Merge pull request #2 from CCInc/conf/trainer

humanpose1 · web-flow · commit 620278173bb4 · 2021-07-20T12:53:31.000+02:00
update
diff --git a/conf/config.yaml b/conf/config.yaml
@@ -0,0 +1,19 @@
+defaults:
+  - task: ???
+  - models: ???
+  - data: ???
+  - training: default
+  - trainer: default
+
+job_name: benchmark
+model_name: ???
+update_lr_scheduler_on: "on_epoch" # ["on_epoch", "on_num_batch", "on_num_sample"]
+selection_stage: ""
+pretty_print: False
+eval_frequency: 1
+
+tracker_options:
+  # Extra options for the tracker
+  full_res: False
+  make_submission: False
+  track_boxes: False
diff --git a/conf/data/default.yaml b/conf/data/default.yaml
diff --git a/conf/data/segmentation/default.yaml b/conf/data/segmentation/default.yaml
diff --git a/conf/data/segmentation/s3dis/s3dis1x1.yaml b/conf/data/segmentation/s3dis/s3dis1x1.yaml
diff --git a/conf/lr_scheduler/cosine.yaml b/conf/lr_scheduler/cosine.yaml
@@ -0,0 +1,4 @@
+# @package _group_
+class: CosineAnnealingLR
+params: 
+  T_max: 10
diff --git a/conf/lr_scheduler/cyclic.yaml b/conf/lr_scheduler/cyclic.yaml
@@ -0,0 +1,5 @@
+# @package _group_
+class: CyclicLR
+params:
+  base_lr: ${training.optim.base_lr}
+  max_lr: 0.1
diff --git a/conf/lr_scheduler/exponential.yaml b/conf/lr_scheduler/exponential.yaml
@@ -0,0 +1,4 @@
+# @package _group_
+class: ExponentialLR
+params:
+  gamma: 0.9885 # = 0.1**(1/200.) divide by 10 every 200 epochs
diff --git a/conf/lr_scheduler/multi_step.yaml b/conf/lr_scheduler/multi_step.yaml
@@ -0,0 +1,5 @@
+# @package _group_
+class: MultiStepLR
+params:
+  milestones: [80,120,160]
+  gamma: 0.2
diff --git a/conf/lr_scheduler/multi_step_reg.yaml b/conf/lr_scheduler/multi_step_reg.yaml
@@ -0,0 +1,5 @@
+# @package _group_
+class: MultiStepLR
+params:
+  milestones: [600, 1200, 1800, 3000]
+  gamma: 0.5
diff --git a/conf/lr_scheduler/plateau.yaml b/conf/lr_scheduler/plateau.yaml
@@ -0,0 +1,4 @@
+# @package _group_
+class: ReduceLROnPlateau
+  params: 
+    mode: "min"
diff --git a/conf/lr_scheduler/poly_lr.yaml b/conf/lr_scheduler/poly_lr.yaml
@@ -0,0 +1,9 @@
+# @package _group_
+class: PolyLR
+params:
+  on_epoch:
+    max_iter: 150
+    power: 0.9
+  on_num_batch:
+    max_iter: 60000
+    power: 2
diff --git a/conf/lr_scheduler/step.yaml b/conf/lr_scheduler/step.yaml
@@ -0,0 +1,6 @@
+# @package _group_
+class: StepLR
+params:
+  step_size: 10
+  gamma: 0.9
+  last_epoch: -1
diff --git a/conf/models/default.yaml b/conf/models/default.yaml
@@ -0,0 +1 @@
+# @package models
diff --git a/conf/models/segmentation/sparseconv3d.yaml b/conf/models/segmentation/sparseconv3d.yaml
@@ -0,0 +1,74 @@
+# @package models
+ResUNet32:
+  class: sparseconv3d.APIModel
+  conv_type: "SPARSE"
+  backend: "torchsparse"
+  backbone:
+    define_constants:
+      in_feat: 32
+      block: ResBlock # Can be any of the blocks in modules/MinkowskiEngine/api_modules.py
+    down_conv:
+      module_name: ResNetDown
+      block: block
+      N: [ 0, 1, 2, 2, 3 ]
+      down_conv_nn:
+        [
+          [ FEAT, in_feat ],
+          [ in_feat, in_feat ],
+          [ in_feat, 2*in_feat ],
+          [ 2*in_feat, 4*in_feat ],
+          [ 4*in_feat, 8*in_feat ],
+        ]
+      kernel_size: 3
+      stride: [ 1, 2, 2, 2, 2 ]
