HOTFIX : use pdal instead of laspy for interpolation

README.md

@@ -49,15 +49,15 @@ source bash/setup_environment/setup_env.sh  # with cudatoolkit
 source bash/setup_environment/setup_env_cpu_only.sh  # cpu version only
 
 # activate using
-conda activate lidar_deep_segmentation
+conda activate lidar_multiclass
 ```
 
 ### Run inference from package
 If you are interested in running inference from anywhere, you can install code as a package in a your virtual environment.
 
 ```
 # activate an env matching ./bash/setup_env.sh requirements.
-conda activate lidar_deep_segmentation
+conda activate lidar_multiclass
 
 # install the package
 pip install --upgrade https://github.com/IGNF/lidar-deep-segmentation/tarball/main  # from github directly

bash/setup_environment/requirements.yml

@@ -1,17 +1,17 @@
-name: lidar_deep_segmentation
+name: lidar_multiclass
 channels:
   - conda-forge
   - anaconda
   - comet_ml
   - pytorch
 dependencies:
-  - python==3.9
+  - python==3.9.*
   - pip
   # --------- numpy --------- #
   - numpy==1.20
   # --------- geo --------- #
-  - laspy
   - pygeos
+  - laspy
   - python-pdal  # useful for data preparation
   # --------- loggers --------- #
   - comet_ml

bash/setup_environment/setup_env.sh

@@ -3,7 +3,7 @@ set -e
 # SETUP
 conda install -y mamba -n base -c conda-forge
 mamba env create -f bash/setup_environment/requirements.yml
-conda activate lidar_deep_segmentation
+conda activate lidar_multiclass
 
 # INSTALL
 export PYTORCHVERSION="1.10.1"
@@ -13,4 +13,4 @@ mamba install -y pytorch==$PYTORCHVERSION torchvision==$TORCHVISIONVERSION cudat
 mamba install -y pyg==2.0.3 -c pytorch -c pyg -c conda-forge
 FORCE_CUDA=1 pip install torch-points-kernels --no-cache
 pip install numba==0.55.1 numpy==1.20.0 # revert inconsistent torch-points-kernel dependencies
-mamba install pytorch-lightning==1.5.9 -c conda-forge
+mamba install -y pytorch-lightning==1.5.9 -c conda-forge

bash/setup_environment/setup_env_cpu_only.sh

@@ -3,7 +3,7 @@ set -e
 # SETUP
 conda install -y mamba -n base -c conda-forge
 mamba env create -f bash/setup_environment/requirements.yml
-conda activate lidar_deep_segmentation
+conda activate lidar_multiclass
 
 # INSTALL
 export PYTORCHVERSION="1.10.1"

lidar_multiclass/datamodules/interpolation.py

@@ -2,7 +2,7 @@
 from tokenize import Number
 from typing import Dict, List, Optional
 
-import laspy
+import pdal
 import numpy as np
 import torch
 from torch_geometric.nn.pool import knn
@@ -47,32 +47,27 @@ def __init__(
 
     def _load_las(self, filepath: str):
         """Load a LAS and add necessary extradim."""
-
-        self.las = laspy.read(filepath)
         self.current_f = filepath
+        pipeline = pdal.Reader.las(filename=filepath)
 
-        coln = ChannelNames.PredictedClassification.value
-        param = laspy.ExtraBytesParams(name=coln, type=int)
-        self.las.add_extra_dim(param)
-        self.las[coln][:] = 0
-
-        param = laspy.ExtraBytesParams(
-            name=ChannelNames.ProbasEntropy.value, type=float
-        )
-        self.las.add_extra_dim(param)
-        self.las[ChannelNames.ProbasEntropy.value][:] = 0.0
-
-        for class_name in self.probas_names:
-            param = laspy.ExtraBytesParams(name=class_name, type=float)
-            self.las.add_extra_dim(param)
-            self.las[class_name][:] = 0.0
-
+        new_dims = self.probas_names + [
+            ChannelNames.PredictedClassification.value,
+            ChannelNames.ProbasEntropy.value,
+        ]
+        for new_dim in new_dims:
+            pipeline |= pdal.Filter.ferry(
+                dimensions=f"=>{new_dim}"
+            ) | pdal.Filter.assign(value=f"{new_dim}=0")
+        pipeline.execute()
+        self.las = pipeline.arrays[0]  # named array
+
+        # TODO: check scale of x versus X ?
         self.pos_las = torch.from_numpy(
             np.asarray(
                 [
-                    self.las.x,
-                    self.las.y,
-                    self.las.z,
+                    self.las["X"],
+                    self.las["Y"],
+                    self.las["Z"],
                 ],
                 dtype=np.float32,
             ).transpose()
@@ -177,7 +172,13 @@ def _write(self, interpolation):
         self.las[ChannelNames.ProbasEntropy.value][:] = entropy
 
         log.info(f"Saving...")
-        self.las.write(out_f)
+
+        pipeline = pdal.Writer.las(
+            filename=out_f,
+            extra_dims=f"all",
+            minor_version=4,
+        ).pipeline(self.las)
+        pipeline.execute()
         log.info(f"Saved.")
 
         return out_f

setup.py

@@ -6,7 +6,7 @@
 
 setup(
     name="lidar_multiclass",
-    version="1.6.3",
+    version="1.6.10",
     description="Multiclass Semantic Segmentation for Lidar Point Cloud",
     author="Charles GAYDON",
     author_email="",