Added model, training and data preparation features

.devcontainer/devcontainer.json

@@ -12,5 +12,3 @@
   "runArgs": ["--gpus", "all", "--shm-size=8g"],
   "remoteUser": "root"
 }
-
-

.dockerignore

@@ -11,4 +11,4 @@ htmlcov
 build/
 dist/
 *.egg-info
-*.egg
+*.egg

.env_example

@@ -7,4 +7,4 @@ COMET_PROJECT_NAME="project_name"
 COMET_WORKSPACE="workspace_name"
 COMET_GIT_DIRECTORY="path/to/dir/"
 
-HYDRA_FULL_ERROR=1
+HYDRA_FULL_ERROR=1

.gitignore

@@ -12,6 +12,10 @@ tmp/
 tests/data/toy_dataset.hdf5
 
 
+# new trained model and config
+trained_model_assets/LargeDataset-LidarHD-100mPatch_7cl_randlanet-Epoch_072.ckpt
+trained_model_assets/LargeDataset-LidarHD-100mPatch_7cl_randlanet-Epoch_072-inference-Myria3DV3.10.0.yaml
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

CHANGELOG.md

@@ -1,7 +1,15 @@
 # CHANGELOG
+## 3.10.0
+- Add config to handle 100 m x 100 m patches
+
+- Add Focal and Lovasz Losses and config
+
+- Add preprocessing steps during hdf5 creation as they do not change at batch selection level (for example,GridSampling will always select the same points as it is deterministic)
+
+- Add config parameters to set model size, i.e a larger model for a larger dataset
 
-### Not yet released
 - Check that las/laz output can be read by pdal when running prediction.
+
 - Clean mamba build to reduce Docker image size (14.6GB to 9.2GB)
 
 ### 3.9.1

README.md

@@ -20,12 +20,12 @@ Myria3D is a deep learning library designed with a focused scope: the multiclass
 The library implements the training of 3D Segmentation neural networks, with optimized data-processing and evaluation logics at fit time. Inference on unseen, large scale point cloud is also supported.
 It allows for the evaluation of single-class IoU on the full point cloud, which results in reliable model evaluation.
 
-Myria3D is built upon [PyTorch](https://pytorch.org/). It keeps the standard data format 
-from [Pytorch-Geometric](https://pytorch-geometric.readthedocs.io/). 
+Myria3D is built upon [PyTorch](https://pytorch.org/). It keeps the standard data format
+from [Pytorch-Geometric](https://pytorch-geometric.readthedocs.io/).
 Its structure was bootstraped from [this code template](https://github.com/ashleve/lightning-hydra-template),
 which heavily relies on [Hydra](https://hydra.cc/) and [Pytorch-Lightning](https://github.com/PyTorchLightning/pytorch-lightning) to enable flexible and rapid iterations of deep learning experiments.
 
-Although the library can be extended with new neural network architectures or new data signatures, it makes some opiniated choices in terms of neural network architecture, data processing logics, and inference logic. Indeed, it is initially built with the [French "Lidar HD" project](https://geoservices.ign.fr/lidarhd) in mind, with the ambition to map France in 3D with 10 pulse/m² aerial Lidar by 2025. The data will be openly available, including a semantic segmentation with a minimal number of classes: ground, vegetation, buildings, vehicles, bridges, others. 
+Although the library can be extended with new neural network architectures or new data signatures, it makes some opiniated choices in terms of neural network architecture, data processing logics, and inference logic. Indeed, it is initially built with the [French "Lidar HD" project](https://geoservices.ign.fr/lidarhd) in mind, with the ambition to map France in 3D with 10 pulse/m² aerial Lidar by 2025. The data will be openly available, including a semantic segmentation with a minimal number of classes: ground, vegetation, buildings, vehicles, bridges, others.
 
 > → For installation and usage, please refer to [**Documentation**](https://ignf.github.io/myria3d/).
 
