Update FasterRCNNProposalGenerator & Merge Detectron2 config

allennlp/common/detectron.py

@@ -9,13 +9,36 @@ class DetectronConfig(NamedTuple):
     overrides: Optional[Dict[str, Any]] = None
 
 
-def apply_dict_to_cfg(d: Dict[str, Any], c: CfgNode) -> None:
-    for key, value in d.items():
-        key = key.upper()
-        if isinstance(value, dict):
-            apply_dict_to_cfg(value, c[key])
+# def apply_dict_to_cfg(d: Dict[str, Any], c: CfgNode) -> None:
+#     for key, value in d.items():
+#         key = key.upper()
+#         if isinstance(value, dict):
+#             try:
+#                 apply_dict_to_cfg(value, c[key])
+#             except KeyError:
+#                 c[key] = {}
+#                 apply_dict_to_cfg(value, c[key])
+#         else:
+#             c[key] = value
+
+import collections.abc
+def update(d, u):
+    for k, v in u.items():
+        if isinstance(v, collections.abc.Mapping):
+            d[k] = update(d.get(k, {}), v)
         else:
-            c[key] = value
+            d[k] = v
+    return d
+
+
+def convert_to_dict(cfg_node, key_list = []):
+    if not isinstance(cfg_node, CfgNode):
+        return cfg_node
+    else:
+        cfg_dict = dict(cfg_node)
+        for k, v in cfg_dict.items():
+            cfg_dict[k] = convert_to_dict(v, key_list + [k])
+        return cfg_dict
 
 
 def get_detectron_cfg(
@@ -27,52 +50,62 @@ def get_detectron_cfg(
     from detectron2.config import get_cfg
 
     cfg = get_cfg()
-    add_attribute_config(cfg)  # TODO: What is this? It's not general, not configurable.
+    # add_attribute_config(cfg)  # TODO: What is this? It's not general, not configurable.
     from detectron2.model_zoo import get_config_file
 
     if builtin_config_file is not None:
         cfg.merge_from_file(get_config_file(builtin_config_file))
     if yaml_config_file is not None:
         cfg.merge_from_file(yaml_config_file)
     if overrides is not None:
-        apply_dict_to_cfg(overrides, cfg)
-    import torch
+        # print('old_dict')
+        # from pprint import pprint
+        old_dict = convert_to_dict(cfg, [])
+        # old_string = cfg.dump()
+        # pprint(old_dict)
+        new_dict = update(old_dict, overrides)
+        # print('new_dict')
+        # pprint(new_dict)
+        cfg = CfgNode(new_dict)
 
+        # apply_dict_to_cfg(overrides, cfg)
+
+    import torch
     if not torch.cuda.is_available():
         cfg.MODEL.DEVICE = "cpu"
     if freeze:
         cfg.freeze()
     return cfg
 
 
-def add_attribute_config(cfg):
-    """
-    Add config for attribute prediction.
-    """
-    # What feature type we want, region or grid:
-    cfg.MODEL.FEATURE_TYPE = "region"
-    # Whether to have attribute prediction
-    cfg.MODEL.ATTRIBUTE_ON = False
-    # Maximum number of attributes per foreground instance
-    cfg.INPUT.MAX_ATTR_PER_INS = 16
-    # ------------------------------------------------------------------------ #
-    # Attribute Head
-    # -----------------------------------------------------------------------  #
-    cfg.MODEL.ROI_ATTRIBUTE_HEAD = CfgNode()
-    # Dimension for object class embedding, used in conjunction with
-    # visual features to predict attributes
-    cfg.MODEL.ROI_ATTRIBUTE_HEAD.OBJ_EMBED_DIM = 256
-    # Dimension of the hidden fc layer of the input visual features
-    cfg.MODEL.ROI_ATTRIBUTE_HEAD.FC_DIM = 512
-    # Loss weight for attribute prediction, 0.2 is best per analysis
-    cfg.MODEL.ROI_ATTRIBUTE_HEAD.LOSS_WEIGHT = 0.2
-    # Number of classes for attributes
-    cfg.MODEL.ROI_ATTRIBUTE_HEAD.NUM_CLASSES = 400
-
-    """
-    Add config for box regression loss adjustment.
-    """
-    # Loss weights for RPN box regression
-    cfg.MODEL.RPN.BBOX_LOSS_WEIGHT = 1.0
-    # Loss weights for R-CNN box regression
-    cfg.MODEL.ROI_BOX_HEAD.BBOX_LOSS_WEIGHT = 1.0
+# def add_attribute_config(cfg):
+#     """
+#     Add config for attribute prediction.
+#     """
+#     # What feature type we want, region or grid:
+#     cfg.MODEL.FEATURE_TYPE = "region"
+#     # Whether to have attribute prediction
+#     cfg.MODEL.ATTRIBUTE_ON = False
+#     # Maximum number of attributes per foreground instance
+#     cfg.INPUT.MAX_ATTR_PER_INS = 16
+#     # ------------------------------------------------------------------------ #
+#     # Attribute Head
+#     # -----------------------------------------------------------------------  #
+#     cfg.MODEL.ROI_ATTRIBUTE_HEAD = CfgNode()
+#     # Dimension for object class embedding, used in conjunction with
+#     # visual features to predict attributes
+#     cfg.MODEL.ROI_ATTRIBUTE_HEAD.OBJ_EMBED_DIM = 256
+#     # Dimension of the hidden fc layer of the input visual features
+#     cfg.MODEL.ROI_ATTRIBUTE_HEAD.FC_DIM = 512
+#     # Loss weight for attribute prediction, 0.2 is best per analysis
+#     cfg.MODEL.ROI_ATTRIBUTE_HEAD.LOSS_WEIGHT = 0.2
+#     # Number of classes for attributes
+#     cfg.MODEL.ROI_ATTRIBUTE_HEAD.NUM_CLASSES = 400
+
+#     """
+#     Add config for box regression loss adjustment.
+#     """
+#     # Loss weights for RPN box regression
+#     cfg.MODEL.RPN.BBOX_LOSS_WEIGHT = 1.0
+#     # Loss weights for R-CNN box regression
+#     cfg.MODEL.ROI_BOX_HEAD.BBOX_LOSS_WEIGHT = 1.0

allennlp/modules/vision/__init__.py

@@ -3,6 +3,6 @@
 from allennlp.modules.vision.grid_embedder import GridEmbedder
 from allennlp.modules.vision.image2image import Image2ImageModule
 from allennlp.modules.vision.proposal_embedder import NullProposalEmbedder
-from allennlp.modules.vision.proposal_generator import RPNProposalGenerator
+from allennlp.modules.vision.proposal_generator import FasterRCNNProposalGenerator
 from allennlp.modules.vision.grid_embedder import NullGridEmbedder
 from allennlp.modules.vision.image2image import NullImage2ImageModule