# Resolved training/model config for the bundled checkpoint (weights/model_best.pth).
# Trimmed down from the original training config: the original `data.train` /
# `data.val` / `data.test` blocks referenced our internal dataset class and
# training-data paths, which aren't needed to run inference and have been
# removed. `segment_scan.py` builds the model straight from `model` below and
# applies its own preprocessing pipeline (see PREPROCESS_TRANSFORM /
# TEST_TIME_AUGMENTATIONS in that script) instead of going through a dataset
# class, so nothing here is unused cruft other than the two notes below.

weight = "weights/model_best.pth"  # relative to this delivery folder's root
resume = False
evaluate = True
test_only = False
seed = 37743921

# save_path is only consulted by the standard Pointcept tools/train.py /
# tools/test.py entry points (dataset-based train/eval), which this delivery
# doesn't use. Left here only so this file stays loadable by those scripts if
# you ever want to wire up your own dataset for standard evaluation.
save_path = "./output"

num_worker = 8
batch_size = 1
batch_size_val = None
batch_size_test = None
epoch = 3000
eval_epoch = 100
sync_bn = False
enable_amp = True
empty_cache = False
empty_cache_per_epoch = False
find_unused_parameters = False
mix_prob = 0.8
param_dicts = [dict(keyword="block", lr=0.0003)]
hooks = [
    dict(type="CheckpointLoader"),
    dict(type="IterationTimer", warmup_iter=2),
    dict(type="InformationWriter"),
    dict(type="SemSegEvaluator"),
    dict(type="CheckpointSaver", save_freq=None),
    dict(type="PreciseEvaluator", test_last=False),
]
train = dict(type="DefaultTrainer")
test = dict(type="SemSegTester", verbose=True)
model = dict(
    type="DefaultSegmentorV2",
    num_classes=10,
    backbone_out_channels=64,
    backbone=dict(
        type="PT-v3m1",
        in_channels=6,
        order=("z", "z-trans", "hilbert", "hilbert-trans"),
        stride=(2, 2, 2, 2),
        enc_depths=(2, 2, 2, 6, 2),
        enc_channels=(32, 64, 128, 256, 512),
        enc_num_head=(2, 4, 8, 16, 32),
        enc_patch_size=(1024, 1024, 1024, 1024, 1024),
        dec_depths=(2, 2, 2, 2),
        dec_channels=(64, 64, 128, 256),
        dec_num_head=(4, 4, 8, 16),
        dec_patch_size=(1024, 1024, 1024, 1024),
        mlp_ratio=4,
        qkv_bias=True,
        qk_scale=None,
        attn_drop=0.0,
        proj_drop=0.0,
        drop_path=0.3,
        shuffle_orders=True,
        pre_norm=True,
        enable_rpe=False,
        enable_flash=False,  # flash-attn deliberately excluded from the minimal install; standard attention is used instead
        upcast_attention=False,
        upcast_softmax=False,
        enc_mode=False,  # renamed from `cls_mode` in newer upstream Pointcept
        pdnorm_bn=False,
        pdnorm_ln=False,
        pdnorm_decouple=True,
        pdnorm_adaptive=False,
        pdnorm_affine=True,
        pdnorm_conditions=("ScanNet", "S3DIS", "Structured3D"),
    ),
    criteria=[
        dict(type="CrossEntropyLoss", loss_weight=1.0, ignore_index=-1),
        dict(type="LovaszLoss", mode="multiclass", loss_weight=1.0, ignore_index=-1),
    ],
)
optimizer = dict(type="AdamW", lr=0.003, weight_decay=0.05)
scheduler = dict(
    type="OneCycleLR",
    max_lr=[0.003, 0.0003],
    pct_start=0.05,
    anneal_strategy="cos",
    div_factor=10.0,
    final_div_factor=1000.0,
)

# Class scheme the model was trained on. Predicted label indices from
# segment_scan.py map directly onto this list (no remap table needed).
data = dict(
    num_classes=10,
    ignore_index=-1,
    names=[
        "clutter",
        "floor",
        "ceiling",
        "wall",
        "column",
        "door",
        "window",
        "stairs",
        "railing",
        "lights",
    ],
)