DETR: CNN-transformer pipeline resolved concretely; matching via Hungarian (hmdetr)

Block costs delegated to morie_geron_block_params; bipartite matching and set-prediction loss to morie_geron_detr_hungarian_matching.

Usage

morie_geron_detr(
  image,
  n_queries = 100,
  n_layers = 6,
  d_model = 256,
  n_heads = 8,
  n_classes = 91,
  backbone_stride = 32,
  pred_boxes = NULL,
  pred_classes = NULL,
  gt_boxes = NULL,
  gt_classes = NULL
)

Arguments

image

(H,W) or (C,H,W).

n_queries, n_layers, d_model, n_heads, n_classes, backbone_stride

int.

pred_boxes, pred_classes, gt_boxes, gt_classes

Optional matching inputs.

Value

List with feature_shape, n_tokens, total_params, matching, loss, max_detections.

Examples

morie_geron_detr(array(0, dim = c(3, 224, 224)), n_queries = 10,
    n_layers = 1, d_model = 8, n_heads = 2, n_classes = 3)
#> $feature_shape
#> [1] 7 7
#> 
#> $n_tokens
#> [1] 49
#> 
#> $n_queries
#> [1] 10
#> 
#> $total_params
#> [1] 18808
#> 
#> $encoder_params
#> [1] 872
#> 
#> $decoder_params
#> [1] 1176
#> 
#> $projection_params
#> [1] 16392
#> 
#> $query_params
#> [1] 80
#> 
#> $head_params
#> [1] 288
#> 
#> $encoder_attention_cost
#> [1] 2401
#> 
#> $max_detections
#> [1] 10
#> 
#> $matching
#> NULL
#> 
#> $loss
#> NULL
#> 
#> $loss_bbox
#> NULL
#> 
#> $loss_class
#> NULL
#> 
#> $estimate
#> [1] 18808
#> 
#> $n
#> [1] 49
#> 
#> $method
#> [1] "DETR pipeline resolved concretely; blocks via hmdctr, set matching via grdetr"
#>