@@ -41,4 +41,4 @@ Please cite Myria3D if it helped your own research. Here is an example BibTex en
   year={2022},
   note={IGN (French Mapping Agency)},
 }
-```
+```

configs/callbacks/default.yaml

@@ -1,4 +1,4 @@
-# Override with "none" to deactivate. 
+# Override with "none" to deactivate.
 
 log_code:
   _target_: myria3d.callbacks.comet_callbacks.LogCode
@@ -8,15 +8,15 @@ log_logs_dir:
   _target_: myria3d.callbacks.comet_callbacks.LogLogsPath
 
 lr_monitor:
-  _target_:  pytorch_lightning.callbacks.LearningRateMonitor
+  _target_: pytorch_lightning.callbacks.LearningRateMonitor
   logging_interval: "step"
   log_momentum: true
 
 model_checkpoint:
   _target_: pytorch_lightning.callbacks.ModelCheckpoint
   monitor: "val/loss_epoch" # name of the logged metric which determines when model is improving
   mode: "min" # can be "max" or "min"
-  save_top_k: 1 # save k best models (determined by above metric)
+  save_top_k: 5 # save k best models (determined by above metric)
   save_last: True # additionaly always save model from last epoch
   verbose: True
   dirpath: "checkpoints/"
@@ -29,7 +29,7 @@ early_stopping:
   mode: "min" # can be "max" or "min"
   patience: 6 # how many validation epochs of not improving until training stops
   min_delta: 0 # minimum change in the monitored metric needed to qualify as an improvement
-  
+
 model_detailed_metrics:
   _target_: myria3d.callbacks.metric_callbacks.ModelMetrics
-  num_classes: ${model.num_classes}
+  num_classes: ${model.num_classes}

configs/callbacks/finetuning.yaml

@@ -1,10 +1,10 @@
 defaults:
   - default.yaml
 
-# Override 
+# Override
 finetune:
   _target_: myria3d.callbacks.finetuning_callbacks.FinetuningFreezeUnfreeze
   d_in: ${model.d_in}
   num_classes: ${model.num_classes}
-  unfreeze_fc_end_epoch: ???  # full end fc is unfrozen later than the last MLP.
-  unfreeze_decoder_train_epoch: ???  # decoder is unfrozen later than last FC layer.
+  unfreeze_fc_end_epoch: ??? # full end fc is unfrozen later than the last MLP.
+  unfreeze_decoder_train_epoch: ??? # decoder is unfrozen later than last FC layer.

configs/config.yaml

@@ -23,19 +23,19 @@ ignore_warnings: True
 
 # specify here default training configuration
 defaults:
-  - _self_  # for hydra legacy reasons
+  - _self_ # for hydra legacy reasons
 
   - trainer: default.yaml
   - datamodule: hdf5_datamodule.yaml
-  - dataset_description: 20220607_151_dalles_proto.yaml  # describes input features and classes
+  - dataset_description: 20220607_151_dalles_proto.yaml # describes input features and classes
   - callbacks: default.yaml # set this to null if you don't want to use callbacks
   - model: pyg_randla_net_model.yaml
-  
+
   - logger: comet # set logger here or use command line (e.g. `python run.py logger=wandb`)
   - task: default.yaml
   - predict: default.yaml
 
-  - experiment: RandLaNetDebug.yaml  # default run is for debugging.
+  - experiment: RandLaNetDebug.yaml # default run is for debugging.
 
   - hydra: default.yaml
 

configs/datamodule/hdf5_datamodule.yaml

@@ -1,10 +1,10 @@
 _target_: myria3d.pctl.datamodule.hdf5.HDF5LidarDataModule
 
 # A standalone precomputed HDF5 dataset can be used without source files by setting data_dir and split_csv_path to null.
-data_dir: null  # data directory containing LAS files
+data_dir: null # data directory containing LAS files
 epsg: null
-split_csv_path: null  # csv specifying split, with schema basename, split (where split is one out of train/val/test)
-hdf5_file_path: "path/to/dataset_file.hdf5"  # where to create a HDF5 dataset file from LAS and CSV sources.
+split_csv_path: null # csv specifying split, with schema basename, split (where split is one out of train/val/test)
+hdf5_file_path: "path/to/dataset_file.hdf5" # where to create a HDF5 dataset file from LAS and CSV sources.
 
