[feat] support for input augmentation bypass mode for Detectron2Dataset

Author: chyomin06

cfgs/dataset/default.yaml

@@ -8,6 +8,7 @@ config:
   seqinfo: "seqinfo.ini"
   ext: "png"
 settings:
+  input_augmentation_bypass: False
   linear_mapper: False
   patch_size: [512, 512]
   ret_name: False

cfgs/vision_model/default.yaml

@@ -27,12 +27,6 @@ mask_rcnn_R_50_FPN_3x:
   integer_conv_weight: False
   splits : "r2" #, "c2" or "fpn"
 
-sam_vit_h_4b8939:
-  model_path_prefix: ${..model_root_path}
-  cfg: "Built-in configurations"
-  weights: "weights/segment_anything/sam_vit_h_4b8939.pth"
-  splits: "imgenc" 
-
 mask_rcnn_X_101_32x8d_FPN_3x:
   model_path_prefix: ${..model_root_path}
   cfg: "models/detectron2/configs/COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml"
@@ -47,6 +41,12 @@ panoptic_rcnn_R_101_FPN_3x:
   integer_conv_weight: False
   splits : "fpn"
 
+sam_vit_h_4b8939:
+  model_path_prefix: ${..model_root_path}
+  cfg: "Built-in configurations"
+  weights: "weights/segment_anything/sam_vit_h_4b8939.pth"
+  splits: "imgenc"
+
 jde_1088x608:
   model_path_prefix: ${..model_root_path}
   cfg: "models/Towards-Realtime-MOT/cfg/yolov3_1088x608.cfg"

compressai_vision/datasets/image.py

@@ -31,18 +31,17 @@
 import base64
 import logging
 import re
-
 from glob import glob
 from pathlib import Path
 from typing import Dict, List
 
 import numpy as np
-
 from detectron2.data import DatasetCatalog, MetadataCatalog
 from detectron2.data.common import DatasetFromList, MapDataset
 from detectron2.data.dataset_mapper import DatasetMapper
 from detectron2.data.datasets import load_coco_json, register_coco_instances
 from detectron2.data.samplers import InferenceSampler
+from detectron2.data.transforms import AugmentationList
 from detectron2.utils.serialize import PicklableWrapper
 from jde.utils.io import read_results
 from PIL import Image
@@ -118,6 +117,14 @@ def __init__(self, root, dataset_name, imgs_folder, **kwargs):
             if kwargs["seqinfo"].lower() != "none":
                 self.seqinfo_path = kwargs["dataset"].seqinfo_path
 
+        self.input_agumentation_bypass = False
+        if "input_augmentation_bypass" in kwargs:
+            self.input_agumentation_bypass = kwargs["input_augmentation_bypass"]
+            if self.input_agumentation_bypass:
+                self.logger.warning(
+                    "The vision model may or may not support the feature of input agumentation bypass\n"
+                )
+
         self.images_folder = Path(root) / imgs_folder
         assert self.images_folder == kwargs["dataset"].imgs_folder_path
 
@@ -251,9 +258,15 @@ def __init__(self, root, dataset_name, imgs_folder, **kwargs):
             ), "A proper mapper information via cfg must be provided"
             mapper = DatasetMapper(kwargs["cfg"], False)
 
-        self.mapDataset = MapDataset(_dataset, mapper)
         self._org_mapper_func = PicklableWrapper(DatasetMapper(kwargs["cfg"], False))
 
+        if self.input_agumentation_bypass:
+            emptyAugList = AugmentationList([])
+            self._org_mapper_func.augmentations = emptyAugList
+            mapper.augmentations = emptyAugList
+
+        self.mapDataset = MapDataset(_dataset, mapper)
+
         metaData = MetadataCatalog.get(dataset_name)
         try:
             self.thing_classes = metaData.thing_classes