NeurIPS 2020 SEVIR
Code for paper: SEVIR : A Storm Event Imagery Dataset for Deep Learning Applications in Radar and Satellite Meteorology
Requirements
To test pretrained models and train on single GPU, this requires
tensorflow 2.1.0
or higherpandas
matplotlib
pytorch 1.4.0
or higher to calculate the LPIPS metric. See Perceptual Similarity Metric and Dataset
Distributed (multi-GPU) training of these models requires
Horovod 0.19.0
or higher for distributed training. See Horovod
To visualize results with statelines as is done in the paper, a geospatial plotting library is required. We recommend either of the following:
basemap
cartopy
To run the rainymotion
benchmark, you'll also need to install this module. See https://rainymotion.readthedocs.io/en/latest/
Downloading pretrained models
To download the models trained in the paper, run the following
cd models/
python download_models.py
See the notebooks
directory for how to apply these models to some sample test data.
Downloading SEVIR
Download information and additional resources for SEVIR data are available at https://registry.opendata.aws/sevir/.
To download, install AWS CLI, and download all of SEVIR (~1TB) to your current directory run
aws s3 sync --no-sign-request s3://sevir .
Extracting training/testing datasets
The models implemented in the paper are implemented on training data collected prior to June 1, 2019, and testing data collected after June 1, 2019. These datasets can be extrated from SEVIR by running the following scripts (one for nowcasting, and one for synrad). Depending on your CPU and speed of your filesystem, these scripts may take several hours to run.
cd src/data
# Generates nowcast training & testing datasets
python make_nowcast_dataset.py --sevir_data ../../data/sevir --sevir_catalog ../../data/CATALOG.csv --output_location ../../data/interim/
# Generate synrad training & testing datasets
python make_synrad_dataset.py --sevir_data ../../data/sevir --sevir_catalog ../../data/CATALOG.csv --output_location ../../data/interim/
Testing pretrained models
Pretrained models used in the paper are located under models/
. To run test metrics on these datasets, run the test_*.py
scripts and point to the pretrained model, and the test dataset. To test, we recommend setting num_test
to a small number, and increasing thereafter (not specifying will use all test data). This shows an example
# Test a trained synrad model
python test_synrad.py --num_test 1000 --model models/synrad_mse.h5 --test_data data/interim/synrad_testing.h5 -output test_output.csv
Also check out the examples in notebooks/
for how to run pretrained models and visualize results.
Model training
This section describes how to train the nowcast
and synthetic weather radar (synrad
) models yourself. Models discussed in the paper were trained using distributed training over 8 NVIDIA Volta V100 GPUs with 32GB of memory. However the code in this repo is setup to train on a single GPU.
The training datasets are pretty large, and running on the full dataset requires a significant amount of RAM. We suggest that you first test the model with --num_train
set to a low number to start, and increase this to the limits of your system. Training with all the data may require writing your own generator that batches the data so that it fits in memory.
nowcast
Training To train the nowcast
model, make sure the nowcast_training.h5
file is created using the previous steps. Below we set num_train
to be only 1024, but this should be increased for better results. Results described in the paper were generated with num_train = 44,760
. When training the model with the mse
loss, the largest batch size possible is 32 and for all other cases, a maximum batch size of 4 must be used. Larger batch sizes will result in out-of-memory errors on the GPU. There are four choices of loss functions configured:
MSE Loss:
python train_nowcast.py --num_train 1024 --nepochs 25 --batch_size 32 --loss_fn mse --logdir logs/mse_`date +yymmddHHMMSS`
Style and Content Loss:
python train_nowcast.py --num_train 1024 --nepochs 25 --batch_size 4 --loss_fn vgg --logdir logs/mse_`date +yymmddHHMMSS`
MSE + Style and Content Loss:
python train_nowcast.py --num_train 1024 --nepochs 25 --batch_size 4 --loss_fn mse+vgg --logdir logs/mse_`date +yymmddHHMMSS`
Conditional GAN Loss:
python train_nowcast.py --num_train 1024 --nepochs 25 --batch_size 32 --loss_fn cgan --logdir logs/mse_`date +yymmddHHMMSS`
Each of these will write several files into the date-stamped directory in logs/
, including tracking of metrics, and a model saved after each epoch. Run python train_nowcast.py -h
for additional input parameters that can be specified.
synrad
Training To train synrad
, make sure the synrad_training.h5
file is created using the previous step above. Below we set num_train
to be only 10,000, but this should be increased for better results. There are three choices of loss functions configured:
MSE Loss:
python train_synrad.py --num_train 10000 --nepochs 100 --loss_fn mse --loss_weights 1.0 --logdir logs/mse_`date +yymmddHHMMSS`
MSE+Content Loss:
python train_synrad.py --num_train 10000 --nepochs 100 --loss_fn mse+vgg --loss_weights 1.0 1.0 --logdir logs/mse_vgg_`date +yymmddHHMMSS`
cGAN + MAE Loss:
python train_synrad.py --num_train 10000 --nepochs 100 --loss_fn gan+mae --loss_weights 1.0 --logdir logs/gan_mae_`date +yymmddHHMMSS`
Each of these will write several files into the date-stamped directory in logs/
, including tracking of metrics, and a model saved after each epoch.
Analyzing results
The notebooks under notebooks
contain code for anaylzing the results of training, and for visualizing the results on sample test cases.