 # functions used to load and preprocess LAS data points into a pytorch geometric Data object.
 points_pre_transform:
@@ -14,7 +14,7 @@ points_pre_transform:
 
 pre_filter:
   _target_: functools.partial
-  _args_: 
+  _args_:
     - "${get_method:myria3d.pctl.dataset.utils.pre_filter_below_n_points}"
   min_num_nodes: 1
 
@@ -26,6 +26,8 @@ subtile_overlap_predict: "${predict.subtile_overlap}"
 batch_size: 32
 num_workers: 3
 prefetch_factor: 3
+pin_memory: true
+persistent_workers: true
 
 defaults:
   - transforms: default.yaml

configs/datamodule/transforms/augmentations/heavy.yaml

@@ -2,7 +2,6 @@
 defaults:
   - light.yaml
 
-
 # Rotation around axis - sample should be centered for this.
 RandomRotate:
   _target_: torch_geometric.transforms.RandomRotate

configs/datamodule/transforms/augmentations/light.yaml

@@ -3,9 +3,9 @@ x_flip:
   _args_:
     - 0
   p: 0.5
-  
+
 y_flip:
   _target_: torch_geometric.transforms.RandomFlip
   _args_:
     - 1
-  p: 0.5
+  p: 0.5

configs/datamodule/transforms/default.yaml

@@ -1,5 +1,5 @@
 defaults:
-  - preparations: points_budget.yaml
+  - preparations: points_budget_patch_100mx100m.yaml
   - augmentations: none.yaml
   - normalizations: default.yaml
 
@@ -13,4 +13,7 @@ preparations_train_list: "${oc.dict.values: datamodule.transforms.preparations.t
 preparations_eval_list: "${oc.dict.values: datamodule.transforms.preparations.eval}"
 preparations_predict_list: "${oc.dict.values: datamodule.transforms.preparations.predict}"
 
+# Deterministic transforms applied once, at HDF5-creation time, for the TRAIN split only.
+preparations_train_bake_list: "${oc.dict.values: datamodule.transforms.preparations.train_bake}"
+
 normalizations_list: "${oc.dict.values: datamodule.transforms.normalizations}"

configs/datamodule/transforms/normalizations/default.yaml

@@ -8,4 +8,4 @@ NormalizePos:
   subtile_width: "${datamodule.subtile_width}"
 
 StandardizeRGBAndIntensity:
-  _target_: myria3d.pctl.transforms.transforms.StandardizeRGBAndIntensity
+  _target_: myria3d.pctl.transforms.transforms.StandardizeRGBAndIntensity

configs/datamodule/transforms/preparations/fixed_num_points.yaml

@@ -1,7 +1,10 @@
 # default preparations with grid sampling and random sampling.
 
-train:
-
+# Deterministic transforms baked into the HDF5 file once, at dataset-creation time,
+# for the TRAIN split only (applied in create_hdf5). They are therefore NOT recomputed
+# every epoch, which is a large speed-up on big datasets.
+# NB: regenerate the HDF5 file after changing this list.
+train_bake:
   TargetTransform:
     _target_: myria3d.pctl.transforms.transforms.TargetTransform
     _args_:
@@ -16,6 +19,7 @@ train:
     _args_:
       - 0.25
 
+train:
   FixedPoints:
     _target_: torch_geometric.transforms.FixedPoints
     _args_:
@@ -27,7 +31,6 @@ train:
     _target_: torch_geometric.transforms.Center
 
 eval:
-
   CopyFullPos:
     _target_: myria3d.pctl.transforms.transforms.CopyFullPos
 
@@ -61,9 +64,8 @@ eval:
 
   Center:
     _target_: torch_geometric.transforms.Center
-
-predict:
 
+predict:
   CopyFullPos:
     _target_: myria3d.pctl.transforms.transforms.CopyFullPos
 
@@ -86,4 +88,4 @@ predict:
     _target_: myria3d.pctl.transforms.transforms.CopySampledPos
 
   Center:
-    _target_: torch_geometric.transforms.Center
+    _target_: torch_geometric.transforms.Center

configs/datamodule/transforms/preparations/points_budget.yaml

@@ -1,7 +1,17 @@
 # default preparations with grid sampling and random sampling.
 
