def symbolic(g, single_level_grid_anchors, base_anchors, base_size, scales,
ratios, anchor_stride, feat, img_tensor, target_stds):
min_size = base_size
max_sizes = []
ars = []
for scale in scales[1:]:
max_sizes.append(scale * scale * min_size)
for ar in ratios:
if ar > 1:
ars.append(ar)
return g.op(add_domain("PriorBox"),
feat,
img_tensor,
min_size_f=[min_size],
max_size_f=max_sizes,
aspect_ratio_f=ars,
flip_i=1,
clip_i=0,
variance_f=list(target_stds),
step_f=anchor_stride[0],
offset_f=0.5,
step_h_f=0,
step_w_f=0,
img_size_i=0,
img_h_i=0,
img_w_i=0)
def symbolic(g, cls_scores, bbox_preds, img_metas, cfg, rescale, priors,
cls_out_channels, use_sigmoid_cls, target_means, target_stds):
return g.op(add_domain("DetectionOutput"),
bbox_preds,
cls_scores,
priors,
num_classes_i=cls_out_channels,
background_label_id_i=cls_out_channels - 1,
top_k_i=cfg['max_per_img'],
keep_top_k_i=cfg['max_per_img'],
confidence_threshold_f=cfg['score_thr'],
nms_threshold_f=cfg['nms']['iou_thr'],
eta_f=1,
share_location_i=1,
code_type_s="CENTER_SIZE",
variance_encoded_in_target_i=0)
def symbolic(g, single_level_grid_anchors, base_anchors, anchors_heights,
anchors_widths, anchor_stride, feat, img_tensor, target_stds):
return g.op(add_domain("PriorBoxClustered"),
feat,
img_tensor,
height_f=anchors_heights,
width_f=anchors_widths,
flip_i=0,
clip_i=0,
variance_f=list(target_stds),
step_f=anchor_stride[0],
offset_f=0.5,
step_h_f=0,
step_w_f=0,
img_size_i=0,
img_h_i=0,
img_w_i=0)
def symbolic(g,
input,
offset,
weight,
stride=1,
padding=0,
dilation=1,
groups=1,
deformable_groups=1,
im2col_step=64):
assert groups == 1
kh, kw = weight.type().sizes()[2:]
return g.op(add_domain('DeformableConv2D'),
input,
offset,
weight,
strides_i=stride,
pads_i=[p for pair in zip(padding, padding) for p in pair],
dilations_i=dilation,
groups_i=groups,
deformable_groups_i=deformable_groups,
kernel_shape_i=[kh, kw])
def roi_feature_extractor_symbolics(g,
rois,
*feats,
output_size=1,
featmap_strides=1,
sample_num=1):
from torch.onnx.symbolic_helper import _slice_helper
rois = _slice_helper(g, rois, axes=[1], starts=[1], ends=[5])
roi_feats = g.op(
add_domain('ExperimentalDetectronROIFeatureExtractor'),
rois,
*feats,
output_size_i=output_size,
pyramid_scales_i=featmap_strides,
sampling_ratio_i=sample_num,
image_id_i=0,
distribute_rois_between_levels_i=1,
preserve_rois_order_i=0,
aligned_i=1,
outputs=1,
)
return roi_feats