Conditional Flow Matching (Lipman et al., 2023) for InSAR
Train a simple conditional flow matching Unet model on MNIST. MNIST data is streamed from FastAI's AWS open data S3 bucket with webdataset.
Training is setup for a SLURM HPC cluster with a Nvidia MIG parition, if you are using a standard Nvidia gpu you can change the following line in launch_cfm_mnist.py:
slurm_additional_parameters={"gres": "gpu:3g.40gb:1"}
To run training for 10 epochs:
uv run launch_cfm_mnist.py train
To run sampling:
uv run launch_cfm_mnist.py sample 7 (this will generate a sample conditioned on MNIST class 7)