Train: MultiSet + Muses Val V2 + Sky-SDF 16ch `78078471`

Standalone launch config based on “Train: MultiSet + Muses Val V2” (config 81). It preserves the V2 multi-source RDS + VKITTI2 training and MUSES sparse-depth validation setup, while enabling the new 16-channel single-pass sky-SDF diffusion pathway. Semantic sky masks are wired for RDS (dataset 7), VKITTI2 (dataset 9), and MUSES validation (split 75); heuristic labels remain disabled. The config uses deterministic SDF VAE encoding, Gaussian segmentation noise, 1/sqrt(3) target initialization, task skip adapters, the latent velocity objective, delayed direct SDF/BCE/Dice supervision, threshold validation, and non-destructive inference exports. Latent standardization is intentionally disabled with identity moments because no completed codec-gate run is available; replace the four channel moments and enable it after preflight calibration. Canonical required RGB/depth placeholders remain intentionally omitted because the inherited V2 template uses explicit multi-source dataset.path bindings.

Output

Launch Output

Waiting for SLURM job to start...

Datasets

Dataset clear_root_rds

Dataset real-drive-sim

Type rgb

Split full

Path /cluster/work/igp_psr/drothenpiele/data/rds/rgb.zip/

Dataset depth_root_rds

Dataset real-drive-sim

Type depth

Split full

Path /cluster/work/igp_psr/drothenpiele/data/rds/depth.zip/

Dataset intrinsics_root_rds

Dataset real-drive-sim

Type intrinsics

Split full

Path /cluster/work/igp_psr/drothenpiele/data/rds/calibration.zip/

Dataset sky_mask_root_rds

Dataset real-drive-sim

Type semantic_segmentation

Split full

Path /cluster/work/igp_psr/drothenpiele/data/rds/semantic_segmentation_2d.zip/

Dataset train_hazy_root_rds

Dataset real-drive-sim (gloomy_final_2)

Type rgb

Split full

Path /cluster/work/igp_psr/drothenpiele/data/rds/gloomy_final_2/foggy_rgb.zip/

Dataset val_hazy_root_rds

Dataset real-drive-sim (gloomy_fixed)

Type rgb

Split val

Path /cluster/work/igp_psr/drothenpiele/data/rds/gloomy_fixed/foggy_rgb.zip/

Dataset clear_root_vkitti2

Dataset VKITTI2

Type rgb

Split full

Path /cluster/work/igp_psr/drothenpiele/data/vkitti_2.0.3_rgb.zip

Dataset depth_root_vkitti2

Dataset VKITTI2

Type depth

Split full

Path /cluster/work/igp_psr/drothenpiele/data/vkitti_2.0.3_depth.zip

Dataset intrinsics_root_vkitti2

Dataset VKITTI2

Type intrinsics

Split full

Path /cluster/work/igp_psr/drothenpiele/data/vkitti_2.0.3_textgt.zip/

Dataset sky_mask_root_vkitti2

Dataset VKITTI2

Type semantic_segmentation

Split full

Path /cluster/work/igp_psr/drothenpiele/data/vkitti_2.0.3_classSegmentation.zip/

Dataset train_hazy_vkitti

Dataset VKITTI2 (gloomy_final_2)

Type rgb

Split train

Path /cluster/work/igp_psr/drothenpiele/data/vkitti/gloomy_final_2/foggy_rgb.zip/

Dataset val_hazy_vkitti

Dataset VKITTI2 (fog_sensor_aug1)

Type rgb

Split val

Path /cluster/work/igp_psr/drothenpiele/data/vkitti_next/gpu_2/foggy_rgb.zip/

Split muses_rgb

Dataset muses

Type rgb

Split fog_day

Path /cluster/work/igp_psr/drothenpiele/data/muses/frame_camera_trainvaltest.zip/

Split muses_sparse_depth

Dataset muses

Type sparse_depth

Split full

Path /cluster/work/igp_psr/drothenpiele/data/muses/lidar_trainvaltest.zip/

Split muses_intrinsics

Dataset muses

Type intrinsics

Split full

Path /cluster/work/igp_psr/drothenpiele/data/muses/frame_camera_trainvaltest.zip/

Split muses_extrinsics

Dataset muses

Type camera_extrinsics

Split full

Path /cluster/work/igp_psr/drothenpiele/data/muses/frame_camera_trainvaltest.zip/

Split muses_sky_mask

Dataset muses

Type semantic_segmentation

Split full

Path /cluster/work/igp_psr/drothenpiele/data/muses/gt_semantic_trainval.zip/

Execution Artifacts

run.sh

Published Exports

Published

This launch does not publish any exports yet.