+    up_conv:
+      block: block
+      module_name: ResNetUp
+      N: [ 1, 1, 1, 1, 0 ]
+      up_conv_nn:
+        [
+          [ 8*in_feat, 4*in_feat ],
+          [ 4*in_feat + 4*in_feat, 4*in_feat ],
+          [ 4*in_feat + 2*in_feat, 3*in_feat ],
+          [ 3*in_feat + in_feat, 3*in_feat ],
+          [ 3*in_feat + in_feat, 3*in_feat ],
+        ]
+      kernel_size: 3
+      stride: [ 2, 2, 2, 2, 1 ]
+
+Res16UNet34:
+  class: sparseconv3d.APIModel
+  conv_type: "SPARSE"
+  backend: "torchsparse"
+  backbone:
+    define_constants:
+      in_feat: 32
+      block: ResBlock # Can be any of the blocks in modules/MinkowskiEngine/api_modules.py
+    down_conv:
+      module_name: ResNetDown
+      block: block
+      N: [ 0, 2, 3, 4, 6 ]
+      down_conv_nn:
+        [
+          [ FEAT, in_feat ],
+          [ in_feat, in_feat ],
+          [ in_feat, 2*in_feat ],
+          [ 2*in_feat, 4*in_feat ],
+          [ 4*in_feat, 8*in_feat ],
+        ]
+      kernel_size: 3
+      stride: [ 1, 2, 2, 2, 2 ]
+    up_conv:
+      block: block
+      module_name: ResNetUp
+      N: [ 1, 1, 1, 1, 1 ]
+      up_conv_nn:
+        [
+          [ 8*in_feat, 4*in_feat ],
+          [ 4*in_feat + 4*in_feat, 4*in_feat ],
+          [ 4*in_feat + 2*in_feat, 3*in_feat ],
+          [ 3*in_feat + in_feat, 3*in_feat ],
+          [ 3*in_feat + in_feat, 3*in_feat ],
+        ]
+      kernel_size: 3
+      stride: [ 2, 2, 2, 2, 1 ]
diff --git a/conf/optimizer/sgd.yaml b/conf/optimizer/sgd.yaml
@@ -0,0 +1,4 @@
+_target_: torch.optim.SGD
+lr: ${training.optim.base_lr}
+momentum: 0.9
+weight_decay: 0.0005
diff --git a/conf/trainer/default.yaml b/conf/trainer/default.yaml
@@ -0,0 +1,52 @@
+_target_: pytorch_lightning.Trainer
+logger: tensorboard
+checkpoint_callback: True
+default_root_dir: null
+gradient_clip_val: 0.0
+process_position: 0
+num_nodes: 1
+num_processes: 1
+gpus: null
+auto_select_gpus: False
+tpu_cores: null
+log_gpu_memory: null
+progress_bar_refresh_rate: 1
+overfit_batches: 0.0
+track_grad_norm: -1
+check_val_every_n_epoch: 1
+fast_dev_run: False
+accumulate_grad_batches: 1
+max_epochs: 10
+min_epochs: 1
+max_steps: null
+min_steps: null
+limit_train_batches: 128
+limit_val_batches: 2
+limit_test_batches: 2
+val_check_interval: 1.0
+flush_logs_every_n_steps: 100
+log_every_n_steps: 50
+accelerator: null
+sync_batchnorm: False
+precision: 32
+weights_summary: 'top'
+weights_save_path: null
+num_sanity_val_steps: 2
+truncated_bptt_steps: null
+resume_from_checkpoint: null
+profiler: null
+benchmark: False
+deterministic: False
+reload_dataloaders_every_epoch: False
+auto_lr_find: False
+replace_sampler_ddp: True
+terminate_on_nan: False
+auto_scale_batch_size: False
+prepare_data_per_node: True
+plugins: null
+amp_backend: 'native'
+amp_level: 'O2'
+distributed_backend: null
+automatic_optimization: null
+move_metrics_to_cpu: False
+enable_pl_optimizer: True 
diff --git a/conf/training/default.yaml b/conf/training/default.yaml
@@ -0,0 +1,43 @@
+# @package training
+# Those arguments defines the training hyper-parameters
+epochs: 100
+num_workers: 6
+batch_size: 16
+shuffle: True
+cuda: 0 # -1 -> no cuda otherwise takes the specified index
+precompute_multi_scale: False # Compute multiscate features on cpu for faster training / inference
