Skip to content

experiments.md

Latest commit

 

History

History
435 lines (364 loc) · 9.36 KB

experiments.md

File metadata and controls

435 lines (364 loc) · 9.36 KB

Experiments

This page lists the experiments that were run for the paper.

Setup

Define the following environment variables in a .env file:

BEAKER_BUDGET=...
BEAKER_WORKSPACE=...
BEAKER_CLUSTERS=...
WEKA_BUCKET=...

Default

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --no-load-weights

Data ablations

All minus Sentinel-1

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands S1

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands S1 \
    --no-load-weights

All minus Sentinel-2

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands S2_RGB S2_Red_Edge S2_NIR_10m S2_NIR_20m S2_SWIR NDVI

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands S2_RGB S2_Red_Edge S2_NIR_10m S2_NIR_20m S2_SWIR NDVI \
    --no-load-weights

All minus ERA5

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands ERA5

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands ERA5 \
    --no-load-weights

All minus TerraClimate

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands TC

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands TC \
    --no-load-weights

All minus SRTM

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands SRTM

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands SRTM \
    --no-load-weights

All minus location

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands location

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --excluded-bands location \
    --no-load-weights

Shape ablations

output_hw=16, num_timesteps=12

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 16 \
    --num-timesteps 12

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 16 \
    --num-timesteps 12 \
    --no-load-weights

output_hw=32, num_timesteps=6

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 32 \
    --num-timesteps 6

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 32 \
    --num-timesteps 6 \
    --no-load-weights

output_hw=32, num_timesteps=3

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 32 \
    --num-timesteps 3

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 32 \
    --num-timesteps 3 \
    --no-load-weights

output_hw=1, patch_size=1, num_timesteps=12

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 1 \
    --patch-size 1 \
    --num-timesteps 12

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 1 \
    --patch-size 1 \
    --num-timesteps 12 \
    --no-load-weights

output_hw=8, patch_size=8, num_timesteps=12

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 8 \
    --patch-size 8 \
    --num-timesteps 12

Random weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --output-hw 8 \
    --patch-size 8 \
    --num-timesteps 12 \
    --no-load-weights

Spatial splits

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --validation-state-regions Colorado "New Mexico" \
    --test-state-regions Texas Michigan

Pretrained weights

python -m experiment.beaker_finetune \
    --image-name ${USER}/lfmc \
    --priority high \
    --gpu-count 1 \
    --model-name tiny \
    --data-folder /presto/data/lfmc/training_tifs \
    --h5py-folder /presto/data/lfmc/h5pys \
    --h5pys-only \
    --validation-state-regions Colorado "New Mexico" \
    --test-state-regions Texas Michigan \
    --no-load-weights