This folder contains example training scripts of AGENIUM SPACE tiny unet model on ALCD Cloud DB. It contains training script, the ALCD DB reprocessed in RGB a docker recipe.
The folder contains a docker recipe and a make file to build an run it, got check the Readme in DOCKER folder. It is assumed you have a NVIDIA GPU on your computer. To build the docker:
cd DOCKER
make build
To run the docker :
cd DOCKER
make bash
You can find the DB in DATA. It is a tiled and rgb format of the ALCD Cloud DB available online here.
This section assume you are using the provided docker.
For training the tiny unet 100k model use :
python ./SCRIPTS/train.py run --model ags_tiny_unet_100k --data_path ./DATA/
For training the tiny unet 50k model use :
python ./SCRIPTS/train.py run --model ags_tiny_unet_50k --data_path ./DATA/
You can access the help info using
python ./SCRIPTS/train.py run -- --help
To use 2 gpus:
# using torchrun
torchrun --nproc_per_node=2 ./SCRIPTS/train.py run --model ags_tiny_unet_100k --data_path ./DATA/ --backend="nccl"
python ./SCRIPTS/train.py run --model ags_tiny_unet_50k --data_path ./DATA/
The table below contains the score results for the the models
Model | Train Score (mean F1-score) | Valid Score (F1-score) |
---|---|---|
tiny_unet_50k | 0.94 - 0.78 | 0.94 - 0.77 |
tiny_unet_100k | 0.94 - 0.79 | 0.84 - 0.78 |