# Copyright (c) OpenMMLab. All rights reserved. # Copyright (c) 2019 Western Digital Corporation or its affiliates. import warnings import torch import torch.nn as nn import torch.nn.functional as F from mmcv.cnn import (ConvModule, bias_init_with_prob, constant_init, is_norm, normal_init) from mmcv.runner import force_fp32 from mmdet.core import (build_assigner, build_bbox_coder, build_prior_generator, build_sampler, images_to_levels, multi_apply, multiclass_nms) from ..builder import HEADS, build_loss from .base_dense_head import BaseDenseHead from .dense_test_mixins import BBoxTestMixin @HEADS.register_module() class YOLOV3Head(BaseDenseHead, BBoxTestMixin): """YOLOV3Head Paper link: https://arxiv.org/abs/1804.02767. Args: num_classes (int): The number of object classes (w/o background) in_channels (List[int]): Number of input channels per scale. out_channels (List[int]): The number of output channels per scale before the final 1x1 layer. Default: (1024, 512, 256). anchor_generator (dict): Config dict for anchor generator bbox_coder (dict): Config of bounding box coder. featmap_strides (List[int]): The stride of each scale. Should be in descending order. Default: (32, 16, 8). one_hot_smoother (float): Set a non-zero value to enable label-smooth Default: 0. conv_cfg (dict): Config dict for convolution layer. Default: None. norm_cfg (dict): Dictionary to construct and config norm layer. Default: dict(type='BN', requires_grad=True) act_cfg (dict): Config dict for activation layer. Default: dict(type='LeakyReLU', negative_slope=0.1). loss_cls (dict): Config of classification loss. loss_conf (dict): Config of confidence loss. loss_xy (dict): Config of xy coordinate loss. loss_wh (dict): Config of wh coordinate loss. train_cfg (dict): Training config of YOLOV3 head. Default: None. test_cfg (dict): Testing config of YOLOV3 head. Default: None. init_cfg (dict or list[dict], optional): Initialization config dict. """ def __init__(self, num_classes, in_channels, out_channels=(1024, 512, 256), anchor_generator=dict( type='YOLOAnchorGenerator', base_sizes=[[(116, 90), (156, 198), (373, 326)], [(30, 61), (62, 45), (59, 119)], [(10, 13), (16, 30), (33, 23)]], strides=[32, 16, 8]), bbox_coder=dict(type='YOLOBBoxCoder'), featmap_strides=[32, 16, 8], one_hot_smoother=0., conv_cfg=None, norm_cfg=dict(type='BN', requires_grad=True), act_cfg=dict(type='LeakyReLU', negative_slope=0.1), loss_cls=dict( type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0), loss_conf=dict( type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0), loss_xy=dict( type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0), loss_wh=dict(type='MSELoss', loss_weight=1.0), train_cfg=None, test_cfg=None, init_cfg=dict( type='Normal', std=0.01, override=dict(name='convs_pred'))): super(YOLOV3Head, self).__init__(init_cfg) # Check params assert (len(in_channels) == len(out_channels) == len(featmap_strides)) self.num_classes = num_classes self.in_channels = in_channels self.out_channels = out_channels self.featmap_strides = featmap_strides self.train_cfg = train_cfg self.test_cfg = test_cfg if self.train_cfg: self.assigner = build_assigner(self.train_cfg.assigner) if hasattr(self.train_cfg, 'sampler'): sampler_cfg = self.train_cfg.sampler else: sampler_cfg = dict(type='PseudoSampler') self.sampler = build_sampler(sampler_cfg, context=self) self.fp16_enabled = False self.one_hot_smoother = one_hot_smoother self.conv_cfg = conv_cfg self.norm_cfg = norm_cfg self.act_cfg = act_cfg self.bbox_coder = build_bbox_coder(bbox_coder) self.prior_generator = build_prior_generator(anchor_generator) self.loss_cls = build_loss(loss_cls) self.loss_conf = build_loss(loss_conf) self.loss_xy = build_loss(loss_xy) self.loss_wh = build_loss(loss_wh) self.num_base_priors = self.prior_generator.num_base_priors[0] assert len( self.prior_generator.num_base_priors) == len(featmap_strides) self._init_layers() @property def