add: mobilenetv4 backbone (#217)

* Fix: data type bug * Optim: model inference display * Fix: cls loss weight too high * add: mobilenetv4 backbone
Seeed-Studio · May 22, 2024 · 343a2cb · 343a2cb
1 parent 40fd17a
commit 343a2cb
Show file tree

Hide file tree

Showing 5 changed files with 1,293 additions and 10 deletions.
diff --git a/configs/swift_yolo/base_arch.py b/configs/swift_yolo/base_arch.py
@@ -147,17 +147,27 @@
  ),
  prior_generator=dict(type='mmdet.YOLOAnchorGenerator', base_sizes=anchors, strides=strides),
  # scaled based on number of detection layers
- loss_cls=dict(type='mmdet.CrossEntropyLoss', use_sigmoid=True, reduction='mean', loss_weight=loss_cls_weight),
+ loss_cls=dict(
+ type='mmdet.CrossEntropyLoss',
+ use_sigmoid=True,
+ reduction='mean',
+ loss_weight=loss_cls_weight * (num_classes / 80 * 3 / num_det_layers),
+ ),
  loss_bbox=dict(
  type='IoULoss',
  iou_mode='ciou',
  bbox_format='xywh',
  eps=1e-7,
  reduction='mean',
- loss_weight=loss_bbox_weight,
+ loss_weight=loss_bbox_weight * (3 / num_det_layers),
  return_iou=True,
  ),
- loss_obj=dict(type='mmdet.CrossEntropyLoss', use_sigmoid=True, reduction='mean', loss_weight=loss_obj_weight),
+ loss_obj=dict(
+ type='mmdet.CrossEntropyLoss',
+ use_sigmoid=True,
+ reduction='mean',
+ loss_weight=loss_obj_weight * ((imgsz[0] / 640) ** 2 * 3 / num_det_layers),
+ ),
  prior_match_thr=prior_match_thr,
  obj_level_weights=obj_level_weights,
  ),

diff --git a/configs/swift_yolo/swift_yolo_labelmatch_coco.py b/configs/swift_yolo/swift_yolo_labelmatch_coco.py
@@ -139,17 +139,27 @@
  ),
  prior_generator=dict(type='mmdet.YOLOAnchorGenerator', base_sizes=anchors, strides=strides),
  # scaled based on number of detection layers
- loss_cls=dict(type='mmdet.CrossEntropyLoss', use_sigmoid=True, reduction='mean', loss_weight=loss_cls_weight),
+ loss_cls=dict(
+ type='mmdet.CrossEntropyLoss',
+ use_sigmoid=True,
+ reduction='mean',
+ loss_weight=loss_cls_weight * (num_classes / 80 * 3 / num_det_layers),
+ ),
  loss_bbox=dict(
  type='IoULoss',
  iou_mode='ciou',
  bbox_format='xywh',
  eps=1e-7,
  reduction='mean',
- loss_weight=loss_bbox_weight,
+ loss_weight=loss_bbox_weight * (3 / num_det_layers),
  return_iou=True,
  ),
- loss_obj=dict(type='mmdet.CrossEntropyLoss', use_sigmoid=True, reduction='mean', loss_weight=loss_obj_weight),
+ loss_obj=dict(
+ type='mmdet.CrossEntropyLoss',
+ use_sigmoid=True,
+ reduction='mean',
+ loss_weight=loss_obj_weight * ((imgsz[0] / 640) ** 2 * 3 / num_det_layers),
+ ),
  prior_match_thr=prior_match_thr,
  obj_level_weights=obj_level_weights,
  ),
@@ -418,7 +428,7 @@
  ),
  dict(type='mmdet.MeanTeacherHook'),
  dict(type="sscma.SemiHook", bure_epoch=200),
- dict(type="sscma.LabelMatchHook",priority=100),
+ dict(type="sscma.LabelMatchHook", priority=100),
 ]
 
 

diff --git a/configs/swift_yolo/swift_yolo_mb2_1xb16_300e_coco.py b/configs/swift_yolo/swift_yolo_mb2_1xb16_300e_coco.py
@@ -145,17 +145,27 @@
  ),
  prior_generator=dict(type='mmdet.YOLOAnchorGenerator', base_sizes=anchors, strides=strides),
  # scaled based on number of detection layers
- loss_cls=dict(type='mmdet.CrossEntropyLoss', use_sigmoid=True, reduction='mean', loss_weight=loss_cls_weight),
+ loss_cls=dict(
+ type='mmdet.CrossEntropyLoss',
+ use_sigmoid=True,
+ reduction='mean',
+ loss_weight=loss_cls_weight * (num_classes / 80 * 3 / num_det_layers),
+ ),
  loss_bbox=dict(
  type='IoULoss',
  iou_mode='ciou',
  bbox_format='xywh',
  eps=1e-7,
  reduction='mean',
- loss_weight=loss_bbox_weight,
+ loss_weight=loss_bbox_weight * (3 / num_det_layers),
  return_iou=True,
  ),
- loss_obj=dict(type='mmdet.CrossEntropyLoss', use_sigmoid=True, reduction='mean', loss_weight=loss_obj_weight),
+ loss_obj=dict(
+ type='mmdet.CrossEntropyLoss',
+ use_sigmoid=True,
+ reduction='mean',
+ loss_weight=loss_obj_weight * ((imgsz[0] / 640) ** 2 * 3 / num_det_layers),
+ ),
  prior_match_thr=prior_match_thr,
  obj_level_weights=obj_level_weights,
  ),