code/train_lpd.sh

#!/bin/bash
# Train LiDAR-Perfect Depth on top of a pretrained PPD checkpoint.
#
# Stage 1 (image pretrain @512²): trains only the sparse-prompt encoder + gate
#   on Hypersim with simulated sparse-LiDAR; DiT backbone frozen.
# Stage 2 (image finetune @1024×768): unfreezes the DiT (optional) and mixes
#   Hypersim + UrbanSyn + UnrealStereo4K + VKITTI2 + TartanAir.
# Stage 3 (video finetune): adds the temporal Kalman filter loop on short clips.
#
# Pre-reqs:
#   * checkpoints/ppd.pth                              <- PPD pretrained weights
#   * checkpoints/depth_anything_v2_vitl.pth           <- DA-V2 ViT-L semantics
#   * datasets extracted under /mnt/sig/datasets/      <- see datasets/README.md

set -e

# Stage 1: image pretrain (Hypersim only, 512×512)
python main.py --cfg_file ppd/configs/lpd_pretrain.yaml pl_trainer.devices=1

# Stage 2 (uncomment after stage 1 produces a checkpoint):
# python main.py --cfg_file ppd/configs/lpd_finetune.yaml pl_trainer.devices=8

# Stage 3 (uncomment for video):
# python main.py --cfg_file ppd/configs/lpd_video.yaml pl_trainer.devices=8