Added helper to ignore iscrowd

aldi/helpers.py

@@ -1,6 +1,8 @@
 import random
 import torch
 
+from detectron2.evaluation import COCOEvaluator
+
 
 class SaveIO:
     """Simple PyTorch hook to save the output of a nn.module."""
@@ -58,4 +60,20 @@ def backward(ctx, grad_output):
         return ctx.weight*grad_input, None
 
 def grad_reverse(x):
-    return _GradientScalarLayer.apply(x, -1.0)
+    return _GradientScalarLayer.apply(x, -1.0)
+
+def _maybe_add_iscrowd_annotations(cocoapi) -> None:
+    for ann in cocoapi.dataset["annotations"]:
+        if "iscrowd" not in ann:
+            ann["iscrowd"] = 0
+
+class Detectron2COCOEvaluatorAdapter(COCOEvaluator):
+    """A COCOEvaluator that makes iscrowd optional."""
+    def __init__(
+        self,
+        dataset_name,
+        output_dir=None,
+        distributed=True,
+    ):
+        super().__init__(dataset_name, output_dir=output_dir, distributed=distributed)
+        _maybe_add_iscrowd_annotations(self._coco_api)

aldi/trainer.py

@@ -5,7 +5,7 @@
 from detectron2.checkpoint.detection_checkpoint import DetectionCheckpointer
 from detectron2.data.build import build_detection_train_loader, get_detection_dataset_dicts
 from detectron2.engine import hooks, BestCheckpointer
-from detectron2.evaluation import COCOEvaluator, DatasetEvaluators
+from detectron2.evaluation import DatasetEvaluators
 from detectron2.modeling.meta_arch.build import build_model
 from detectron2.solver import build_optimizer
 from detectron2.utils.events import get_event_storage
@@ -17,6 +17,7 @@
 from aldi.distill import Distiller
 from aldi.dropin import DefaultTrainer, AMPTrainer, SimpleTrainer
 from aldi.dataloader import SaveWeakDatasetMapper, UnlabeledDatasetMapper, WeakStrongDataloader
+from aldi.helpers import Detectron2COCOEvaluatorAdapter
 from aldi.ema import EMA
 
 
@@ -158,7 +159,7 @@ def build_evaluator(cls, cfg, dataset_name, output_folder=None):
         """Just do COCO Evaluation."""
         if output_folder is None:
             output_folder = os.path.join(cfg.OUTPUT_DIR, "inference")
-        return DatasetEvaluators([COCOEvaluator(dataset_name, output_dir=output_folder)])
+        return DatasetEvaluators([Detectron2COCOEvaluatorAdapter(dataset_name, output_dir=output_folder)])
 
      def build_hooks(self):
           ret = super(DATrainer, self).build_hooks()

tools/slurm_train_net.sh

@@ -31,8 +31,8 @@ module load anaconda/2023a
 # run training script inside anaconda environment
 srun -N$SLURM_JOB_NUM_NODES bash -c "\
     CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES \
-    conda run --no-capture-output -n aldi \
-        python train_net.py \
+    conda run --no-capture-output -n aldi4 \
+        python tools/train_net.py \
             --machine-rank \$SLURM_PROCID \
             --num-machines $SLURM_JOB_NUM_NODES \
             --num-gpus 2 \