# kpconv-ev

**Repository Path**: designer7458/kpconv-ev

## Basic Information

- **Project Name**: kpconv-ev
- **Description**: KPConv实验复现
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-07-19
- **Last Updated**: 2026-07-21

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# KPConv / KPFCNN on EV-Flying

This project adapts Open3D-ML's PyTorch `KPFCNN` semantic segmentation model to
the EV-Flying point-level foreground/background task.

## Scope

- Data: `/home/zikun/jzw/experiments/cetus/EV-Flying-dataset`
- Environment: existing conda env `randla_ev`
- Model: `open3d.ml.torch.models.KPFCNN`
- Input points: normalized `[x, y, t]` from `evs_norm[:, 0:3]`
- Input feature: polarity from `evs_norm[:, 3:4]`
- Label: foreground/background from `evs_norm[:, 4]`
- Track id: never used as an input feature
- Open3D kernel cache: `/tmp/kpconv_open3d_kernel_cache`

Training uses a fixed validation threshold of `0.9` and saves
`best_loss_seed37.pt` and `best_iou_seed37.pt`. Testing exports full original
point predictions and the point-level evaluator chooses the best test threshold
from the exported probabilities.

## CPU Checks

```bash
CUDA_VISIBLE_DEVICES= MPLCONFIGDIR=/tmp/kpconv_mpl \
/home/zikun/anaconda3/envs/randla_ev/bin/python tools/check_ev_flying_dataset.py \
  --config configs/ev_flying_kpconv.yaml

CUDA_VISIBLE_DEVICES= MPLCONFIGDIR=/tmp/kpconv_mpl \
/home/zikun/anaconda3/envs/randla_ev/bin/python tools/smoke_open3d_kpconv_cpu.py \
  --config configs/ev_flying_kpconv.yaml --forward \
  --max-sample-points 4096 --min-in-points 128 --max-in-points 256 --batch-limit 256
```

The command above uses a reduced CPU-only model/patch size for shape smoke.
The default config can be instantiated on CPU without a forward pass:

```bash
CUDA_VISIBLE_DEVICES= MPLCONFIGDIR=/tmp/kpconv_mpl \
/home/zikun/anaconda3/envs/randla_ev/bin/python -c \
"from pathlib import Path; import sys, torch; sys.path.insert(0, 'tools'); from kpconv_common import resolve_config, build_model; cfg=resolve_config('configs/ev_flying_kpconv.yaml'); model=build_model(cfg, torch.device('cpu')); print(type(model).__name__, model.cfg.first_features_dim, model.cfg.max_in_points, torch.cuda.is_available())"
```

## GPU Smoke

Run only after GPU use is explicitly allowed.

```bash
/home/zikun/anaconda3/envs/randla_ev/bin/python tools/train_kpconv_ev_flying.py \
  --config configs/ev_flying_kpconv.yaml --device cuda:0 \
  --epochs 1 --limit-train 2 --limit-val 2
```

## Test And Eval

After training creates both checkpoints:

```bash
/home/zikun/anaconda3/envs/randla_ev/bin/python tools/evaluate_checkpoints.py \
  --config configs/ev_flying_kpconv.yaml --device cuda:0
```

This tests `best_loss_seed37.pt` and `best_iou_seed37.pt`, writes
`test_best_loss/predictions.txt` and `test_best_iou/predictions.txt`, runs
point-level threshold eval on each, and selects the higher test IoU in
`outputs/ev_flying/checkpoint_eval_summary.json`.