Raw Artifacts

Parameters

299

Typed Parameters

Dataset clear_root_rds

real-drive-sim / full (4)

Dataset depth_root_rds

real-drive-sim / full (5)

Dataset intrinsics_root_rds

real-drive-sim / full (6)

Dataset sky_mask_root_rds

real-drive-sim / full (7)

Dataset train_hazy_root_rds

real-drive-sim / full (300)

Dataset val_hazy_root_rds

real-drive-sim / val (196)

Dataset clear_root_vkitti2

VKITTI2 / full (1)

Dataset depth_root_vkitti2

VKITTI2 / full (2)

Dataset intrinsics_root_vkitti2

VKITTI2 / full (12)

Dataset sky_mask_root_vkitti2

VKITTI2 / full (11)

Dataset train_hazy_vkitti

VKITTI2 / train (303)

Dataset val_hazy_vkitti

VKITTI2 / val (238)

Split muses_rgb

muses / fog_day (77)

Split muses_sparse_depth

muses / full (76)

Split muses_intrinsics

muses / full (82)

Split muses_extrinsics

muses / full (81)

Split muses_sky_mask

muses / full (75)

Simple Parameters

clip

true

crop

center

seed

42

tracker

wandb

use_ema

false

gpu_type

rtx_4090:1

job_name

sd21_sky_sdf

norm_max

1

norm_min

-1

run_name

multiset-muses-v2-sky-sdf

tmp_size

10G

adam_8bit

true

data_kind

real_drive_sim

ema_decay

0.9999

ema_dtype

float32

log_every

10

max_depth

800

min_depth

1

adam_beta1

0.9

adam_beta2

0.999

batch_size

24

clear_root

4

depth_mode

metric_log

depth_root

5

image_size

[384, 768]

joint_mode

single_pass

log_images

true

num_epochs

200

output_dir

/cluster/scratch/drothenpiele/SD21/multiset_muses_v2_sky_sdf

time_limit

2-00:00:00

mem_per_cpu

8G

num_workers

4

adam_epsilon

1e-8

adam_foreach

false

aspect_ratio

[1, 2]

