Demo.py

.gitignore

@@ -138,13 +138,16 @@ dmypy.json
 .pyre/
 
 .idea/
-weights/
 *.pth
 *.onnx
 
 .vscode/
 output/
 datasets/
-weights/
 vendor/vendor/
 vendor/
+log2.md
+vendor/
+.DS_Store
+wandb/
+artifacts/

configs/Base-YOLOv7.yaml

@@ -1,5 +1,5 @@
 MODEL:
-  META_ARCHITECTURE: "YOLO" # default is YOLO, can be YOLOV7, YOLOX, YOLOMASK as well
+  META_ARCHITECTURE: "YOLOV7" # default is YOLO, can be YOLOV7, YOLOX, YOLOMASK as well
   PIXEL_MEAN: [0.406, 0.485, 0.456] # same value as PP-YOLOv2, BGR order
   PIXEL_STD: [0.225, 0.229, 0.224]
   PADDED_VALUE: 114.0

configs/canaries/detrt_256_6_6_regnetx_0.4g.yaml

@@ -0,0 +1,57 @@
+MODEL:
+  META_ARCHITECTURE: "Detr"
+  WEIGHTS: "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
+  PIXEL_MEAN: [123.675, 116.280, 103.530]
+  PIXEL_STD: [58.395, 57.120, 57.375]
+  MASK_ON: False
+
+  BACKBONE:
+    NAME: "build_regnet_backbone"
+  REGNETS:
+    TYPE: "RegNetX_400MF"
+    OUT_FEATURES: ["s2", "s3", "s4"] # fpn produce 4 levels, only using 3 for now
+  # RESNETS:
+  #   DEPTH: 50
+  #   STRIDE_IN_1X1: False
+  #   OUT_FEATURES: ["res2", "res3", "res4", "res5"]
+  DETR:
+    GIOU_WEIGHT: 2.0
+    L1_WEIGHT: 5.0
+    NUM_OBJECT_QUERIES: 100
+    ENC_LAYERS: 6
+    DEC_LAYERS: 6
+    HIDDEN_DIM: 256
+
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+
+SOLVER:
+  IMS_PER_BATCH: 28
+  BASE_LR: 0.00005
+  STEPS: (369600,)
+  MAX_ITER: 554400
+  WARMUP_FACTOR: 1.0
+  WARMUP_ITERS: 10
+  WEIGHT_DECAY: 0.0001
+  OPTIMIZER: "ADAMW"
+  BACKBONE_MULTIPLIER: 0.1
+  CLIP_GRADIENTS:
+    ENABLED: True
+    CLIP_TYPE: "full_model"
+    CLIP_VALUE: 0.01
+    NORM_TYPE: 2.0
+INPUT:
+  MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
+  CROP:
+    ENABLED: True
+    TYPE: "absolute_range"
+    SIZE: (384, 600)
+  FORMAT: "RGB"
+TEST:
+  EVAL_PERIOD: 4000
+DATALOADER:
+  FILTER_EMPTY_ANNOTATIONS: False
+  NUM_WORKERS: 2
+VERSION: 2
+OUTPUT_DIR: "output/coco_detr_regx"

configs/coco-keypoints/yolox_kpts.yaml

@@ -0,0 +1,85 @@
+_BASE_: "../Base-YOLOv7.yaml"
+MODEL:
+  PIXEL_MEAN: [0.485, 0.456, 0.406] # same value as PP-YOLOv2, RGB order
+  PIXEL_STD: [0.229, 0.224, 0.225]
+
+  WEIGHTS: ""
+  KEYPOINT_ON: True
+  META_ARCHITECTURE: "YOLOX"
+  BACKBONE:
+    NAME: "build_cspdarknetx_backbone"
+
+  DARKNET:
+    WEIGHTS: ""
+    DEPTH_WISE: False
+    OUT_FEATURES: ["dark3", "dark4", "dark5"]
+
+  YOLO:
+    CLASSES: 80
+    IN_FEATURES: ["dark3", "dark4", "dark5"]
+    CONF_THRESHOLD: 0.001
+    NMS_THRESHOLD: 0.65
+    IGNORE_THRESHOLD: 0.7
+    WIDTH_MUL: 0.50
+    DEPTH_MUL: 0.33
+    LOSS_TYPE: "v7"
+    LOSS:
+      LAMBDA_IOU: 1.5
+
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  # TEST: ("coco_2014_val_mini",)
+  TEST: ("coco_2017_val",)
+
+INPUT:
+  # FORMAT: "RGB" # using BGR default
+  MIN_SIZE_TRAIN: (416, 512, 608, 768)
+  MAX_SIZE_TRAIN: 800 # force max size train to 800?
+  MIN_SIZE_TEST: 640
+  MAX_SIZE_TEST: 800
+  # open all augmentations
+  JITTER_CROP:
+    ENABLED: False
+  RESIZE:
+    ENABLED: False
+    # SHAPE: (540, 960)
+  DISTORTION:
+    ENABLED: True
+  COLOR_JITTER:
+    BRIGHTNESS: True
+    SATURATION: True
+  # MOSAIC:
+  #   ENABLED: True
+  #   NUM_IMAGES: 4
+  #   DEBUG_VIS: True
+  #   # MOSAIC_WIDTH: 960
+  #   # MOSAIC_HEIGHT: 540
+  MOSAIC_AND_MIXUP:
+    ENABLED: True
+    # ENABLED: False
+    DEBUG_VIS: False
+    ENABLE_MIXUP: False
+    DISABLE_AT_ITER: 120000
+
+
+SOLVER:
+  # enable fp16 training
+  AMP:
+    ENABLED: true
+  IMS_PER_BATCH: 112
+  BASE_LR: 0.027
+  STEPS: (60000, 80000)
+  WARMUP_FACTOR: 0.00033333
+  WARMUP_ITERS: 1200
+  MAX_ITER: 230000
+  LR_SCHEDULER_NAME: "WarmupCosineLR"
+
+TEST:
+  EVAL_PERIOD: 10000
+  # EVAL_PERIOD: 0
+OUTPUT_DIR: "output/coco_yolox_s_kpts"
+VIS_PERIOD: 5000
+
+DATALOADER:
+  # proposals are part of the dataset_dicts, and take a lot of RAM
+  NUM_WORKERS: 3

