Usage: reconstruct_phema.py [OPTIONS]

  Perform post-hoc EMA reconstruction.

  Examples:

  # Download raw snapshots for the pre-trained edm2-img512-xs model
  rclone copy --progress --http-url https://nvlabs-fi-cdn.nvidia.com/edm2 \
      :http:raw-snapshots/edm2-img512-xs/ raw-snapshots/edm2-img512-xs/

  # Reconstruct a new EMA profile with std=0.150
  python reconstruct_phema.py --indir=raw-snapshots/edm2-img512-xs \
      --outdir=out --outstd=0.150

  # Reconstruct a set of 31 EMA profiles, streaming over the input data 4 times
  python reconstruct_phema.py --indir=raw-snapshots/edm2-img512-xs \
      --outdir=out --outstd=0.010,0.015,...,0.250 --batch=8

  # Perform reconstruction for the latest snapshot of a given training run
  python reconstruct_phema.py --indir=training-runs/00000-edm2-img512-xs \
      --outdir=out --outstd=0.150

Options:
  --indir DIR      Directory containing the input pickles  [required]
  --inprefix STR   Filter inputs based on filename prefix
  --instd LIST     Filter inputs based on standard deviations
  --outdir DIR     Where to save the reconstructed network pickles  [required]
  --outprefix STR  Output filename prefix  [default: phema]
  --outstd LIST    List of desired relative standard deviations  [required]
  --outkimg KIMG   Training time of the snapshot to reconstruct  [x>=1]
  --skip           Skip output files that already exist
  --batch INT      Maximum simultaneous reconstructions  [default: 8; x>=1]
  --help           Show this message and exit.