-train:
-
+# Maximum number of nodes (points) kept per subtile after sampling.
+# This is the ONLY value that differs between points budget variants, so it is
+# exposed here as a single parameter. Override it on the CLI with:
+#   datamodule.transforms.preparations.max_num_nodes=120000
+# or inherit this file and override just this value (see points_budget_patch_100mx100m.yaml).
+max_num_nodes: 40000
+
+# Deterministic transforms baked into the HDF5 file once, at dataset-creation time,
+# for the TRAIN split only (applied in create_hdf5). They are therefore NOT recomputed
+# every epoch, which is a large speed-up on big datasets.
+# NB: regenerate the HDF5 file after changing this list.
+train_bake:
   TargetTransform:
     _target_: myria3d.pctl.transforms.transforms.TargetTransform
     _args_:
@@ -15,7 +25,8 @@ train:
     _target_: torch_geometric.transforms.GridSampling
     _args_:
       - 0.25
-
+
+train:
   MinimumNumNodes:
     _target_: myria3d.pctl.transforms.transforms.MinimumNumNodes
     _args_:
@@ -24,13 +35,12 @@ train:
   MaximumNumNodes:
     _target_: myria3d.pctl.transforms.transforms.MaximumNumNodes
     _args_:
-      - 40000
+      - ${datamodule.transforms.preparations.max_num_nodes}
 
   Center:
     _target_: torch_geometric.transforms.Center
 
 eval:
-
   TargetTransform:
     _target_: myria3d.pctl.transforms.transforms.TargetTransform
     _args_:
@@ -59,16 +69,15 @@ eval:
   MaximumNumNodes:
     _target_: myria3d.pctl.transforms.transforms.MaximumNumNodes
     _args_:
-      - 40000
+      - ${datamodule.transforms.preparations.max_num_nodes}
 
   CopySampledPos:
     _target_: myria3d.pctl.transforms.transforms.CopySampledPos
 
   Center:
     _target_: torch_geometric.transforms.Center
-
-predict:
 
+predict:
   DropPointsByClass:
     _target_: myria3d.pctl.transforms.transforms.DropPointsByClass
 
@@ -88,10 +97,10 @@ predict:
   MaximumNumNodes:
     _target_: myria3d.pctl.transforms.transforms.MaximumNumNodes
     _args_:
-      - 40000
+      - ${datamodule.transforms.preparations.max_num_nodes}
 
   CopySampledPos:
     _target_: myria3d.pctl.transforms.transforms.CopySampledPos
 
   Center:
-    _target_: torch_geometric.transforms.Center
+    _target_: torch_geometric.transforms.Center

configs/datamodule/transforms/preparations/points_budget_patch_100mx100m.yaml

@@ -0,0 +1,8 @@
+# Same preparations as points_budget.yaml, but with a larger points budget suited
+# to 100m x 100m patches. It inherits everything from points_budget.yaml and only
+# overrides the single value that differs between the two: max_num_nodes.
+defaults:
+  - points_budget
+  - _self_
+
+max_num_nodes: 120000

configs/dataset_description/20220204_BuildingValidation_and_Ground.yaml

@@ -1,9 +1,9 @@
-_convert_: all  # For omegaconf struct to be converted to python dictionnaries
+_convert_: all # For omegaconf struct to be converted to python dictionnaries
 # classification_preprocessing_dict = {source_class_code_int: target_class_code_int},
-classification_preprocessing_dict: {59: 6, 50: 1}
+classification_preprocessing_dict: { 59: 6, 50: 1 }
 # classification_dict = {code_int: name_str, ...} and MUST be sorted (increasing order).
-classification_dict: {1: "unclassified", 2: "ground", 6: "building"}
+classification_dict: { 1: "unclassified", 2: "ground", 6: "building" }
 
 # Input and output dims of neural net are dataset dependant:
 d_in: 12
-num_classes: 3
+num_classes: 3