configs/coco/darknet53.yaml

@@ -1,4 +1,4 @@
-_BASE_: "../Base-YoloV7.yaml"
+_BASE_: "../Base-YOLOv7.yaml"
 MODEL:
   WEIGHTS: ""
   MASK_ON: False
@@ -23,6 +23,9 @@ MODEL:
     CONF_THRESHOLD: 0.2
     NMS_THRESHOLD: 0.1
     IGNORE_THRESHOLD: 0.6
+    NECK:
+      TYPE: "fpn"
+      WITH_SPP: True
 DATASETS:
   TRAIN: ("coco_2017_train",)
   TEST: ("coco_2017_val",)

configs/coco/detr/anchordetr_origin.yaml

@@ -0,0 +1,67 @@
+MODEL:
+  META_ARCHITECTURE: "AnchorDetr"
+  WEIGHTS: "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
+  # WEIGHTS: "weights/R-50.pkl"
+  PIXEL_MEAN: [123.675, 116.280, 103.530]
+  PIXEL_STD: [58.395, 57.120, 57.375]
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 50
+    STRIDE_IN_1X1: False
+    OUT_FEATURES: ["res2", "res3", "res4", "res5"]
+  DETR:
+    GIOU_WEIGHT: 2.0
+    L1_WEIGHT: 5.0
+    NUM_QUERY_POSITION: 300
+    DIM_FEEDFORWARD: 1024
+    DROPOUT: 0.0
+    ENC_LAYERS: 6
+    DEC_LAYERS: 6
+    HIDDEN_DIM: 256
+    NUM_CLASSES: 80
+  YOLO:
+    CONF_THRESHOLD: 0.001
+    IGNORE_THRESHOLD: 0.07
+
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+
+SOLVER:
+  AMP:
+    ENABLED: true
+  IMS_PER_BATCH: 16
+  # BASE_LR: 0.0002 # 0.00025 is better
+  BASE_LR: 0.0001 # 0.00025 is better
+  STEPS: (295720, )
+  # MAX_ITER: 369650
+  MAX_ITER: 409650
+  WARMUP_FACTOR: 1.0
+  WARMUP_ITERS: 10
+  WEIGHT_DECAY: 0.0001
+  OPTIMIZER: "ADAMW"
+  LR_MULTIPLIER_OVERWRITE:
+    [{ "backbone": 0.1 }, { "reference_points": 0.1, "sampling_offsets": 0.1 }]
+  CLIP_GRADIENTS:
+    ENABLED: True
+    CLIP_TYPE: "full_model"
+    # CLIP_TYPE: "norm"
+    CLIP_VALUE: 0.1
+    NORM_TYPE: 2.0
+INPUT:
+  MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
+  CROP:
+    ENABLED: True
+    TYPE: "absolute_range"
+    SIZE: (384, 600)
+    # SIZE: (384, 632)
+  FORMAT: "RGB"
+TEST:
+  EVAL_PERIOD: 7393 # 1 epoch same as bs=2
+DATALOADER:
+  FILTER_EMPTY_ANNOTATIONS: False
+  NUM_WORKERS: 2
+VERSION: 2
+
+VIS_PERIOD: 100
+OUTPUT_DIR: "output/coco_anchordetr"