camera_model

pinhole

conditioning

concat

lambda_depth

1

lambda_point

0.0

min_lr_ratio

0.01

project_name

joint-dehazing

use_geometry

false

warmup_ratio

0.1

weight_decay

0

cpus_per_task

8

lambda_camera

0.0

lambda_dehaze

1

lambda_eg_ssi

0.0

learning_rate

0.00004

max_depth_rds

800

max_grad_norm

1

min_depth_rds

0.00001

val_hazy_root

104

camera_enabled

true

freeze_encoder

false

log_visibility

true

num_log_images

4

use_depth_head

true

val_batch_size

10

camera_fallback

invalid

depth_head_type

base_residual

enable_xformers

true

euler_train_dir

/cluster/work/igp_psr/drothenpiele/data/out/train/sd21-joint

lambda_depth_tv

0.0002

mixed_precision

fp16

prediction_type

v_prediction

source_kind_rds

real_drive_sim

train_hazy_root

103

visibility_cmap

viridis

camera_embedding

sine

camera_emit_rays

false

camera_mlp_ratio

4

camera_num_heads

4

camera_token_dim

256

depth_noise_type

annealed_multires

lambda_log_depth

0.0

lr_schedule_type

constant_with_warmup

min_max_quantile

0.02

no_decay_enabled

false

pretrained_model

sd2-community/stable-diffusion-2-1

sky_depth_meters

800

camera_num_layers

2

camera_num_tokens

4

conv_in_init_mode

joint_marigold

geometry_lambda_z

0.15

geometry_timestep

0

joint_weight_psnr

1

joint_weight_ssim

10

lambda_depth_head

0.2

lambda_lambda_mse

0.0

lambda_visibility

0.05

max_depth_vkitti2

300

min_depth_vkitti2

0.00001

sampling_strategy

source_balanced

source_weight_rds

1

depth_ensemble_tol

0.001

lambda_uncertainty

0.0

model_depth_output

vae_decode

sky_labels_enabled

true

val_every_n_epochs

1

camera_feature_pool

avg

depth_ensemble_size

1

geometry_lambda_phi

1

lambda_segmentation

1

num_inference_steps

25

sampling_epoch_size

all

save_every_n_epochs

1

source_kind_vkitti2

vkitti2

use_visibility_head

true

warmup_start_factor

0.001

camera_metadata_root

null

conv_in_target_scale

0.5773502691896258

external_camera_mode

none

geometry_loss_robust

charbonnier

lambda_vae_depth_rec

0.0

lambda_visibility_tv

0.005

segmentation_enabled

true

sky_depth_meters_rds

800

camera_branch_enabled

true

camera_embedding_size

latent

depth_gradient_robust

charbonnier

depth_multires_levels

4

encoder_learning_rate

0.00001

geometry_head_enabled

false

geometry_lambda_theta

1

lambda_depth_gradient

0.05

lambda_geometry_total

0.0

source_weight_vkitti2

1

use_cross_task_fusion

true

use_deterministic_vae

true

vae_decoder_trainable

false

visibility_rank_pairs

128

weight_decay_backbone

0

weight_decay_no_decay

0

zero_grad_set_to_none

true

camera_feature_sources

auto

camera_metadata_format

euler_loading

depth_ensemble_max_res

1024

depth_smoothness_space

depth_head

freeze_cross_attention

true

gradient_checkpointing

true

inference_depth_output

vae_decode

lambda_geo_consistency

0.0

lambda_visibility_rank

0

use_task_skip_adapters

false

visibility_rank_margin

0.05

camera_embedding_stride

8

camera_prediction_space

residual_intrinsics

depth_ensemble_max_iter

2

depth_multires_strength

0.9

depth_smoothness_scales

4

geometric_pairs_enabled

false

joint_weight_delta1_pct

0.5

keep_last_n_checkpoints

5

lambda_segmentation_bce

0.5

lambda_segmentation_sdf

1

segmentation_latent_std

[1,1,1,1]

segmentation_noise_type

gaussian

segmentation_sdf_robust

charbonnier

validation_depth_output

both

visibility_target_gamma

4

visibility_warmup_steps

1000

cross_task_fusion_blocks

[1,2,3]

cross_task_fusion_detach

false

depth_ensemble_reduction

median

depth_normalization_type

log_depth

geometry_camera_encoding

sine

geometry_head_num_blocks

2

joint_weight_abs_rel_pct

0.5

lambda_segmentation_dice

0.5

segmentation_depth_cap_m

null

segmentation_latent_mean

[0,0,0,0]

sky_depth_meters_vkitti2

800

camera_intrinsics_loading

hierarchical

depth_diagnostics_enabled

true

depth_head_residual_scale

0.05

geometry_depth_convention

z_depth

geometry_metric_depth_max

800

geometry_metric_depth_min

1

num_inference_steps_depth

20

segmentation_latent_space

post_vae_scaling_factor

sky_semantic_label_weight

1

vae_decoder_learning_rate

0.000001

visibility_apply_to_depth

false

camera_residual_logit_clip

4

depth_head_hidden_channels

64

depth_resize_interpolation

bilinear

depth_smoothness_normalize

true

detach_rays_for_point_loss

true

enable_efficient_attention

true

lambda_visibility_preserve

0.05

sky_heuristic_label_weight

0.5

sky_labels_allow_heuristic

false

visibility_apply_to_dehaze

true

visibility_apply_to_fusion

false

visibility_hidden_channels

32

visibility_target_quantile

0.9

weight_decay_depth_decoder

0

weight_decay_geometry_head

0

camera_require_for_training

true

camera_sine_num_frequencies

null

checkpoint_selection_metric

valset/muses/mae

depth_decoder_learning_rate

0.00003

emit_camera_full_resolution

true

geometry_eg_ssi_edge_source

hazy

geometry_encoder_input_mode

hazy_hazy

geometry_head_learning_rate

0.00004

geometry_output_uncertainty

true

geometry_ray_representation

unit

gradient_accumulation_steps

1

segmentation_boundary_scale

0.125

validation_geometry_enabled

false

weight_decay_dehaze_decoder

0

camera_output_representation

angles

camera_valid_for_approximate

false

cross_task_fusion_directions

["rgb_to_depth","depth_to_rgb"]

