Autoencoder and Latent Components

The repository contains a depth-band autoencoder, latent diffusion model code, configs, and dry-run tests. These are experimental building blocks: there is a maintained autoencoder training command, but no dedicated end-to-end latent diffusion launcher.

Autoencoder contract

src/depth_recon/configs/lat_space/ae_config.yaml currently configures:

• model type depth_band_ae;
• 50 input depth channels and 12 latent channels;
• band-only compression with spatial downsample 1;
• encoder widths [64, 96, 128];
• decoder widths [128, 96, 64].

For temperature training, ae.in_channels must match the active dataset's 50 target channels. Reconstruction uses the configured L1/L2 weights and valid-mask policy.

Train the autoencoder

/work/envs/depth/bin/python train_autoencoder.py \
  --data-config src/depth_recon/configs/px_space/training_super_config.yaml \
  --train-config src/depth_recon/configs/lat_space/training_config.yaml \
  --ae-config src/depth_recon/configs/lat_space/ae_config.yaml

The script unwraps the data section of the pixel super-config before building the active dataset.

Latent diffusion status

Latent configs and latent_cond_dif model code describe compression, latent conditioning, and decoding. Older shell launchers were removed during the pixel config cleanup. A new supported workflow would need a dedicated launcher that resolves the current scenario/data contract and validates an autoencoder checkpoint before training.

Until that exists, do not present latent diffusion as a reproducible production workflow. Existing components remain useful for architecture tests and controlled development.

Constraints

• Diffusion cannot recover depth information discarded by the autoencoder.
• Changing latent channels or normalization invalidates downstream checkpoints.
• Spatial downsampling would require explicit output alignment and mask behavior.
• Joint temperature/salinity latent semantics are not defined by the current autoencoder preset.