Better Detectors

mart/configs/attack/gain/rcnn_training_loss.yaml

@@ -1,8 +1,4 @@
 _target_: mart.nn.CallWith
 module:
   _target_: mart.nn.Sum
-_call_with_args_:
-  - "losses_and_detections.training.loss_objectness"
-  - "losses_and_detections.training.loss_rpn_box_reg"
-  - "losses_and_detections.training.loss_classifier"
-  - "losses_and_detections.training.loss_box_reg"
+_call_with_args_: ${model.training_sequence.seq100.loss}

mart/configs/experiment/ArmoryCarlaOverObjDet_TorchvisionFasterRCNN.yaml

@@ -11,11 +11,12 @@ tags: ["regular_training"]
 optimized_metric: "test_metrics/map"
 
 model:
+  load_state_dict: null
+
   modules:
-    losses_and_detections:
-      model:
+    roi_heads:
+      box_predictor:
         num_classes: 3
-        weights: null
 
   optimizer:
     lr: 0.0125

mart/configs/experiment/COCO_TorchvisionFasterRCNN.yaml

@@ -28,13 +28,6 @@ datamodule:
   world_size: 1
 
 model:
-  modules:
-    losses_and_detections:
-      model:
-        # Inferred by torchvision.
-        num_classes: null
-        weights: COCO_V1
-
   optimizer:
     lr: 0.0125
     momentum: 0.9

mart/configs/experiment/COCO_TorchvisionRetinaNet.yaml

@@ -20,19 +20,18 @@ callbacks:
 trainer:
   # 117,266 training images, 6 epochs, batch_size=2, 351798
   max_steps: 351798
-  precision: 16
+  # FIXME: "nms_kernel" not implemented for 'BFloat16', torch.ops.torchvision.nms().
+  precision: 32
 
 datamodule:
   ims_per_batch: 2
   world_size: 1
 
 model:
   modules:
-    losses_and_detections:
-      model:
-        # Inferred by torchvision.
-        num_classes: null
-        weights: COCO_V1
+    detector:
+      num_classes: null # inferred by torchvision
+      weights: COCO_V1
 
   optimizer:
     lr: 0.0125

mart/configs/model/torchvision_dual_faster_rcnn.yaml

@@ -0,0 +1,82 @@
+# We simply wrap a torchvision object detection model for validation.
+defaults:
+  - torchvision_object_detection
+
+# log all losses separately in losses.
+training_step_log:
+  loss_objectness: "losses.loss_objectness"
+  loss_rpn_box_reg: "losses.loss_rpn_box_reg"
+  loss_classifier: "losses.loss_classifier"
+  loss_box_reg: "losses.loss_box_reg"
+
+training_sequence:
+  seq010:
+    preprocessor: ["input"]
+
+  seq020:
+    detector:
+      _name_: losses
+      _train_mode_: True
+      images: preprocessor
+      targets: target
+
+  seq030:
+    detector:
+      _name_: preds
+      _train_mode_: False
+      _inference_mode_: True
+      images: preprocessor
+      targets: target
+
+  seq100:
+    loss:
+      # Sum up the losses.
+      [
+        "losses.loss_objectness",
+        "losses.loss_rpn_box_reg",
+        "losses.loss_classifier",
+        "losses.loss_box_reg",
+      ]
+
+validation_sequence:
+  seq010:
+    preprocessor: ["input"]
+
+  seq020:
+    detector:
+      _name_: losses
+      _train_mode_: True
+      images: preprocessor
+      targets: target
+
+  seq030:
+    detector:
+      _name_: preds
+      _train_mode_: False
+      _inference_mode_: True
+      images: preprocessor
+      targets: target
+
+test_sequence:
+  seq010:
+    preprocessor: ["input"]
+
+  seq020:
+    detector:
+      _name_: losses
+      _train_mode_: True
+      images: preprocessor
+      targets: target
+
+  seq030:
+    detector:
+      _name_: preds
+      _train_mode_: False
+      _inference_mode_: True
+      images: preprocessor
+      targets: target
+
+modules:
+  detector:
+    _target_: torchvision.models.detection.fasterrcnn_resnet50_fpn
+    weights: "COCO_V1"