updated the head and hyperparameters

efficient_net_v2/build.py

@@ -198,6 +198,7 @@ def main(args):
         register_dataset(dataset_name='coco_2017_val_20', is_train=False)
 
     trainer = Trainer(cfg)
+
     # trainer = DefaultTrainer(cfg)
     trainer.resume_or_load(resume=args.resume)
     return trainer.train()

efficient_net_v2/config/effnet_coco.yaml

@@ -11,12 +11,17 @@ MODEL:
     IN_FEATURES: ["s2", "s3", "s5", "s6"]
   RPN:
     IN_FEATURES: ["p2", "p3", "p4", "p5"]
-    PRE_NMS_TOPK_TRAIN: 12000
-    POST_NMS_TOPK_TRAIN: 1000    
+    PRE_NMS_TOPK_TRAIN: 2000
+    POST_NMS_TOPK_TRAIN: 1000
+    CONV_DIMS: [256, 512, 256]
   PROPOSAL_GENERATOR:
-    NAME: "RPN"    
+    NAME: "RPN"
+    MIN_SIZE: 4
+  ANCHOR_GENERATOR:
+    SIZES: [[32], [64], [256], [512]]
+    ASPECT_RATIOS: [[0.5, 1.0, 2.0]] 
   RETINANET:
-    NUM_CLASSES: &num_classes 20
+    NUM_CLASSES: &num_classes 80
     IOU_THRESHOLDS: [0.4, 0.5]
     IOU_LABELS: [0, -1, 1]
     SMOOTH_L1_LOSS_BETA: 0.0
@@ -26,8 +31,8 @@ MODEL:
     NAME: "StandardROIHeads"
   ROI_BOX_HEAD:
     NAME: "FastRCNNConvFCHead"
-    NUM_FC: 2
-    NUM_CONV: 1
+    NUM_FC: 0
+    NUM_CONV: 3
     NORM: ""
     POOLER_RESOLUTION: 7
   ROI_MASK_HEAD:
@@ -41,19 +46,24 @@ MODEL:
   SEM_SEG_HEAD:
     NUM_CLASSES: *num_classes
   MASK_ON: False
-  LOAD_PROPOSALS: False    
+  LOAD_PROPOSALS: False
   WEIGHTS: ""    
 SOLVER:
-  IMS_PER_BATCH: 12
-  BASE_LR: 0.0005
-  STEPS: (60000, 800000)
+  AMP:
+    ENABLED: True
+  IMS_PER_BATCH: 20
+  BASE_LR: 0.02
+  STEPS: (30000, 60000)
   MAX_ITER: 100000
+  NESTEROV: False
   WARMUP_ITERS: 0
   CHECKPOINT_PERIOD: 5000
 DATALOADER:
   NUM_WORKERS: 8
 DATASETS:
-  TRAIN: ("coco_2017_train_20", )  
-  TEST: ("coco_2017_val_20", )
+  TRAIN: ("coco_2017_train", ) 
+  TEST: ("coco_2017_val", )
+  # PROPOSAL_FILES_TRAIN: ("detectron2://COCO-Detection/rpn_R_50_FPN_1x/137258492/coco_2017_train_box_proposals_21bc3a.pkl", )
+  # PROPOSAL_FILES_TEST: ("detectron2://COCO-Detection/rpn_R_50_FPN_1x/137258492/coco_2017_val_box_proposals_ee0dad.pkl", )
   # TRAIN: ("penn_fudan", )
   # TEST: ("penn_fudan_test",)

efficient_net_v2/datasets/coco_dataset.py

@@ -11,7 +11,7 @@
 
 logger = logging.getLogger(__name__)
 
-__labels__ = (
+LABELS = (
     'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus',
     'car', 'cat', 'chair', 'cow', 'diningtable', 'dog',
     'horse', 'motorbike', 'person', 'pottedplant', 'sheep',
@@ -35,7 +35,7 @@ def __post_init__(self):
         self.class_label = {}
         self.match_index = {}
         for i, line in enumerate(lines, 1):
-            if line not in __labels__:
+            if line not in LABELS:
                 continue
             self.class_label[line] = i
             self.match_index[i] = [len(self.match_index) + 1, line]
@@ -68,6 +68,6 @@ def __call__(self):
 
     import sys
     c = CocoDataset(coco_names=sys.argv[1])
-    
+
     import IPython
     IPython.embed()

efficient_net_v2/model/backbone.py

@@ -45,7 +45,6 @@ def output_shape(self):
         }
 
     def forward(self, x):
-        # return {'stage7': super().forward(x)}
         features = self.stage_forward(x)
         return {
             name: features[name] for name in self.out_features