Model Settings

This page maps key configuration flags to their runtime behavior in code.

Primary config files used in current EO setup:
- configs/px_space/data_ostia_argo_disk.yaml (OSTIA + Argo + GLORYS GeoTIFF disk preset)
- configs/px_space/model_config.yaml
- configs/px_space/model_config_ambient.yaml
- configs/px_space/training_config.yaml

Latent-space config set:
- configs/lat_space/model_config.yaml
- configs/lat_space/training_config.yaml
- configs/lat_space/data_config.yaml
- configs/lat_space/ae_config.yaml

See Autoencoder + Latent Diffusion for latent architecture and training workflow.

Major Settings

Conditioning channels

Config (model_config_*):
- model.generated_channels
- model.condition_channels
- model.condition_mask_channels
- model.condition_include_eo
- model.condition_use_valid_mask

Runtime effect:
- controls how condition = [eo?, x, x_valid_mask?] is assembled
- channel count is validated against expected condition_channels

Diffusion target parameterization

Config:
- model.parameterization: epsilon or x0

Runtime effect:
- defines target in diffusion loss and sampler conversions
- current EO config uses x0

Masked loss

Config:
- model.mask_loss_with_valid_pixels

Runtime effect:
- if enabled, loss uses the task-valid support instead of the old missing-pixel mask
- standard mode: y_valid_mask over the full y target
- ambient mode: x_valid_mask ∩ y_valid_mask over the x target
- mask alignment preserves per-band semantics (B x C x H x W) unless a single shared mask channel is explicitly used

Inference output composition

Runtime effect:
- direct y prediction keeps the generated field and masks invalid y_valid_mask support to NaN
- ambient x completion leaves known-pixel enforcement to clamp_known_pixels during sampling
- both modes then mask invalid y_valid_mask support to NaN

Known-pixel clamping during sampling

Config:
- model.clamp_known_pixels

Runtime effect:
- if enabled and known masks/values are available, known pixels are overwritten each reverse step
- useful for inpainting-style stability

Illustration:

Coordinate/date FiLM conditioning

Config:
- data: dataset.output.return_coords
- model:
- coord_conditioning.enabled
- coord_conditioning.encoding
- coord_conditioning.include_date
- coord_conditioning.date_encoding
- coord_conditioning.embed_dim

Runtime effect:
- creates a coordinate/date embedding and injects it via FiLM in ConvNeXt blocks
- details: Data + Coordinate Injection

Training and Optimization Settings

Noise schedule and diffusion steps

Config (training.noise):
- num_timesteps
- schedule: linear, cosine, quadratic, sigmoid
- beta_start, beta_end

Validation sampling mode

Config (training.validation_sampling):
- sampler: ddpm or ddim
- ddim_num_timesteps, ddim_eta
- log_intermediates

Runtime effect:
- training loss still uses forward noising objective
- full reverse sampling diagnostics use chosen validation sampler

Learning-rate warmup and plateau scheduler

Config (scheduler):
- warmup.enabled, warmup.steps, warmup.start_ratio
- reduce_on_plateau.enabled
- reduce_on_plateau.monitor, mode, factor, patience, threshold, cooldown

Runtime effect:
- warmup is applied per optimizer step in optimizer_step
- plateau scheduler is applied on epoch-level monitored metric

Trainer/Runtime Controls

Config (trainer):
- hardware/precision: accelerator, devices, optional num_gpus, precision
- logging/checkpoint cadence: log_every_n_steps, ckpt_monitor, lr_logging_interval
- validation load: val_batches_per_epoch or limit_val_batches
- stability knobs: gradient_clip_val, warning suppressions

Dataloader Settings

Config (dataloader):
- batch_size, val_batch_size
- num_workers, val_num_workers
- shuffle, val_shuffle
- pin_memory, persistent_workers, prefetch_factor

Runtime notes:
- prefetch_factor is only applied when num_workers > 0
- validation shuffle defaults to true in DataModule unless explicitly changed

Logging Settings (W&B)

Config (wandb):
- project/entity/run naming
- model logging policy
- watch toggles (watch_gradients, watch_parameters)
- scalar/image logging intervals

Runtime notes:
- watch mode is resolved from explicit gradient/parameter toggles
- config files used in the run are uploaded to W&B run files when possible

FUll settings documentation

This section contains the complete key-by-key configuration reference previously documented on the separate Configs page.

Dataset Configs (configs/px_space/data_ostia_argo_disk.yaml)

Dataset settings are grouped by intent (core, validity, degradation, conditioning, augmentation, output, runtime).
Defaults below refer to configs/px_space/data_ostia_argo_disk.yaml unless noted.

