All configurations available in Detectron2 are not supported by D2go

[shivadvg19]

• I am able to apply Object detection with Mask-RCNN using detectron2 and getting good results for my custom dataset. but when i try to use the same architecture with d2go to create a mobile deployable model am not able to use the same config file.

• can we have a list of configurations/architectures supported in d2go with quantization aware training and deployable in android mobiles.

[zhanghang1989]

Standard detectron models are not deployable. The supported backbones are available at

d2go/d2go/modeling/modeldef/modeldef.py

Lines 414 to 522 in 82295db

	MODEL_ARCH_BUILTIN = {
	"default": {
	"trunk": DEFAULT_STAGES[0:4],
	"rpn": [[_repeat_last(DEFAULT_STAGES[3])]],
	"bbox": LARGE_BOX_HEAD_STAGES,
	"mask": LARGE_UPSAMPLE_HEAD_STAGES,
	"kpts": LARGE_UPSAMPLE_HEAD_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	"default_dsmask": {
	"trunk": DEFAULT_STAGES[0:4],
	"rpn": [[_repeat_last(DEFAULT_STAGES[3])]],
	"bbox": SMALL_BOX_HEAD_STAGES,
	"mask": SMALL_DS_UPSAMPLE_HEAD_STAGES,
	"kpts": SMALL_DS_UPSAMPLE_HEAD_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_A": {
	"trunk": FBNetV3_A[0:4],
	"rpn": [[_repeat_last(FBNetV3_A[3])]],
	"bbox": [FBNetV3_A[4]],
	"mask": SMALL_DS_UPSAMPLE_HEAD_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_B": {
	"trunk": FBNetV3_B[0:4],
	"rpn": [[_repeat_last(FBNetV3_B[3])]],
	"bbox": [FBNetV3_B[4]],
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_C": {
	"trunk": FBNetV3_C[0:4],
	"rpn": [[_repeat_last(FBNetV3_C[3])]],
	"bbox": [FBNetV3_C[4]],
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_D": {
	"trunk": FBNetV3_D[0:4],
	"rpn": [[_repeat_last(FBNetV3_D[3])]],
	"bbox": [FBNetV3_D[4]],
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_E": {
	"trunk": FBNetV3_E[0:4],
	"rpn": [[_repeat_last(FBNetV3_E[3])]],
	"bbox": [FBNetV3_E[4]],
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_F": {
	"trunk": FBNetV3_F[0:4],
	"rpn": [[_repeat_last(FBNetV3_F[3])]],
	"bbox": [FBNetV3_F[4]],
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_G": {
	"trunk": FBNetV3_G[0:4],
	"rpn": [[_repeat_last(FBNetV3_G[3])]],
	"bbox": [FBNetV3_G[4]],
	"mask": LARGE_UPSAMPLE_HEAD_STAGES,
	"kpts": LARGE_UPSAMPLE_HEAD_D21_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_H": {
	"trunk": FBNetV3_H[0:4],
	"rpn": [[_repeat_last(FBNetV3_H[3])]],
	"bbox": [FBNetV3_H[4]],
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_A_dsmask_C5": {
	"trunk": FBNetV3_A_dsmask,
	"rpn": [[_repeat_last(FBNetV3_A_dsmask[3])]],
	"bbox": SMALL_BOX_HEAD_STAGES,
	"mask": SMALL_DS_UPSAMPLE_HEAD_STAGES,
	"kpts": SMALL_DS_UPSAMPLE_HEAD_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_A_dsmask": {
	"trunk": FBNetV3_A_dsmask[0:4],
	"rpn": [[_repeat_last(FBNetV3_A_dsmask[3])]],
	"bbox": SMALL_BOX_HEAD_STAGES,
	"mask": SMALL_DS_UPSAMPLE_HEAD_STAGES,
	"kpts": SMALL_DS_UPSAMPLE_HEAD_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_A_dsmask_tiny": {
	"trunk": FBNetV3_A_dsmask_tiny[0:4],
	"rpn": [[_repeat_last(FBNetV3_A_dsmask_tiny[3])]],
	"bbox": TINY_BOX_HEAD_STAGES,
	"mask": TINY_DS_UPSAMPLE_HEAD_STAGES,
	"kpts": TINY_DS_UPSAMPLE_HEAD_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_B_light_large": {
	"trunk": FBNetV3_B_light_no_se[0:4],
	"rpn": [[_repeat_last(FBNetV3_B_light_no_se[3])]],
	"bbox": SMALL_BOX_HEAD_STAGES,
	"mask": SMALL_DS_UPSAMPLE_HEAD_STAGES,
	"kpts": LARGE_UPSAMPLE_HEAD_D21_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	"FBNetV3_G_fpn": {
	"trunk": FBNetV3_G[0:5], # FPN uses all 5 stages
	"rpn": [[_repeat_last(FBNetV3_G[3], n=1)]],
	"bbox": [FBNetV3_G[4]],
	"mask": FPN_UPSAMPLE_HEAD_STAGES,
	"kpts": LARGE_UPSAMPLE_HEAD_D21_STAGES,
	"basic_args": _BASIC_ARGS,
	},
	}

Please also used the built-in faster rcnn / mask rcnn model configs https://github.com/facebookresearch/d2go/tree/master/configs