+optim:
+  base_lr: 0.001
+  # accumulated_gradient: -1 # Accumulate gradient accumulated_gradient * batch_size
+  grad_clip: -1
+  optimizer:
+    class: Adam
+    params:
+      lr: ${training.optim.base_lr} # The path is cut from training
+  lr_scheduler: ${lr_scheduler}
+  bn_scheduler:
+    bn_policy: "step_decay"
+    params:
+      bn_momentum: 0.1
+      bn_decay: 0.9
+      decay_step: 10
+      bn_clip: 1e-2
+weight_name: "latest" # Used during resume, select with model to load from [miou, macc, acc..., latest]
+enable_cudnn: True
+checkpoint_dir: ""
+
+# Those arguments within experiment defines which model, dataset and task to be created for benchmarking
+# parameters for Weights and Biases
+wandb:
+  entity: ""
+  project: default
+  log: True
+  notes:
+  name:
+  public: True # It will be display the model within wandb log, else not.
+  config:
+    model_name: ${model_name}
+
+  # parameters for TensorBoard Visualization
+tensorboard:
+  log: True
diff --git a/lightning_transformers/trainer.py b/lightning_transformers/trainer.py
diff --git a/torch_points3d/.#trainer.py b/torch_points3d/.#trainer.py
@@ -0,0 +1 @@
+admincaor@admincaor.20945:1624465837
diff --git a/torch_points3d/core/instantiator.py b/torch_points3d/core/instantiator.py
diff --git a/torch_points3d/dataset/base_dataset.py b/torch_points3d/dataset/base_dataset.py
diff --git a/torch_points3d/dataset/geometric_dataset.py b/torch_points3d/dataset/geometric_dataset.py
diff --git a/torch_points3d/trainer.py b/torch_points3d/trainer.py
@@ -0,0 +1,43 @@
+from typing import Any, Optional
+
+import hydra
+from omegaconf import DictConfig, OmegaConf
+from pytorch_lightning import LightningDataModule
+from pytorch_lightning.utilities.distributed import rank_zero_info
+from hydra.utils.instantiate as hydra_instantiate
+
+from torch_points3d.datasets.dataset_factory import instantiate_dataset, convert_to_lightning_data_module
+from torch_points3d.datasets.dataset_factory import instantiate_dataset
+
+
+class LitTrainer:
+    def __init__(self, cfg: DictConfig):
+        self._cfg = cfg
+
+    def instantiate_trainer(self):
+        trainer = hydra_instantiate(self._cfg.trainer)
+        return trainer
+
+    def instantiate_dataset_and_model(self):
+        dataset: BaseDataset = instantiate_dataset(self._cfg.data)
+        model: BaseModel = instantiate_model(copy.deepcopy(cfg), dataset) 
+        model.instantiate_optimizers(cfg) # we will change it and instantiate the optimizers separately
+        model.set_pretrained_weights()
+        dataset.create_dataloaders(
+            cfg.training.batch_size,
+            cfg.training.shuffle,
+            cfg.training.num_workers,
+            model.conv_type == "PARTIAL_DENSE" and getattr(cfg.training, "precompute_multi_scale", False),
+        )
+        data_module = convert_to_lightning_data_module(dataset)
+        return model, data_module
+
+    def train(self):
+
+        
+        # model.tracker_options = cfg.get("tracker_options", {})
+        # model.trackers = data_module.trackers
+        model, data_module = self.instantiate_dataset_and_model()
+        trainer = self.instantiate_trainer()
+        trainer.fit(model, data_module)
+