Config key	Default value	Explanation
dataset.core.dataset_variant	"ostia_argo_disk"	Selects dataset in train.py ("eo_4band", "ostia", or "ostia_argo_disk").
dataset.core.dataloader_type	"light"	The current training runner supports only "light" loading.
dataset.core.manifest_csv_path	"/work/data/depth_prod/ostia_argo_tiff_index.csv"	Manifest CSV used by OstiaArgoTiffDataset; paths inside it are resolved relative to the manifest location.
dataset.output.return_info	false	Returns per-sample metadata under batch["info"].
dataset.output.return_coords	true	Returns patch-center coordinates under batch["coords"].
dataset.runtime.random_seed	7	Seed used for deterministic split and random dataset sampling behavior.
split.val_fraction	0.2	Fraction of dataset reserved for validation.

configs/px_space/model_config.yaml

Config key	Default value	Explanation
model.model_type	"cond_px_dif"	Model type ("cond_px_dif" for pixel diffusion, "latent_cond_dif" for latent diffusion with AE bridge).
model.resume_checkpoint	false	false/null starts from scratch; checkpoint path resumes training.
model.load_checkpoint	false	false/null disables warm start; checkpoint path loads model state_dict only (no Lightning optimizer/trainer resume).
model.generated_channels	50	Number of predicted GLORYS depth channels.
model.condition_channels	52	Condition channel count: OSTIA EO (1) + corrupted Argo stack (50) + collapsed x_valid_mask (1).
model.condition_mask_channels	1	Number of x_valid_mask condition channels.
model.condition_include_eo	true	Includes batch["eo"] as condition input.
model.condition_use_valid_mask	true	Includes x_valid_mask in condition input.
model.clamp_known_pixels	false	Clamps known pixels each reverse step for inpainting-style stability.
model.mask_loss_with_valid_pixels	true	Computes loss on the task-valid supervision mask (y_valid_mask in standard mode, x_valid_mask ∩ y_valid_mask in ambient mode).
model.parameterization	"x0"	Diffusion training target ("epsilon" or "x0").
model.log_intermediates	true	Default validation intermediate logging behavior.
model.ambient_occlusion.enabled	false	Enables ambient-diffusion style occlusion objective (further-corrupt input, supervise x on x_valid_mask ∩ y_valid_mask).
model.ambient_occlusion.further_drop_prob	0.1	Additional drop probability delta applied on already observed pixels during training.
model.ambient_occlusion.apply_to_noisy_branch	true	Applies the further mask to the noisy target branch in p_loss (~A x_t).
model.ambient_occlusion.shared_spatial_mask	true	Uses one spatial further-mask per sample and shares it across channels.
model.ambient_occlusion.min_kept_observed_pixels	1	Guarantees a minimum number of observed pixels kept after further corruption.
model.ambient_occlusion.require_x0_parameterization	true	Enforces model.parameterization == "x0" when ambient objective is enabled.
model.post_process.gaussian_blur.enabled	false	Enables final denormalized Gaussian blur post-process.
model.post_process.gaussian_blur.sigma	0.5	Gaussian blur sigma in pixels.
model.post_process.gaussian_blur.kernel_size	3	Blur kernel size; even values are adjusted to odd.
model.coord_conditioning.enabled	true	Enables coordinate conditioning with FiLM.
model.coord_conditioning.encoding	"unit_sphere"	Coordinate encoding type ("unit_sphere", "sincos", "raw").
model.coord_conditioning.include_date	true	Includes date encoding with coordinates.
model.coord_conditioning.date_encoding	"day_of_year_sincos"	Date encoding mode (day-of-year sin/cos, denominator 365).
model.coord_conditioning.embed_dim	null	FiLM embedding dimension; defaults to unet.dim when null.
model.unet.dim	64	Base channel width of U-Net denoiser.
model.unet.dim_mults	[1, 2, 4, 8]	Per-stage width multipliers; controls depth/width scaling.
model.unet.with_time_emb	true	Enables timestep embeddings in denoiser.
model.unet.output_mean_scale	false	Optional output mean correction for diffusion variants.
model.unet.residual	false	If enabled, predicts residual added to input.

Detailed objective math, implementation mapping, visualization, and citation: Ambient Occlusion Objective.

