add video training and infer scripts for GLEE-Pro and fix improve infer speed

projects/GLEE/configs/videos/Pro/BURST_Pro.yaml

@@ -3,6 +3,7 @@ MODEL:
   CROSS_TRACK: False
   PSEUDO_VIDEO: False
   FREEZE_WHOLE: False
+  VIDEO_WINDOW_SIZE: 2
   BACKBONE:
     NAME: "D2_EVA02"
   EVA02:

projects/GLEE/configs/videos/Pro/TAO_Pro.yaml

@@ -3,6 +3,7 @@ MODEL:
   CROSS_TRACK: False
   PSEUDO_VIDEO: False
   FREEZE_WHOLE: False
+  VIDEO_WINDOW_SIZE: 2
   BACKBONE:
     NAME: "D2_EVA02"
   EVA02:

projects/GLEE/configs/videos/Pro/ovis_Pro.yaml

@@ -3,6 +3,7 @@ MODEL:
   CROSS_TRACK: False
   PSEUDO_VIDEO: False
   FREEZE_WHOLE: False
+  VIDEO_WINDOW_SIZE: 2
   BACKBONE:
     NAME: "D2_EVA02"
   EVA02:

projects/GLEE/glee/GLEE.py

@@ -245,6 +245,7 @@ class GLEE(nn.Module):
         self.num_queries = cfg.MODEL.MaskDINO.NUM_OBJECT_QUERIES
         self.instance_on = True
         self.visaul_prompt = cfg.MODEL.VISUAL_PROMPT
+        self.video_window_size = cfg.MODEL.VIDEO_WINDOW_SIZE
 
         self.is_lsj = cfg.INPUT.DATASET_MAPPER_NAME == 'coco_instance_lsj'
 
@@ -754,7 +755,7 @@ class GLEE(nn.Module):
         video_len = len(batched_inputs[0]['file_names'])
 
 
-        clip_length = 5 # self.batch_infer_len
+        clip_length = self.video_window_size
         batch_name_list = self.dataset_name_dicts[task]
 
         #split long video into clips to form a batch input 

projects/GLEE/glee/config.py

@@ -21,6 +21,7 @@ def add_glee_config(cfg):
     cfg.MODEL.CROSS_TRACK = False
     cfg.MODEL.TRACK_VERSION = 'v3'
     cfg.MODEL.EARLYFUSION = True
+    cfg.MODEL.VIDEO_WINDOW_SIZE = 10
     
     cfg.INPUT.SAMPLING_FRAME_NUM = 1
     cfg.INPUT.SAMPLING_FRAME_RANGE = 10