Use fvdb.nn.PointTransformerV3 in the PTv3 example

point_transformer_v3/README.md

@@ -1,6 +1,6 @@
 # Point Transformer V3 (PT-v3) - FVDB Implementation
 
-This repository contains a minimal implementation of Point Transformer V3 using the FVDB library for scalable 3D point cloud processing.
+This example trains and runs Point Transformer V3 with `fvdb.nn.PointTransformerV3` from fVDB, through Pointcept.
 
 ## Setup
 
@@ -40,8 +40,8 @@ python scripts/test/compute_difference.py --stats_path_1 data/scannet_samples_sm
 
 ## Project Structure
 
-- `fvdb_extensions/models/ptv3_fvdb.py` - Core FVDB implementation
-- `fvdb_extensions/models/point_transformer_v3m1_fvdb.py` - Pointcept framework adapter
+- `fvdb_extensions/models/point_transformer_v3m1_fvdb.py` - Pointcept backbone (`PT-v3fvdb`) wrapping `fvdb.nn.PointTransformerV3`
+- `fvdb_extensions/configs/` - Pointcept training configs for the fvdb backbone and the official PT-v3m1
 - `scripts/data/` - Data download and preprocessing scripts
 - `scripts/test/` - Inference and comparison scripts
 

point_transformer_v3/fvdb_extensions/configs/semseg-pt-v3m1-0-fvdb-test.py

@@ -19,23 +19,34 @@
     num_classes=20,
     backbone_out_channels=64,
     backbone=dict(
+        # fvdb.nn.PointTransformerV3; arguments follow Pointcept's PT-v3m1 (decoder lists shallow to deep).
         type="PT-v3fvdb",
         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_heads=(2, 4, 8, 16, 32),
+        enc_num_head=(2, 4, 8, 16, 32),
+        enc_patch_size=(1024, 1024, 1024, 1024, 1024),
         dec_depths=(2, 2, 2, 2),
-        dec_channels=(256, 128, 64, 64),
-        dec_num_heads=(16, 8, 4, 4),
-        patch_size=1024,
-        drop_path=0.3,
-        proj_drop=0.0,
-        no_conv_in_cpe=False,
-        order_type=("z", "z-trans", "hilbert", "hilbert-trans"),
-        enable_batch_norm=True,
+        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,
+        proj_drop=0.0,
+        drop_path=0.3,
+        pre_norm=True,
         shuffle_orders=True,
+        # Serialize the raw grid_coord, as Pointcept does. SphereCrop leaves samples whose
+        # coordinates do not start at zero, and the curve orders depend on the origin.
+        serialization_offset="none",
+        # The settings fvdb-examples has compared against Pointcept with: a linear stem,
+        # batch norm in the stem and pooling layers, and convolution in the position encoding.
         embedding_mode="linear",
+        norm="batch",
+        cpe_conv=True,
     ),
     criteria=[
         dict(type="CrossEntropyLoss", loss_weight=1.0, ignore_index=-1),

point_transformer_v3/fvdb_extensions/models/__init__.py

@@ -2,6 +2,5 @@
 # SPDX-License-Identifier: Apache-2.0
 
 
-# Lazy imports - only import when explicitly requested
-# This allows ptv3_fvdb to be imported without pulling in pointcept dependencies
-__all__ = ["point_transformer_v3m1_fvdb", "ptv3_fvdb"]
+# The Pointcept backbone is imported explicitly, so importing this package does not pull in Pointcept.
+__all__ = ["point_transformer_v3m1_fvdb"]