#!/usr/bin/env bash
#
# Single-scene overfit smoke test on ARKitScenes (scene 40958756, idx 0).
#
# ARKitScenes analog of finetune_wildrgbd_overfit.sh. Loads DFoT_RE10K.ckpt +
# re10k.ckpt as initialization (same as train.sh / train_re10k.sh / the WildRGBD
# overfit) then finetunes the encoder on a single ARKitScenes (ctxt, trgt, intm)
# sample with overfit_to_index=0 so we can verify the loss actually drops.
#
# Camera convention: verify_arkitscenes_pose_convention.py confirms CUT3R's
# preprocessed ARKitScenes camera_pose is c2w OpenCV (x right, y down, z fwd) —
# the convention the Cut3rAdapter assumes. Multi-view consistency at 94% <5cm
# for the OpenCV hypothesis vs 53% for the OpenGL flip. NO pose flip needed.
#
# Depth: ARKitScenes ships per-frame metric iPad-LiDAR depth (lowres_depth, mm,
# 0=invalid). Depth supervision + the depth mask are ENABLED (use_depth_supervision
# + model.encoder.use_depth_mask=true), per the SPOC-style setup.
#
# Required env vars (with sensible defaults):
#   DATASET_ROOT  - path to processed_arkitscenes/  (the dir with Training/)
#   CKPT_DIR      - directory with DFoT_RE10K.ckpt + re10k.ckpt
#   CUDA_VISIBLE_DEVICES - which GPU (default: 3)

set -e

# ---- environment ----
# NOTE: use the `3d-belief` env (torch 2.11), NOT `3d-belief-release`. The
# release env's diff_gaussian_rasterization _C.so is stale (built against an
# older libtorch) and fails with `undefined symbol ...decref_pyobject`. The
# `3d-belief` env has a working rasterizer build. (quaternion is broken there
# under numpy 2.x but is only needed by the CUT3R *preprocessing* scripts, not
# the dataloader/training path.)
eval "$(conda shell.bash hook)"
conda activate 3d-belief

nvidia-smi
REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
cd "$REPO_ROOT"
# Avoid PYTHONPATH shadowing by an older splat_belief copy
# (see memory: 3d-belief-training-script-gotchas).
export PYTHONPATH="${REPO_ROOT}:${PYTHONPATH}"

DATASET_ROOT="${DATASET_ROOT:-/home/ubuntu/tianmin-neurips/zwen19/data/ARKitScenes/processed_arkitscenes}"
CKPT_DIR="${CKPT_DIR:-${REPO_ROOT}/checkpoints}"
STAGE="${STAGE:-train}"

export MASTER_PORT=$((12000 + RANDOM % 1000))
export PATH=$CONDA_PREFIX/bin:$PATH
export CUDA_HOME=$CONDA_PREFIX
export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-3}"
export TORCH_CUDA_ARCH_LIST="8.6;9.0"

# Tunables for the smoke test — override in the env to scale up.
RESULTS_FOLDER="${RESULTS_FOLDER:-outputs/training/arkitscenes_overfit_scene0}"
TRAIN_NUM_STEPS="${TRAIN_NUM_STEPS:-500}"
SAMPLE_EVERY="${SAMPLE_EVERY:-100}"
SAVE_EVERY="${SAVE_EVERY:-500}"
# wandb_every must divide sample_every (assert in diffusion.py:1514).
WANDB_EVERY="${WANDB_EVERY:-${SAMPLE_EVERY}}"
NUM_INTERMEDIATE="${NUM_INTERMEDIATE:-3}"
OVERFIT_TO_INDEX="${OVERFIT_TO_INDEX:-0}"   # idx 0 -> scene 40958756, group 0
# The base config warmup_period=10000 (with lr=2e-5 => +2e-9/step). That keeps
# the LR near-zero for a short overfit, so the loss can't move. Use a short
# warmup for the overfit smoke (the multi-scene training script keeps the
# canonical 10000).
WARMUP_PERIOD="${WARMUP_PERIOD:-200}"
LR="${LR:-2e-5}"

mkdir -p "${RESULTS_FOLDER}"

# DFoT init (same as train.sh / train_re10k.sh / WildRGBD overfit).
CKPT_PATH="${CKPT_PATH:-${CKPT_DIR}/DFoT_RE10K.ckpt}"
LOAD_ENC="${LOAD_ENC:-true}"
ENCODER_CKPT="${ENCODER_CKPT:-${CKPT_DIR}/re10k.ckpt}"
DATASET_VIZ_TYPE="${DATASET_VIZ_TYPE:-}"  # leave empty to keep YAML default (interpolation)

CUDA_LAUNCH_BLOCKING=1 torchrun --nnodes 1 --nproc_per_node 1 --master_port $MASTER_PORT \
    splat_belief/experiment/train.py \
    dataset=arkitscenes_cut3r \
    dataset.root_dir="${DATASET_ROOT}" \
    dataset.vggt_alignment_loss_weight=2.0 \
    dataset.intermediate_weight=5.0 \
    dataset.depth_smooth_loss_weight=0.1 \
    setting_name=debug \
    stage="${STAGE}" \
    use_depth_supervision=true \
    results_folder="${RESULTS_FOLDER}" \
    semantic_config=splat_belief/config/semantic/onehot.yaml \
    checkpoint_path="${CKPT_PATH}" \
    $( [ -n "${DATASET_VIZ_TYPE}" ] && echo "dataset.viz_type=${DATASET_VIZ_TYPE}" ) \
    ngpus=1 \
    image_size=128 \
    train_num_steps=${TRAIN_NUM_STEPS} \
    warmup_period=${WARMUP_PERIOD} \
    lr=${LR} \
    sample_every=${SAMPLE_EVERY} \
    save_every=${SAVE_EVERY} \
    wandb_every=${WANDB_EVERY} \
    overfit_to_index=${OVERFIT_TO_INDEX} \
    ctxt_min=5 \
    ctxt_max=15 \
    model/encoder=uvitmvsplat \
    model.encoder.use_image_condition=true \
    model.encoder.depth_predictor_time_embed=true \
    model.encoder.use_camera_pose=true \
    model.encoder.use_semantic=false \
    model.encoder.use_reg_model=false \
    model.encoder.d_semantic=512 \
    model.encoder.d_semantic_reg=384 \
    model.encoder.gaussians_per_pixel=1 \
    model.encoder.evolve_ctxt=false \
    model.encoder.use_depth_mask=true \
    model.encoder.encoder_ckpt="${ENCODER_CKPT}" \
    model.encoder.freeze_depth_predictor=false \
    model/encoder/backbone=u_vit3d_pose \
    model.encoder.backbone.use_vggt_alignment=true \
    model.encoder.backbone.use_repa=true \
    model.encoder.backbone.input_size='[128, 128]' \
    alignment.latents_info=-1 \
    ctxt_losses_factor=0.9 \
    repa_encoder_resolution=512 \
    model_type=uvit_pose \
    name=arkitscenes_overfit \
    wandb=local \
    clean_target=false \
    use_identity=true \
    intermediate=true \
    num_intermediate=${NUM_INTERMEDIATE} \
    load_optimizer=false \
    load_enc=${LOAD_ENC} \
    finetune_component=encoder \
    finetune_steps=5 \
    use_depth_smoothness=true \
    adjacent_angle=0.785 \
    adjacent_distance=1.0