'code'

experiments/activitynet/dot+4conv.yaml

@@ -0,0 +1,97 @@
+WORKERS: 16
+
+MODEL_DIR: ./checkpoints/checkpoints-paper
+RESULT_DIR: ./infer_results
+LOG_DIR: ./log/log-paper
+DATA_DIR: ./data/ActivityNet
+EXPERIMENT_CFG: dot+4conv
+
+DATASET:
+  NAME: ActivityNet
+  VIS_INPUT_TYPE: c3d
+  NO_VAL: True
+  NUM_SAMPLE_CLIPS: 256
+  TARGET_STRIDE: 4
+  NORMALIZE: True
+  RANDOM_SAMPLING: False
+
+TEST:
+  BATCH_SIZE: 32
+  RECALL: 1,5
+  TIOU: 0.3,0.5,0.7
+  EVAL_TRAIN: False
+  NMS_THRESH: 0.5
+
+CUDNN:
+  DETERMINISTIC: False
+  BENCHMARK: True
+
+TRAIN:
+  BATCH_SIZE: 32
+  LR: 0.001
+  WEIGHT_DECAY: 0.0000
+  MAX_EPOCH: 15
+  CONTINUE: False
+  # STEPSIZE: 15
+  GAMMA: 0.1
+  MILESTONES: [5,9]
+
+LOSS1:
+  NAME: bce_rescale_loss
+  PARAMS:
+    MIN_IOU: 0.5
+    MAX_IOU: 1.0
+    BIAS: 0.0
+
+RANET:
+
+  ENCODER_LAYER:
+    NAME: FeatureEncoder
+    PARAMS:
+      FRAME:
+        INPUT_SIZE: 500
+        HIDDEN_SIZE: 512
+        KERNEL_SIZE: 4
+        STRIDE: 4
+        USE_POSITION: True
+        NUM_CLIPS: 64
+      GCNEXT:
+        INPUT_SIZE: 512
+        OUTPUT_SIZE: 512
+        K_NUM: 5
+        GROUP_NUM: 32
+        WIDTH_GROUP: 4
+      LSTM:
+        TXT_INPUT_SIZE: 300
+        TXT_HIDDEN_SIZE: 512
+        NUM_LAYERS: 3
+        BIDIRECTIONAL: True
+
+  GNERATOR_LAYER:
+    NAME: SparseBoundaryCat
+    PARAMS:
+      NUM_CLIPS: 64
+      NUM_SCALE_LAYERS: [15,8,8]
+
+  INTERACTOR_LAYER:
+    NAME: DotFuse
+    PARAMS:
+      HIDDEN_SIZE: 512
+      TXT_INPUT_SIZE: 512
+      VIS_INPUT_SIZE: 1024
+
+  RELATION_LAYER:
+    NAME: MapConv
+    PARAMS:
+      INPUT_SIZE: 512
+      HIDDEN_SIZES: [512, 512, 512, 512]
+      KERNEL_SIZES: [9, 9, 9, 9]
+      STRIDES: [1, 1, 1, 1]
+      PADDINGS: [16, 0, 0, 0]
+      DILATIONS: [1, 1, 1, 1]
+
+  PRED_INPUT_SIZE: 512
+
+MODEL:
+  NAME: RaNet
+  CHECKPOINT: ./checkpoints/best/ActivityNet/dot+4conv-2778.pkl

