Awesome
Pytorch -- Multitemporal Land Cover Classification Network
A (yet barebone) Pytorch port of Rußwurm & Körner (2018) Tensorflow implementation
Please consider citing
Rußwurm M., Körner M. (2018). Multi-Temporal Land Cover Classification with
Sequential Recurrent Encoders. ISPRS International Journal of Geo-Information, 2018.
if you use this repository
Activations while encoding sequence:
<p align="center"> <img src="doc/lstm.gif" width="500" /> </p>Dependencies
Python dependencies
pip install numpy
pip install pandas>=0.23.4
pip install visdom==0.1.8.4
pip install rasterio>=1.0.2
# install pytorch 0.4.1 (https://pytorch.org/)
pip3 install torch torchvision
Download dataset to src/data and model checkpoint to src/checkpoints
bash download.sh
Train 10 epochs (batchsize 16, dataloader-workers 16) with initialized weights
from checkpoint file checkpoints/model_00.pth
# add src folder to python path
export PYTHONPATH="$PATHONPATH:$PWD/src"
# train
python src/train.py data -b 16 -w 16 -s checkpoints/model_00.pth
Visdom Support
Start visdom server in screen, tmux or other terminal with $ visdom
and open http://localhost:8097 in the browser while training
Comparison to Tensorflow implementation
not yet implemented features compared to the Tensorflow version
- ConvGRU integration in
train.py - bidirectional RNN loop
- masking of the background class
Source of ConvLSTM and ConvGRU implementations
- ConvLSTM cell implementation from
https://github.com/ndrplz/ConvLSTM_pytorch - (not used yet) ConvGRU cell implementation from
https://github.com/bionick87/ConvGRUCell-pytorch