|
| 1 | +# Copyright (c) OpenMMLab. All rights reserved. |
| 2 | +import torch |
| 3 | +import torch.nn.functional as F |
| 4 | + |
| 5 | +from mmdeploy.codebase.mmdet import (get_post_processing_params, |
| 6 | + multiclass_nms, pad_with_value) |
| 7 | +from mmdeploy.core import FUNCTION_REWRITER |
| 8 | +from mmdeploy.utils import Backend, get_backend, is_dynamic_shape |
| 9 | + |
| 10 | + |
| 11 | +@FUNCTION_REWRITER.register_rewriter( |
| 12 | + func_name='mmdet.models.dense_heads.gfl_head.' |
| 13 | + 'GFLHead.get_bboxes') |
| 14 | +def gfl_head__get_bbox(ctx, |
| 15 | + self, |
| 16 | + cls_scores, |
| 17 | + bbox_preds, |
| 18 | + score_factors=None, |
| 19 | + img_metas=None, |
| 20 | + cfg=None, |
| 21 | + rescale=False, |
| 22 | + with_nms=True, |
| 23 | + **kwargs): |
| 24 | + """Rewrite `get_bboxes` of `GFLHead` for default backend. |
| 25 | +
|
| 26 | + Rewrite this function to deploy model, transform network output for a |
| 27 | + batch into bbox predictions. |
| 28 | +
|
| 29 | + Args: |
| 30 | + ctx (ContextCaller): The context with additional information. |
| 31 | + self: The instance of the original class. |
| 32 | + cls_scores (list[Tensor]): Classification scores for all |
| 33 | + scale levels, each is a 4D-tensor, has shape |
| 34 | + (batch_size, num_priors * num_classes, H, W). |
| 35 | + bbox_preds (list[Tensor]): Box energies / deltas for all |
| 36 | + scale levels, each is a 4D-tensor, has shape |
| 37 | + (batch_size, num_priors * 4, H, W). |
| 38 | + score_factors (list[Tensor], Optional): Score factor for |
| 39 | + all scale level, each is a 4D-tensor, has shape |
| 40 | + (batch_size, num_priors * 1, H, W). Default None. |
| 41 | + img_metas (list[dict], Optional): Image meta info. Default None. |
| 42 | + cfg (mmcv.Config, Optional): Test / postprocessing configuration, |
| 43 | + if None, test_cfg would be used. Default None. |
| 44 | + rescale (bool): If True, return boxes in original image space. |
| 45 | + Default False. |
| 46 | + with_nms (bool): If True, do nms before return boxes. |
| 47 | + Default True. |
| 48 | +
|
| 49 | + Returns: |
| 50 | + If with_nms == True: |
| 51 | + tuple[Tensor, Tensor]: tuple[Tensor, Tensor]: (dets, labels), |
| 52 | + `dets` of shape [N, num_det, 5] and `labels` of shape |
| 53 | + [N, num_det]. |
| 54 | + Else: |
| 55 | + tuple[Tensor, Tensor, Tensor]: batch_mlvl_bboxes, |
| 56 | + batch_mlvl_scores, batch_mlvl_centerness |
| 57 | + """ |
| 58 | + deploy_cfg = ctx.cfg |
| 59 | + is_dynamic_flag = is_dynamic_shape(deploy_cfg) |
| 60 | + backend = get_backend(deploy_cfg) |
| 61 | + num_levels = len(cls_scores) |
| 62 | + |
| 63 | + featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores] |
| 64 | + mlvl_priors = self.prior_generator.grid_priors( |
| 65 | + featmap_sizes, dtype=bbox_preds[0].dtype, device=bbox_preds[0].device) |
| 66 | + |
| 67 | + mlvl_cls_scores = [cls_scores[i].detach() for i in range(num_levels)] |
| 68 | + mlvl_bbox_preds = [bbox_preds[i].detach() for i in range(num_levels)] |
| 69 | + if score_factors is None: |
| 70 | + with_score_factors = False |
| 71 | + mlvl_score_factor = [None for _ in range(num_levels)] |
| 72 | + else: |
| 73 | + with_score_factors = True |
| 74 | + mlvl_score_factor = [ |
| 75 | + score_factors[i].detach() for i in range(num_levels) |
| 76 | + ] |
| 77 | + mlvl_score_factors = [] |
| 78 | + assert img_metas is not None |
| 79 | + img_shape = img_metas[0]['img_shape'] |
| 80 | + |
| 81 | + assert len(cls_scores) == len(bbox_preds) == len(mlvl_priors) |
| 82 | + batch_size = cls_scores[0].shape[0] |
| 83 | + cfg = self.test_cfg |
| 84 | + pre_topk = cfg.get('nms_pre', -1) |
| 85 | + |
| 86 | + mlvl_valid_bboxes = [] |
| 87 | + mlvl_valid_scores = [] |
| 88 | + mlvl_valid_priors = [] |
| 89 | + |
| 90 | + for cls_score, bbox_pred, score_factors, priors, stride in zip( |
| 91 | + mlvl_cls_scores, mlvl_bbox_preds, mlvl_score_factor, mlvl_priors, |
| 92 | + self.prior_generator.strides): |
| 93 | + assert cls_score.size()[-2:] == bbox_pred.size()[-2:] |
| 94 | + assert stride[0] == stride[1] |
