Run 2026-02-06_15-18-05_db24 - MERGE

Run Dossier

Training run for MERGE

8 metrics · 11,672 train · 0 val

Chart Overlays

Compare this run against its peers

0/12

Add peer runs here to layer their train and val traces onto the current metric charts.

completed Feb 6, 15:18 Started 85d 21h ago Python 3.9.18

merge/train_dehaze_depth.py --pretrained_model_name_or_path=PixArt-alpha/PixArt-XL-2-512x512 --config=config/train_dehaze_depth.yaml --output_dir=/cluster/scratch/drothenpiele/MERGE/static_prompt_unfrozen_2_checkpoint --train_batch_size=1 --gradient_accumulation_steps=16 --num_train_epochs=1000 --learning_rate=1e-4 --lr_scheduler=linear --lr_warmup_steps=50 --mixed_precision=fp16 --seed=42 --num_layers=14 --num_attention_heads=16 --attention_head_dim=72 --ff_mult=1 --converter_init_type=pretrained --multi_res_noise=True --lambda_dehaze=1.0 --lambda_depth=1.0 --checkpointing_steps=1000 --validation_steps=10 --report_to=wandb --wandb_name=dehaze_depth_vkitti --tracker_project_name=merge_dehaze_depth --base_data_dir=/cluster/work/igp_psr/drothenpiele/data/ --dataloader_num_workers=2 --max_grad_norm=1.0 --num_unfrozen_backbone_layers=2 --gradient_checkpointing --run_name=Unfrozen-2-with-grad-checkpoint --runlog_dir=/cluster/work/igp_psr/drothenpiele/out/train/MERGE

lr 1e-4precision fp16 seed 42grad accum 16grad clip 1

ID 56385486 Job MERGE_d_d Part gpuhe.120h Node eu-g6-030 CPU 8 GPU nvidia_geforce_rtx_4090:1

2026-02-06_15-18-05_db24 train (smoothed) —