weather4cast2021_Stage1

3rd place solution for the Weather4cast 2021 Stage 1 Challenge

Dependencies

The code can be executed from a fresh environment using the provided list of requirements: conda env create -f environment.yml.

Inference

A script has been create to made predictions using a trained model on new data as per requirements detailed in competition: (#https://github.com/iarai/weather4cast#code-and-abstract-submission)

The model weights of the final submission for both core and transfer learning can be downloaded from https://drive.google.com/file/d/1mX4HMbf4QAW12DgAq1Bge1WT_5nAnz7N/view?usp=sharing

To run predictions on a test dataset use ('inference.py'). This should fine on a CPU machine

examples of usage:

- inference for Region R1

R=R1
INPUT_PATH=data
WEIGHTS=weights/Lomb_14.pth
OUT_PATH=.
python weather4cast/inference.py -d $INPUT_PATH -r $R -w $WEIGHTS -o $OUT_PATH -g 'cuda'

Train/evaluate a UNet

To replicate the training for the 3rd place solutions we recommend use of the training notebook (Training.ipynb).

Alternatively, a single script has also been provided (train.py).

In the actual competitions, the authors actually included an extra step to preprocess all the data and save them all on disk. This reduced the training time.