| 95 | + |
| 96 | + scores = cls_score.permute(0, 2, 3, 1).reshape(batch_size, -1, |
| 97 | + self.cls_out_channels) |
| 98 | + if self.use_sigmoid_cls: |
| 99 | + scores = scores.sigmoid() |
| 100 | + nms_pre_score = scores |
| 101 | + else: |
| 102 | + scores = scores.softmax(-1) |
| 103 | + nms_pre_score = scores |
| 104 | + if with_score_factors: |
| 105 | + score_factors = score_factors.permute(0, 2, 3, |
| 106 | + 1).reshape(batch_size, |
| 107 | + -1).sigmoid() |
| 108 | + score_factors = score_factors.unsqueeze(2) |
| 109 | + bbox_pred = batched_integral(self.integral, |
| 110 | + bbox_pred.permute(0, 2, 3, 1)) * stride[0] |
| 111 | + if not is_dynamic_flag: |
| 112 | + priors = priors.data |
| 113 | + priors = priors.expand(batch_size, -1, priors.size(-1)) |
| 114 | + if pre_topk > 0: |
| 115 | + if with_score_factors: |
| 116 | + nms_pre_score = nms_pre_score * score_factors |
| 117 | + if backend == Backend.TENSORRT: |
| 118 | + priors = pad_with_value(priors, 1, pre_topk) |
| 119 | + bbox_pred = pad_with_value(bbox_pred, 1, pre_topk) |
| 120 | + scores = pad_with_value(scores, 1, pre_topk, 0.) |
| 121 | + nms_pre_score = pad_with_value(nms_pre_score, 1, pre_topk, 0.) |
| 122 | + if with_score_factors: |
| 123 | + score_factors = pad_with_value(score_factors, 1, pre_topk, |
| 124 | + 0.) |
| 125 | + |
| 126 | + # Get maximum scores for foreground classes. |
| 127 | + if self.use_sigmoid_cls: |
| 128 | + max_scores, _ = nms_pre_score.max(-1) |
| 129 | + else: |
| 130 | + max_scores, _ = nms_pre_score[..., :-1].max(-1) |
| 131 | + _, topk_inds = max_scores.topk(pre_topk) |
| 132 | + batch_inds = torch.arange( |
| 133 | + batch_size, |
| 134 | + device=bbox_pred.device).view(-1, 1).expand_as(topk_inds) |
| 135 | + priors = priors[batch_inds, topk_inds, :] |
| 136 | + bbox_pred = bbox_pred[batch_inds, topk_inds, :] |
| 137 | + scores = scores[batch_inds, topk_inds, :] |
| 138 | + if with_score_factors: |
| 139 | + score_factors = score_factors[batch_inds, topk_inds, :] |
| 140 | + |
| 141 | + mlvl_valid_bboxes.append(bbox_pred) |
| 142 | + mlvl_valid_scores.append(scores) |
| 143 | + priors = self.anchor_center(priors) |
| 144 | + mlvl_valid_priors.append(priors) |
| 145 | + if with_score_factors: |
| 146 | + mlvl_score_factors.append(score_factors) |
| 147 | + |
| 148 | + batch_mlvl_bboxes_pred = torch.cat(mlvl_valid_bboxes, dim=1) |
| 149 | + batch_scores = torch.cat(mlvl_valid_scores, dim=1) |
| 150 | + batch_priors = torch.cat(mlvl_valid_priors, dim=1) |
| 151 | + batch_bboxes = self.bbox_coder.decode( |
| 152 | + batch_priors, batch_mlvl_bboxes_pred, max_shape=img_shape) |
| 153 | + if with_score_factors: |
| 154 | + batch_score_factors = torch.cat(mlvl_score_factors, dim=1) |
| 155 | + |
| 156 | + if not self.use_sigmoid_cls: |
| 157 | + batch_scores = batch_scores[..., :self.num_classes] |
| 158 | + |
| 159 | + if with_score_factors: |
| 160 | + batch_scores = batch_scores * batch_score_factors |
| 161 | + if not with_nms: |
| 162 | + return batch_bboxes, batch_scores |
| 163 | + post_params = get_post_processing_params(deploy_cfg) |
| 164 | + max_output_boxes_per_class = post_params.max_output_boxes_per_class |
| 165 | + iou_threshold = cfg.nms.get('iou_threshold', post_params.iou_threshold) |
| 166 | + score_threshold = cfg.get('score_thr', post_params.score_threshold) |
| 167 | + pre_top_k = post_params.pre_top_k |
| 168 | + keep_top_k = cfg.get('max_per_img', post_params.keep_top_k) |
| 169 | + return multiclass_nms( |
| 170 | + batch_bboxes, |
| 171 | + batch_scores, |
| 172 | + max_output_boxes_per_class, |
| 173 | + iou_threshold=iou_threshold, |
| 174 | + score_threshold=score_threshold, |
| 175 | + pre_top_k=pre_top_k, |
| 176 | + keep_top_k=keep_top_k) |
| 177 | + |
| 178 | + |
| 179 | +def batched_integral(intergral, x): |
| 180 | + batch_size = x.size(0) |
| 181 | + x = F.softmax(x.reshape(batch_size, -1, intergral.reg_max + 1), dim=2) |
| 182 | + x = F.linear(x, |
| 183 | + intergral.project.type_as(x).unsqueeze(0)).reshape( |
| 184 | + batch_size, -1, 4) |
| 185 | + return x |
0 commit comments