improvements).
So, there are a lot of hardcoded magic numbers and strings and there may be some inconsistensies and differences between different models. Sometimes, it was indentended to get more accurate predictions and there wasn't enough resources to check if changes improve score for other classes after they were introduced for some of them. Sometimes, it just slipped from our attention.
Also, inherent stochasticity of neural networks training on many different levels (random initialization of weights, random cropping of patches into minibatch and so on) makes it impossible to reproduce exact submission from scratch. We went extra mile and reimplemented solution and training procedure from scratch as much as possible in the last two weeks after competition final. We've got up to 20% extra performance for some classes with abundant training data like buildings, tracks and so on. However, some classes proven more difficult to reliably reproduce because of lack of training data and small amount of time. Such classes show high variance of results between epochs. For competition we used our best performing combinations of epoch/model for those classes, which may not be exactly the same as trained for fixed number of epochs (as in this particular code). However, we believe that our model is equally capable to segment any classes, given enough data and/or clear definitions what exactly consists of each class (it wasn't clear how segmentation was performed in the first place for some classes, like road/tracks).