dehaze_decoder_learning_rate

0.00003

depth_head_context_dilations

[1, 2, 4]

depth_smoothness_edge_source

clear

depth_smoothness_edge_weight

10

geometry_head_residual_scale

0.05

lambda_depth_edge_smoothness

0.002

segmentation_boundary_weight

4

weight_decay_geometry_camera

0

depth_diagnostics_edge_source

dehazed

depth_diagnostics_edge_weight

10

depth_head_base_upsample_mode

bilinear

geometry_camera_learning_rate

0.00004

geometry_head_hidden_channels

128

geometry_sine_num_frequencies

64

visibility_decoder_grad_scale

1.0

camera_approximate_fov_degrees

60

checkpoint_selection_direction

min

cross_task_fusion_learned_gate

true

geometry_use_deterministic_vae

true

segmentation_codec_encode_mode

deterministic

depth_multires_downscale_factor

2

lambda_depth_gt_edge_smoothness

0

segmentation_direct_start_epoch

3

validation_geometry_camera_mode

predicted

validation_geometry_output_mode

geometry_head

camera_include_input_in_encoding

null

camera_intrinsics_metadata_scope

intrinsics

cross_task_fusion_gate_init_bias

-4

cross_task_fusion_sender_lowpass

false

depth_head_use_depthwise_context

true

geometric_pairs_views_per_sample

2

geometry_condition_on_visibility

false

geometry_detach_encoder_features

true

segmentation_depth_output_policy

raw

segmentation_inference_threshold

0

segmentation_sdf_truncation_type

short_side_fraction

cross_task_fusion_hidden_channels

64

depth_diagnostics_normalize_depth

true

segmentation_head_hidden_channels

64

segmentation_sdf_truncation_value

0.0625

segmentation_validation_threshold

0

geometric_pairs_emit_pair_metadata

true

geometry_include_input_in_encoding

false

lambda_depth_multiscale_smoothness

0.001

depth_ensemble_regularizer_strength

0.02

segmentation_direct_timestep_weight

alpha2

segmentation_head_context_dilations

[1,2,4]

segmentation_latent_standardization

false

segmentation_detach_encoder_features

true

segmentation_probability_temperature

0.1

depth_diagnostics_road_edge_threshold

0.08

geometry_head_detach_camera_embedding

true

depth_diagnostics_road_bottom_fraction

0.4

geometry_encoder_no_grad_when_detached

true

cross_task_fusion_sender_lowpass_kernel

3

depth_diagnostics_high_frequency_kernel

5

segmentation_head_use_depthwise_context

true

validation_geometry_compute_edge_metrics

true

validation_geometry_compute_point_metrics

false

validation_geometry_max_points_per_sample

20000

validation_geometry_save_geometry_samples

true

validation_geometry_compute_camera_metrics

true

validation_geometry_compute_resolution_sweep

false

validation_geometry_max_point_metric_samples

8

validation_geometry_save_point_cloud_samples

false

validation_geometry_compute_uncertainty_metrics

true

Raw JSON

{
  "clip": "true",
  "crop": "center",
  "seed": 42,
  "tracker": "wandb",
  "use_ema": "false",
  "gpu_type": "rtx_4090:1",
  "job_name": "sd21_sky_sdf",
  "norm_max": 1,
  "norm_min": -1,
  "run_name": "multiset-muses-v2-sky-sdf",
  "tmp...

Events

Launch Events

No launch events recorded.

Output

Datasets

Execution Artifacts

Published Exports

Launch-Owned Exports

Parameters

Typed Parameters

Simple Parameters

Events

Launch Events