Change datamodule input to [0,1] scale

mart/configs/batch_c15n/input_tensor_float01.yaml

@@ -0,0 +1,5 @@
+# Scale pixel input in the range of [0,1] to [0,255] for Adversary.
+defaults:
+  - tuple
+  - transform: times_255_and_round
+  - transform@untransform: divided_by_255

mart/configs/batch_c15n/input_tuple_float01.yaml

@@ -0,0 +1,11 @@
+# Scale pixel input in the range of [0,1] to [0,255] for Adversary.
+defaults:
+  - tuple
+  - transform@transform.transforms: times_255_and_round
+  - transform@untransform.transforms: divided_by_255
+
+transform:
+  _target_: mart.transforms.TupleTransforms
+
+untransform:
+  _target_: mart.transforms.TupleTransforms

mart/configs/batch_c15n/transform/divided_by_255.yaml

@@ -0,0 +1,3 @@
+_target_: torchvision.transforms.Normalize
+mean: 0
+std: 255

mart/configs/batch_c15n/transform/times_255_and_round.yaml

@@ -0,0 +1,13 @@
+_target_: torchvision.transforms.Compose
+transforms:
+  - _target_: mart.transforms.Denormalize
+    center: 0
+    scale: 255
+  # Fix potential numeric error.
+  - _target_: torch.fake_quantize_per_tensor_affine
+    _partial_: true
+    # (x/1+0).round().clamp(0, 255) * 1
+    scale: 1
+    zero_point: 0
+    quant_min: 0
+    quant_max: 255

mart/configs/datamodule/cifar10.yaml

@@ -17,16 +17,6 @@ train_dataset:
         padding_mode: reflect
       - _target_: torchvision.transforms.RandomHorizontalFlip
       - _target_: torchvision.transforms.ToTensor
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
   target_transform: null
   download: true
 
@@ -38,16 +28,6 @@ val_dataset:
     _target_: torchvision.transforms.Compose
     transforms:
       - _target_: torchvision.transforms.ToTensor
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
   target_transform: null
   download: true
 

mart/configs/datamodule/coco.yaml

@@ -12,16 +12,6 @@ train_dataset:
       - _target_: mart.transforms.ConvertCocoPolysToMask
       - _target_: mart.transforms.RandomHorizontalFlip
         p: 0.5
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 val_dataset:
   _target_: mart.datamodules.coco.CocoDetection
@@ -32,16 +22,6 @@ val_dataset:
     transforms:
       - _target_: torchvision.transforms.ToTensor
       - _target_: mart.transforms.ConvertCocoPolysToMask
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 test_dataset:
   _target_: mart.datamodules.coco.CocoDetection
@@ -52,16 +32,6 @@ test_dataset:
     transforms:
       - _target_: torchvision.transforms.ToTensor
       - _target_: mart.transforms.ConvertCocoPolysToMask
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 num_workers: 2
 collate_fn:

mart/configs/datamodule/coco_perturbable_mask.yaml

@@ -10,16 +10,6 @@ train_dataset:
       - _target_: mart.transforms.RandomHorizontalFlip
         p: 0.5
       - _target_: mart.transforms.ConvertInstanceSegmentationToPerturbable
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 val_dataset:
   transforms:
@@ -28,16 +18,6 @@ val_dataset:
       # ConvertCocoPolysToMask must be prior to ConvertInstanceSegmentationToPerturbable.
       - _target_: mart.transforms.ConvertCocoPolysToMask
       - _target_: mart.transforms.ConvertInstanceSegmentationToPerturbable
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 test_dataset:
   transforms:
@@ -46,13 +26,3 @@ test_dataset:
       # ConvertCocoPolysToMask must be prior to ConvertInstanceSegmentationToPerturbable.
       - _target_: mart.transforms.ConvertCocoPolysToMask
       - _target_: mart.transforms.ConvertInstanceSegmentationToPerturbable
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255

mart/configs/datamodule/fiftyone.yaml

@@ -14,16 +14,6 @@ train_dataset:
       - _target_: mart.transforms.ConvertCocoPolysToMask
       - _target_: mart.transforms.RandomHorizontalFlip
         p: 0.5
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 val_dataset:
   _target_: mart.datamodules.fiftyone.FiftyOneDataset
@@ -36,16 +26,6 @@ val_dataset:
     transforms:
       - _target_: torchvision.transforms.ToTensor
       - _target_: mart.transforms.ConvertCocoPolysToMask
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 test_dataset:
   _target_: mart.datamodules.fiftyone.FiftyOneDataset
@@ -58,16 +38,6 @@ test_dataset:
     transforms:
       - _target_: torchvision.transforms.ToTensor
       - _target_: mart.transforms.ConvertCocoPolysToMask
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 num_workers: 2
 collate_fn:

