Bidirectional Variational Autoencoder with Inverted Autoregressive Flows:

1. Requirements

The model is built in Python 3.9 using Tensorflow 2.7.0. Use the following command to install the requirements:

pip install -r requirements.txt

2. Dataset

Dataset for CESM can be downloaded at: https://sdrbench.github.io/

Sample download instruction:

wget https://97235036-3749-11e7-bcdc-22000b9a448b.e.globus.org/ds131.2/Data-Reduction-Repo/raw-data/CESM-ATM/SDRBENCH-CESM-ATM-26x1800x3600.tar.gz
tar -xvzf SDRBENCH-CESM-ATM-26x1800x3600.tar.gz -C cesm_data_2

3. Running the training scripts

MNIST

python main.py --use_se --num_initial_channel 16 --num_process_blocks 2 \
    --num_preprocess_cells 1 --num_postprocess_cells 1 --num_cell_per_group_enc 1 \
    --num_cell_per_group_dec 1 --num_groups_per_scale 1 --num_scales 2 --batch_size 256 \
    --learning_rate 0.001 --learning_rate_min 0.000005 --epochs 100 \
    --model_path ./model_output/mnist_iaf

CESM-Cloud

python main.py --use_se --num_initial_channel 16 --num_process_blocks 3 \
    --num_preprocess_cells 1 --num_postprocess_cells 1 --num_cell_per_group_enc 1 \
    --num_cell_per_group_dec 1 --num_groups_per_scale 1 --num_scales 2 --batch_size 128 \
    --learning_rate 0.001 --learning_rate_min 0.000005 --epochs 200 \
    --model_path ./model_output/cesm_iaf --data_path ./data --dataset cesm \
    --tile_size 64

4. Evaluating:

GitHub URL

https://github.com/hieutrungle

License

This program is created by Hieu Le