[Example Code] Minimal Pointcept to Verify fVDB PTV3 Against Official PTV3

point_transformer_v3/external/pointcept_minimal/.github/workflows/formatter.yml

@@ -0,0 +1,20 @@
+name: Formatter
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+  pull_request:
+    types: [opened, reopened, synchronize]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  formatter:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - uses: psf/black@stable

point_transformer_v3/external/pointcept_minimal/.gitignore

@@ -0,0 +1,16 @@
+image/
+__pycache__
+**/build/
+**/*.egg-info/
+**/dist/
+*.so
+exp
+weights
+data
+*log*
+outputs/
+.vscode
+.idea
+*/.DS_Store
+**/*.out
+Dockerfile

point_transformer_v3/external/pointcept_minimal/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Pointcept
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

point_transformer_v3/external/pointcept_minimal/README.md

@@ -0,0 +1,53 @@
+# Pointcept Minimal
+
+A minimal version of [Pointcept](https://github.com/Pointcept/Pointcept) for comparing official PTv3 and fVDB PTv3 training results on ScanNet.
+
+## Environment Setup
+
+1. **fVDB-core**: Follow fVDB's installation instructions and install `env/dev_environment.yml`. Tested with fVDB commit `51cec3d3e90d7d571e22862d17ae9051cbd13afd`.
+2. Activate the environment and install dependencies:
+   ```bash
+   conda activate fvdb
+   pip install -r requirements_pointceptminimal.txt
+   ```
+
+## Dataset Preparation
+
+Follow the [Pointcept scannet dataset preparation](https://github.com/Pointcept/Pointcept?tab=readme-ov-file#scannet-v2) to download ScanNet and prepare the dataset in the correct location.
+
+## Training Configurations
+
+### With Convolution (CPE)
+Both configurations train PTv3 on ScanNet with convolution enabled. Training loss curves will diverge. 
+
+**fVDB PTv3:**
+```bash
+CUDA_VISIBLE_DEVICES=4,5,6,7 sh scripts/train.sh -g 4 -d scannet -c semseg-pt-v3m1-0-fvdb-test-4g-conv -n semseg-pt-v3m1-0-fvdb-test-4g-conv
+```
+
+**Official PTv3:**
+```bash
+CUDA_VISIBLE_DEVICES=4,5,6,7 sh scripts/train.sh -g 4 -d scannet -c semseg-pt-v3m1-0-test-4g-conv -n semseg-pt-v3m1-0-test-4g-conv
+```
+
+### Without Convolution (CPE)
+Both configurations train PTv3 on ScanNet without convolution. Training loss curves are identical. 
+
+**fVDB PTv3:**
+```bash
+CUDA_VISIBLE_DEVICES=4,5,6,7 sh scripts/train.sh -g 4 -d scannet -c semseg-pt-v3m1-0-fvdb-test-4g -n semseg-pt-v3m1-0-fvdb-test-4g
+```
+
+**Official PTv3:**
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 sh scripts/train.sh -g 4 -d scannet -c semseg-pt-v3m1-0-test-4g -n semseg-pt-v3m1-0-test-4g
+```
+
+## Weights & Biases (Optional)
+
+To enable W&B logging, set the following in your config file:
+```python
+enable_wandb = True
+wandb_project = "your_project"
+wandb_key = "your_key"
+```

point_transformer_v3/external/pointcept_minimal/configs/_base_/dataset/scannet.py

@@ -0,0 +1,26 @@
+class_names = [
+    "wall",
+    "floor",
+    "cabinet",
+    "bed",
+    "chair",
+    "sofa",
+    "table",
+    "door",
+    "window",
+    "bookshelf",
+    "picture",
+    "counter",
+    "desk",
+    "curtain",
+    "refridgerator",
+    "shower curtain",
+    "toilet",
+    "sink",
+    "bathtub",
+    "otherfurniture",
+]
+
+data = dict(
+    names=class_names,
+)

point_transformer_v3/external/pointcept_minimal/configs/_base_/default_runtime.py

