nbrightandClaude Opus 5
f39b093106
Adds STATUS.md as the handoff document: benchmark numbers, the bare-earth
metrics that actually matter for this project, the Korean-data domain gap that
retraining will not fix, and what to do on the 24 GB machine.
The memory problem was in how a prediction's colours were turned back into
class indices. Every consumer built an (N, 13, 3) float64 temporary:
d = ((rgb[:, None, :] - COLOR_MAP[None, :, :]) ** 2).sum(axis=2)
That is ~250 MB of intermediates per 800k-point tile, several live at once, and
a full 4.7M-point block pushes it into gigabytes. main.py writes exact palette
entries, so an exact hash lookup resolves nearly every point with no large
temporary; only leftovers fall back to a chunked distance search. Peak RSS on a
470k-point tile drops to 61 MB. Extracted to sumparts_palette.py and shared by
coarse_eval.py and split_by_class.py.
Also from this round:
- patch_cm_mutation.sh: ConfusionMatrix.update() rewrote the caller's pred
tensor in place, folding every ignore_index point into class num_classes-1.
test() saves its visualization from that same tensor afterwards, so an
unlabelled tile came out 100% wall and the model looked degenerate when it
was not.
- patch_class_mask.sh: SUMPARTS_MASK_CLASSES drops known-absent classes from
the argmax. Measured on Seosan and it does not help - the runner-up for
"water" is "wall", not "terrain" - but the experiment is worth keeping.
- split_by_class.py now writes .ply alongside .obj. A vertex-only OBJ has zero
faces and most viewers render nothing, which is why the first export looked
broken.
- verify_outputs.sh reads exported files back with a parser, so "here are your
files" can be checked rather than asserted.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
59 lines
1.9 KiB
Bash
59 lines
1.9 KiB
Bash
#!/usr/bin/env bash
|
|
# SUM Parts - run inference on the converted Korean tile
|
|
#
|
|
# This answers one question: does a SUM Parts model accept our data and emit
|
|
# per-point predictions? It does NOT measure anything. The checkpoint is the
|
|
# 1-epoch smoke-test model and the tile carries label=0 everywhere, so the
|
|
# reported mIoU/OA are meaningless by construction. Look at whether it runs and
|
|
# what the predicted class histogram looks like, nothing else.
|
|
set -euo pipefail
|
|
|
|
CONDA_ROOT="$HOME/miniconda3"
|
|
ENV_NAME="sumparts"
|
|
REPO="$HOME/sum-parts/semantic_segmentation/PointNeXt_bundle"
|
|
SEG="$REPO/examples/segmentation"
|
|
|
|
TILE="$HOME/sum-parts/data/korea_poc/seosan_BlockYBA_tile0.ply"
|
|
TRACK="$HOME/sum-parts/data/korea_poc_track"
|
|
LOG="${LOG:-/tmp/sumparts_poc_infer.log}"
|
|
|
|
source "$CONDA_ROOT/etc/profile.d/conda.sh"
|
|
conda activate "$ENV_NAME"
|
|
export WANDB_MODE=disabled WANDB_SILENT=true
|
|
export CUDA_HOME="$CONDA_PREFIX"
|
|
|
|
[ -f "$TILE" ] || { echo "error: $TILE missing -- run poc_korea.sh first" >&2; exit 1; }
|
|
|
|
# The dataset class globs {train,val,test}/*.ply and errors on a missing split,
|
|
# so all three have to exist even for a test-only run.
|
|
for split in train val test; do
|
|
mkdir -p "$TRACK/$split"
|
|
ln -f "$TILE" "$TRACK/$split/$(basename "$TILE")"
|
|
done
|
|
rm -rf "$TRACK/processed"
|
|
|
|
# The cfg has to match the architecture that wrote the checkpoint; hardcoding
|
|
# one here loads the weights into the wrong model and torch raises on the
|
|
# state_dict.
|
|
source "/mnt/d/MYCLAUDE_PROJECT/sum-parts-test/scripts/resolve_ckpt.sh"
|
|
|
|
cd "$SEG"
|
|
|
|
set +e
|
|
python -u main.py \
|
|
--cfg "../../cfgs/sumv2_triangle/${CKPT_CFG}.yaml" \
|
|
mode=test \
|
|
--pretrained_path "$CKPT" \
|
|
dataset.common.data_root="$TRACK" \
|
|
wandb.use_wandb=False \
|
|
batch_size=2 \
|
|
val_batch_size=1 \
|
|
> "$LOG" 2>&1
|
|
rc=$?
|
|
set -e
|
|
|
|
echo "=== exit=$rc | last 25 lines ==="
|
|
tail -25 "$LOG"
|
|
[ "$rc" -eq 0 ] && echo "POC INFER DONE" || echo "POC INFER FAILED (rc=$rc)"
|
|
exit "$rc"
|