mart/configs/datamodule/fiftyone_perturbable_mask.yaml

@@ -16,16 +16,6 @@ train_dataset:
       - _target_: mart.transforms.RandomHorizontalFlip
         p: 0.5
       - _target_: mart.transforms.ConvertInstanceSegmentationToPerturbable
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 val_dataset:
   _target_: mart.datamodules.fiftyone.FiftyOneDataset
@@ -40,16 +30,6 @@ val_dataset:
       # ConvertCocoPolysToMask must be prior to ConvertInstanceSegmentationToPerturbable.
       - _target_: mart.transforms.ConvertCocoPolysToMask
       - _target_: mart.transforms.ConvertInstanceSegmentationToPerturbable
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 test_dataset:
   _target_: mart.datamodules.fiftyone.FiftyOneDataset
@@ -64,13 +44,3 @@ test_dataset:
       # ConvertCocoPolysToMask must be prior to ConvertInstanceSegmentationToPerturbable.
       - _target_: mart.transforms.ConvertCocoPolysToMask
       - _target_: mart.transforms.ConvertInstanceSegmentationToPerturbable
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255

mart/configs/datamodule/imagenet.yaml

@@ -13,16 +13,6 @@ train_dataset:
         size: 224
       - _target_: torchvision.transforms.RandomHorizontalFlip
       - _target_: torchvision.transforms.ToTensor
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 val_dataset:
   _target_: torchvision.datasets.ImageNet
@@ -34,16 +24,6 @@ val_dataset:
       - _target_: torchvision.transforms.CenterCrop
         size: 224
       - _target_: torchvision.transforms.ToTensor
-      - _target_: mart.transforms.Denormalize
-        center: 0
-        scale: 255
-      - _target_: torch.fake_quantize_per_tensor_affine
-        _partial_: true
-        # (x/1+0).round().clamp(0, 255) * 1
-        scale: 1
-        zero_point: 0
-        quant_min: 0
-        quant_max: 255
 
 test_dataset: ${.val_dataset}
 

mart/configs/experiment/CIFAR10_CNN_Adv.yaml

@@ -3,6 +3,7 @@
 defaults:
   - /attack@callbacks.adversary_connector.train_adversary: classification_fgsm_linf
   - /attack@callbacks.adversary_connector.test_adversary: classification_pgd_linf
+  - override /batch_c15n@callbacks.adversary_connector.batch_c15n: input_tensor_float01
   - override /datamodule: cifar10
   - override /model: classifier_cifar10_cnn
   - override /metric: accuracy

mart/configs/experiment/COCO_TorchvisionFasterRCNN_Adv.yaml

@@ -3,6 +3,7 @@
 defaults:
   - COCO_TorchvisionFasterRCNN
   - /attack@callbacks.adversary_connector.test_adversary: object_detection_mask_adversary
+  - override /batch_c15n@callbacks.adversary_connector.batch_c15n: input_tuple_float01
   - override /datamodule: coco_perturbable_mask
   - override /callbacks: [model_checkpoint, lr_monitor, adversary_connector]
 

mart/configs/model/classifier_cifar10_cnn.yaml

@@ -4,7 +4,7 @@ defaults:
 
 modules:
   preprocessor:
-    # Normalize [0, 255] input.
+    # Normalize [0, 1] input.
     _target_: torchvision.transforms.Normalize
-    mean: [125.307, 122.961, 113.8575]
-    std: [51.5865, 50.847, 51.255]
+    mean: [0.4914, 0.4822, 0.4465]
+    std: [0.2023, 0.1994, 0.2010]

mart/configs/model/classifier_timm.yaml

@@ -3,10 +3,11 @@ defaults:
 
 modules:
   preprocessor:
-    # Convert [0, 255] input to [0, 1]
+    # Use normalization parameters from timm
+    # https://github.com/huggingface/pytorch-image-models/blob/3d05c0e86f2f4c57bf495468aa3f8a7d3487c986/timm/data/constants.py#L3
     _target_: torchvision.transforms.Normalize
-    mean: 0
-    std: 255
+    mean: [0.485, 0.456, 0.406]
+    std: [0.229, 0.224, 0.225]
 
   logits:
     _target_: timm.models.convnext.convnext_tiny

mart/configs/model/torchvision_faster_rcnn.yaml

@@ -11,12 +11,9 @@ training_step_log:
 
 training_sequence:
   seq010:
-    preprocessor: ["input"]
+    losses_and_detections: ["input", "target"]
 
   seq020:
-    losses_and_detections: ["preprocessor", "target"]
-
-  seq030:
     loss:
       # Sum up the losses.
       [
@@ -28,17 +25,11 @@ training_sequence:
 
 validation_sequence:
   seq010:
-    preprocessor: ["input"]
-
-  seq020:
-    losses_and_detections: ["preprocessor", "target"]
+    losses_and_detections: ["input", "target"]
 
 test_sequence:
   seq010:
-    preprocessor: ["input"]
-
-  seq020:
-    losses_and_detections: ["preprocessor", "target"]
+    losses_and_detections: ["input", "target"]
 
 modules:
   losses_and_detections:

mart/configs/model/torchvision_object_detection.yaml

@@ -12,13 +12,6 @@ test_sequence: ???
 output_preds_key: "losses_and_detections.eval"
 
 modules:
-  preprocessor:
-    _target_: mart.transforms.TupleTransforms
-    transforms:
-      _target_: torchvision.transforms.Normalize
-      mean: 0
-      std: 255
-
   losses_and_detections:
     # Return losses in the training mode and predictions in the eval mode in one pass.
     _target_: mart.models.DualMode