experiments/activitynet/dynamic+gat.yaml

@@ -0,0 +1,97 @@
+WORKERS: 16
+
+MODEL_DIR: ./checkpoints/checkpoints-paper
+RESULT_DIR: ./infer_results
+LOG_DIR: ./log/log-paper
+DATA_DIR: ./data/ActivityNet
+EXPERIMENT_CFG: dynamic+gat
+
+DATASET:
+  NAME: ActivityNet
+  VIS_INPUT_TYPE: c3d
+  NO_VAL: True
+  NUM_SAMPLE_CLIPS: 256
+  TARGET_STRIDE: 4
+  NORMALIZE: True
+  RANDOM_SAMPLING: False
+
+TEST:
+  BATCH_SIZE: 64
+  RECALL: 1,5
+  TIOU: 0.3,0.5,0.7
+  EVAL_TRAIN: False
+  NMS_THRESH: 0.52
+
+CUDNN:
+  DETERMINISTIC: False
+  BENCHMARK: True
+
+TRAIN:
+  BATCH_SIZE: 64
+  LR: 0.001
+  WEIGHT_DECAY: 0.0000
+  MAX_EPOCH: 15
+  CONTINUE: False
+  # STEPSIZE: 15
+  GAMMA: 0.1
+  MILESTONES: [5,9]
+
+LOSS1:
+  NAME: bce_rescale_loss
+  PARAMS:
+    MIN_IOU: 0.5
+    MAX_IOU: 1.0
+    BIAS: 0.0
+
+RANET:
+
+  ENCODER_LAYER:
+    NAME: FeatureEncoder
+    PARAMS:
+      FRAME:
+        INPUT_SIZE: 500
+        HIDDEN_SIZE: 512
+        KERNEL_SIZE: 4
+        STRIDE: 4
+        USE_POSITION: True
+        NUM_CLIPS: 64
+      GCNEXT:
+        INPUT_SIZE: 512
+        OUTPUT_SIZE: 512
+        K_NUM: 5
+        GROUP_NUM: 32
+        WIDTH_GROUP: 4
+      LSTM:
+        TXT_INPUT_SIZE: 300
+        TXT_HIDDEN_SIZE: 512
+        NUM_LAYERS: 3
+        BIDIRECTIONAL: True
+
+  GNERATOR_LAYER:
+    NAME: SparseBoundaryCat
+    PARAMS:
+      NUM_CLIPS: 64
+      NUM_SCALE_LAYERS: [15,8,8]
+
+  INTERACTOR_LAYER:
+    NAME: DynamicFuse
+    PARAMS:
+      HIDDEN_SIZE: 512
+      TXT_INPUT_SIZE: 512
+      VIS_INPUT_SIZE: 1024
+
+  RELATION_LAYER:
+    NAME: GATModule
+    PARAMS:
+      INPUT_SIZE: 512
+      OUTPUT_SIZE: 512
+      INTER_PROP: 4
+      BLOCK_NUM: 1
+      LOOP_NUM: 2
+
+  PRED_INPUT_SIZE: 512
+
+
+MODEL:
+  NAME: RaNet
+  CHECKPOINT: ./checkpoints/best/ActivityNet/dynamic+gat-2867.pkl

experiments/charades/i3d-finetune-dynamic+gat.yaml

@@ -0,0 +1,97 @@
+WORKERS: 16
+
+MODEL_DIR: ./checkpoints/checkpoints-paper
+RESULT_DIR: ./infer_results
+LOG_DIR: ./log/log-paper
+DATA_DIR: ./data/Charades-STA
+EXPERIMENT_CFG: i3d-finetune-dynamic+gat
+
+DATASET:
+  NAME: Charades
+  VIS_INPUT_TYPE: I3D/finetune
+  NO_VAL: True
+  NUM_SAMPLE_CLIPS: 256
+  TARGET_STRIDE: 4
+  NORMALIZE: True
+  RANDOM_SAMPLING: False
+
+TEST:
+  BATCH_SIZE: 32
+  RECALL: 1,5
+  TIOU: 0.3,0.5,0.7
+  EVAL_TRAIN: False
+  NMS_THRESH: 0.44
+
+CUDNN:
+  DETERMINISTIC: False
+  BENCHMARK: True
+
+TRAIN:
+  BATCH_SIZE: 32
+  LR: 0.001
+  WEIGHT_DECAY: 0.0000
+  MAX_EPOCH: 15
+  CONTINUE: False
+  STEPSIZE: 15
+  GAMMA: 0.1
+  MILESTONES: [7,12]
+
+LOSS1:
+  NAME: bce_rescale_loss
+  PARAMS:
+    MIN_IOU: 0.5
+    MAX_IOU: 1.0
+    BIAS: 0.0
+
+RANET:
+
+  ENCODER_LAYER:
+    NAME: FeatureEncoder
+    PARAMS:
+      FRAME:
+        INPUT_SIZE: 1024
+        HIDDEN_SIZE: 512
+        KERNEL_SIZE: 4
+        STRIDE: 4
+        USE_POSITION: False
+        NUM_CLIPS: 64
+      GCNEXT:
+        INPUT_SIZE: 512
+        OUTPUT_SIZE: 512
+        K_NUM: 5
+        GROUP_NUM: 32
+        WIDTH_GROUP: 4
+      LSTM:
+        TXT_INPUT_SIZE: 300
+        TXT_HIDDEN_SIZE: 512
+        NUM_LAYERS: 3
+        BIDIRECTIONAL: True
+
+  GNERATOR_LAYER:
+    NAME: SparseBoundaryCat
+    PARAMS:
+      NUM_CLIPS: 64
+      NUM_SCALE_LAYERS: [15,8,8]
+
+  INTERACTOR_LAYER:
+    NAME: DynamicFuse
+    PARAMS:
+      HIDDEN_SIZE: 512
+      TXT_INPUT_SIZE: 512
+      VIS_INPUT_SIZE: 1024
+
+  RELATION_LAYER:
+    NAME: GATModule
+    PARAMS:
+      INPUT_SIZE: 512
+      OUTPUT_SIZE: 512
+      INTER_PROP: 4
+      BLOCK_NUM: 2
+      LOOP_NUM: 2
+
+  PRED_INPUT_SIZE: 512
+
+
+MODEL:
+  NAME: RaNet
+  CHECKPOINT: ./checkpoints/best/Charades/i3d-dynamic+gat-3965.pkl