configs/px_space/training_config.yaml

Config key	Default value	Explanation
training.lr	1.0e-4	Optimizer learning rate.
training.batch_size	4	Informational training batch size (dataloader section is source of truth).
training.noise.num_timesteps	1000	Number of diffusion timesteps.
training.noise.schedule	"cosine"	Noise schedule: linear, cosine, quadratic, sigmoid.
training.noise.beta_start	1.0e-4	First-step noise level (must be positive and below beta_end).
training.noise.beta_end	2.0e-2	Final-step noise level (must be below 1 and above beta_start).
training.validation_sampling.sampler	"ddim"	Validation sampler (ddpm full chain, ddim faster).
training.validation_sampling.ddim_num_timesteps	100	DDIM steps when sampler="ddim".
training.validation_sampling.ddim_eta	0.0	DDIM eta; 0.0 is deterministic DDIM.
training.validation_sampling.log_intermediates	false	Captures/logs denoising intermediate images in validation.
training.validation_sampling.skip_full_reconstruction_in_sanity_check	true	Skips expensive full reconstruction during Lightning sanity checks when true.
training.validation_sampling.max_full_reconstruction_samples	2	Max first-batch val samples used for full reconstruction pass.
trainer.max_epochs	1500	Maximum training epochs.
trainer.accelerator	"auto"	Lightning accelerator backend selection.
trainer.devices	"auto"	Device selection (auto, int, list).
trainer.num_gpus	null	Legacy explicit GPU count override; null leaves accelerator/devices in control.
trainer.strategy	"auto"	Distributed strategy selection.
trainer.precision	"16-mixed"	Mixed precision mode.
trainer.matmul_precision	"high"	torch.set_float32_matmul_precision mode.
trainer.suppress_accumulate_grad_stream_mismatch_warning	true	Suppresses PyTorch stream mismatch warning noise.
trainer.suppress_lightning_pytree_warning	true	Suppresses Lightning LeafSpec deprecation warning noise.
trainer.ckpt_monitor	"val/loss_ckpt"	Metric monitored for best-checkpoint saving.
trainer.lr_logging_interval	"step"	Learning-rate logging cadence (step or epoch).
trainer.log_every_n_steps	25	Trainer logging interval in steps.
trainer.num_sanity_val_steps	1	Number of startup sanity-validation steps.
trainer.limit_val_batches	4	Number/fraction of validation batches per epoch.
trainer.enable_model_summary	true	Enables Lightning model summary printout.
trainer.gradient_clip_val	1.0	Gradient clipping threshold (0.0 disables).
wandb.project	"DepthDif_Simon"	W&B project name.
wandb.entity	"esa-phi-lab"	W&B entity/team (null uses default account).
wandb.run_name	"ostia_argo_disk_px"	Explicit run name.
wandb.log_model	"false"	W&B model artifact logging policy.
wandb.verbose	true	Enables extra metric/image logging.
wandb.watch_gradients	false	Enables gradient history logging via wandb.watch.
wandb.watch_parameters	false	Enables parameter history logging via wandb.watch.
wandb.watch_log_freq	100	wandb.watch logging frequency in steps.
wandb.watch_log_graph	false	Logs computation graph when watch is enabled.
wandb.log_stats_every_n_steps	200	Step interval for scalar debug stats.
wandb.log_images_every_n_steps	200	Step interval for validation preview images.
dataloader.batch_size	4	Training dataloader batch size.
dataloader.val_batch_size	2	Validation batch size (falls back to batch_size if omitted).
dataloader.num_workers	4	Number of training dataloader workers.
dataloader.val_num_workers	0	Validation workers (0 avoids h5netcdf sanity-check instability).
dataloader.persistent_workers	true	Keeps train workers alive across epochs when true.
dataloader.val_persistent_workers	false	Validation worker persistence (when val_num_workers > 0).
dataloader.prefetch_factor	2	Prefetched batches per worker (only used when workers > 0).
dataloader.shuffle	true	Shuffles training dataset each epoch.
dataloader.val_shuffle	false	Shuffles validation set (often used with limited val batches).
dataloader.pin_memory	true	Enables pinned host memory for faster H2D transfer.
scheduler.warmup.enabled	true	Enables linear warmup before plateau scheduling.
scheduler.warmup.steps	2000	Warmup step count to ramp LR from start_ratio to base LR.
scheduler.warmup.start_ratio	0.2	Initial warmup LR as ratio of training.lr.
scheduler.reduce_on_plateau.enabled	true	Enables ReduceLROnPlateau.
scheduler.reduce_on_plateau.monitor	"val/loss_ckpt"	Metric monitored for LR reduction.
scheduler.reduce_on_plateau.mode	"min"	Plateau mode (min or max).
scheduler.reduce_on_plateau.factor	0.5	Multiplicative LR decay factor on plateau.
scheduler.reduce_on_plateau.patience	20	Validation epochs with no improvement before reducing LR.
scheduler.reduce_on_plateau.threshold	1.0e-4	Minimum significant metric change.
scheduler.reduce_on_plateau.threshold_mode	"rel"	Threshold mode (rel or abs).
scheduler.reduce_on_plateau.cooldown	0	Epoch cooldown after LR reduction.
scheduler.reduce_on_plateau.min_lr	1.0e-6	Lower bound for LR.
scheduler.reduce_on_plateau.eps	1.0e-8	Minimum effective LR change.