configs/coco/detr/anchordetr_origin_bs64.yaml

@@ -0,0 +1,68 @@
+MODEL:
+  META_ARCHITECTURE: "AnchorDetr"
+  WEIGHTS: "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
+  # WEIGHTS: "weights/R-50.pkl"
+  PIXEL_MEAN: [123.675, 116.280, 103.530]
+  PIXEL_STD: [58.395, 57.120, 57.375]
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 50
+    STRIDE_IN_1X1: False
+    OUT_FEATURES: ["res2", "res3", "res4", "res5"]
+  DETR:
+    GIOU_WEIGHT: 2.0
+    L1_WEIGHT: 5.0
+    NUM_QUERY_POSITION: 300
+    DIM_FEEDFORWARD: 1024
+    DROPOUT: 0.0
+    ENC_LAYERS: 6
+    DEC_LAYERS: 6
+    HIDDEN_DIM: 256
+    NUM_CLASSES: 80
+  YOLO:
+    CONF_THRESHOLD: 0.001
+    IGNORE_THRESHOLD: 0.07
+
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+
+SOLVER:
+  AMP:
+    ENABLED: true
+  IMS_PER_BATCH: 64
+  # BASE_LR: 0.0002 # 0.00025 is better
+  BASE_LR: 0.00025 # lr should be a little bit larger
+  # STEPS: (295720, )
+  STEPS: (73930, )
+  # MAX_ITER: 369650
+  MAX_ITER: 409650
+  WARMUP_FACTOR: 1.0
+  WARMUP_ITERS: 10
+  WEIGHT_DECAY: 0.0001
+  OPTIMIZER: "ADAMW"
+  LR_MULTIPLIER_OVERWRITE:
+    [{ "backbone": 0.1 }, { "reference_points": 0.1, "sampling_offsets": 0.1 }]
+  CLIP_GRADIENTS:
+    ENABLED: True
+    CLIP_TYPE: "full_model"
+    # CLIP_TYPE: "norm"
+    CLIP_VALUE: 0.1
+    NORM_TYPE: 2.0
+INPUT:
+  MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
+  CROP:
+    ENABLED: True
+    TYPE: "absolute_range"
+    SIZE: (384, 600)
+    # SIZE: (384, 632)
+  FORMAT: "RGB"
+TEST:
+  EVAL_PERIOD: 7393 # 1 epoch same as bs=2
+DATALOADER:
+  FILTER_EMPTY_ANNOTATIONS: False
+  NUM_WORKERS: 2
+VERSION: 2
+
+VIS_PERIOD: 100
+OUTPUT_DIR: "output/coco_anchordetr"

configs/coco/detr/d2go/detr_bs16.yaml

@@ -0,0 +1,62 @@
+MODEL:
+  META_ARCHITECTURE: "DetrD2go"
+  PIXEL_MEAN: [123.675, 116.280, 103.530]
+  PIXEL_STD: [58.395, 57.120, 57.375]
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 50
+    STRIDE_IN_1X1: False
+    OUT_FEATURES: ["res2", "res3", "res4", "res5"]
+  DETR:
+    NUM_CLASSES: 80
+    CLS_WEIGHT: 2.0
+    DIM_FEEDFORWARD: 1024
+    GIOU_WEIGHT: 2.0
+    L1_WEIGHT: 5.0
+    NUM_OBJECT_QUERIES: 300
+    CENTERED_POSITION_ENCODIND: True
+    USE_FOCAL_LOSS: True
+    NUM_FEATURE_LEVELS: 1
+    ATTENTION_TYPE: 'DETR'
+
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+
+SOLVER:
+  AMP:
+    ENABLED: true
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.0001
+  STEPS: (887040,)
+  MAX_ITER: 1108800
+  WARMUP_FACTOR: 1.0
+  WARMUP_ITERS: 10
+  WEIGHT_DECAY: 0.0001
+  OPTIMIZER: "ADAMW"
+  CLIP_GRADIENTS:
+    ENABLED: True
+    CLIP_TYPE: "full_model"
+    CLIP_VALUE: 0.1
+    NORM_TYPE: 2.0
+  LR_MULTIPLIER_OVERWRITE: [{'backbone': 0.1}, {'reference_points': 0.1, 'sampling_offsets': 0.1}]
+  # BACKBONE_MULTIPLIER: 0.1
+
+INPUT:
+  MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
+  CROP:
+    ENABLED: True
+    TYPE: "absolute_range"
+    SIZE: (384, 600)
+  FORMAT: "RGB"
+# D2GO_DATA:
+#   MAPPER:
+#     NAME: "DETRDatasetMapper"
+
+TEST:
+  EVAL_PERIOD: 4000
+DATALOADER:
+  FILTER_EMPTY_ANNOTATIONS: False
+  NUM_WORKERS: 4
+VERSION: 2
+