@@ -0,0 +1,48 @@
+weight = None  # path to model weight
+resume = False  # whether to resume training process
+evaluate = True  # evaluate after each epoch training process
+test_only = False  # test process
+
+seed = None  # train process will init a random seed and record
+save_path = "exp/default"
+num_worker = 16  # total worker in all gpu
+batch_size = 16  # total batch size in all gpu
+gradient_accumulation_steps = 1  # total steps to accumulate gradients for
+batch_size_val = None  # auto adapt to bs 1 for each gpu
+batch_size_test = None  # auto adapt to bs 1 for each gpu
+epoch = 100  # total epoch, data loop = epoch // eval_epoch
+eval_epoch = 100  # sche total eval & checkpoint epoch
+clip_grad = None  # disable with None, enable with a float
+
+sync_bn = False
+enable_amp = False
+amp_dtype = "float16"
+empty_cache = False
+empty_cache_per_epoch = False
+find_unused_parameters = False
+
+enable_wandb = False
+wandb_project = "pointcept"  # custom your project name e.g. Sonata, PTv3
+wandb_key = None  # wandb token, default is None. If None, login with `wandb login` in your terminal
+
+mix_prob = 0
+param_dicts = None  # example: param_dicts = [dict(keyword="block", lr_scale=0.1)]
+
+memory_snapshot_enabled = False
+
+# hook
+hooks = [
+    dict(type="CheckpointLoader"),
+    dict(type="ModelHook"),
+    dict(type="IterationTimer", warmup_iter=2),
+    dict(type="InformationWriter"),
+    dict(type="SemSegEvaluator"),
+    dict(type="CheckpointSaver", save_freq=None),
+    dict(type="PreciseEvaluator", test_last=False),
+]
+
+# Trainer
+train = dict(type="DefaultTrainer")
+
+# Tester
+test = dict(type="SemSegTester", verbose=True)

point_transformer_v3/external/pointcept_minimal/configs/_base_/fvdb_runtime.py

@@ -0,0 +1,55 @@
+# Copyright Contributors to the OpenVDB Project
+# SPDX-License-Identifier: Apache-2.0
+
+"""
+Runtime configuration for FVDB-based Point Transformer V3 models.
+"""
+
+from __future__ import annotations
+
+weight = None  # path to model weight
+resume = False  # whether to resume training process
+evaluate = True  # evaluate after each epoch training process
+test_only = False  # test process
+
+seed = None  # train process will init a random seed and record
+save_path = "exp/default"
+num_worker = 16  # total worker in all gpu
+batch_size = 16  # total batch size in all gpu
+gradient_accumulation_steps = 1  # total steps to accumulate gradients for
+batch_size_val = None  # auto adapt to bs 1 for each gpu
+batch_size_test = None  # auto adapt to bs 1 for each gpu
+epoch = 100  # total epoch, data loop = epoch // eval_epoch
+eval_epoch = 100  # sche total eval & checkpoint epoch
+clip_grad = None  # disable with None, enable with a float
+
+sync_bn = False
+enable_amp = False
+amp_dtype = "float16"
+empty_cache = False
+empty_cache_per_epoch = False
+find_unused_parameters = False
+
+enable_wandb = True
+wandb_project = "pointcept"  # custom your project name e.g. Sonata, PTv3
+wandb_key = None  # wandb token, default is None. If None, login with `wandb login` in your terminal
+
+mix_prob = 0
+param_dicts = None  # example: param_dicts = [dict(keyword="block", lr_scale=0.1)]
+
+# hook
+hooks = [
+    dict(type="CheckpointLoader"),
+    dict(type="ModelHook"),
+    dict(type="IterationTimer", warmup_iter=2),
+    dict(type="InformationWriter"),
+    dict(type="SemSegEvaluator"),
+    dict(type="CheckpointSaver", save_freq=None),
+    dict(type="PreciseEvaluator", test_last=False),
+]
+
+# Trainer
+train = dict(type="DefaultTrainer")
+
+# Tester
+test = dict(type="SemSegTester", verbose=True)