anchor_generator(self): warnings.warn('DeprecationWarning: `anchor_generator` is deprecated, ' 'please use "prior_generator" instead') return self.prior_generator @property def num_anchors(self): """ Returns: int: Number of anchors on each point of feature map. """ warnings.warn('DeprecationWarning: `num_anchors` is deprecated, ' 'please use "num_base_priors" instead') return self.num_base_priors @property def num_levels(self): return len(self.featmap_strides) @property def num_attrib(self): """int: number of attributes in pred_map, bboxes (4) + objectness (1) + num_classes""" return 5 + self.num_classes def _init_layers(self): self.convs_bridge = nn.ModuleList() self.convs_pred = nn.ModuleList() for i in range(self.num_levels): conv_bridge = ConvModule( self.in_channels[i], self.out_channels[i], 3, padding=1, conv_cfg=self.conv_cfg, norm_cfg=self.norm_cfg, act_cfg=self.act_cfg) conv_pred = nn.Conv2d(self.out_channels[i], self.num_base_priors * self.num_attrib, 1) self.convs_bridge.append(conv_bridge) self.convs_pred.append(conv_pred) def init_weights(self): for m in self.modules(): if isinstance(m, nn.Conv2d): normal_init(m, mean=0, std=0.01) if is_norm(m): constant_init(m, 1) # Use prior in model initialization to improve stability for conv_pred, stride in zip(self.convs_pred, self.featmap_strides): bias = conv_pred.bias.reshape(self.num_base_priors, -1) # init objectness with prior of 8 objects per feature map # refer to https://github.com/ultralytics/yolov3 nn.init.constant_(bias.data[:, 4], bias_init_with_prob(8 / (608 / stride)**2)) nn.init.constant_(bias.data[:, 5:], bias_init_with_prob(0.01)) def forward(self, feats): """Forward features from the upstream network. Args: feats (tuple[Tensor]): Features from the upstream network, each is a 4D-tensor. Returns: tuple[Tensor]: A tuple of multi-level predication map, each is a 4D-tensor of shape (batch_size, 5+num_classes, height, width). """ assert len(feats) == self.num_levels pred_maps = [] for i in range(self.num_levels): x = feats[i] x = self.convs_bridge[i](x) pred_map = self.convs_pred[i](x) pred_maps.append(pred_map) return tuple(pred_maps), @force_fp32(apply_to=('pred_maps', )) def get_bboxes(self, pred_maps, img_metas, cfg=None, rescale=False, with_nms=True): """Transform network output for a batch into bbox predictions. It has been accelerated since PR #5991. Args: pred_maps (list[Tensor]): Raw predictions for a batch of images. img_metas (list[dict]): Meta information of each image, e.g., image size, scaling factor, etc. cfg (mmcv.Config | None): Test / postprocessing configuration, if None, test_cfg would be used. Default: None. rescale (bool): If True, return boxes in original image space. Default: False. with_nms (bool): If True, do nms before return boxes. Default: True. Returns: list[tuple[Tensor, Tensor]]: Each item in result_list is 2-tuple. The first item is an (n, 5) tensor, where 5 represent (tl_x, tl_y, br_x, br_y, score) and the score between 0 and 1. The shape of the second tensor in the tuple is (n,), and each element represents the class label of the corresponding box. """ assert len(pred_maps) == self.num_levels cfg = self.test_cfg if cfg is None else cfg scale_factors = [img_meta['scale_factor'] for img_meta in img_metas] num_imgs = len(img_metas) featmap_sizes = [pred_map.shape[-2:] for pred_map in pred_maps] mlvl_anchors = self.prior_generator.grid_priors( featmap_sizes, device=pred_maps[0].device) flatten_preds = [] flatten_strides = [] for pred, stride in zip(pred_maps, self.featmap_strides): pred = pred.permute(0, 2, 3, 1).reshape(num_imgs, -1, self.num_attrib) pred[..., :2].sigmoid_() flatten_preds.append(pred) flatten_strides.append( pred.new_tensor(stride).expand(pred.size(1))) flatten_preds = torch.cat(flatten_preds, dim=1) flatten_bbox_preds = flatten_preds[..., :4] flatten_objectness = flatten_preds[..., 4].sigmoid() flatten_cls_scores = flatten_preds[..., 5:].sigmoid() flatten_anchors = torch.cat(mlvl_anchors) flatten_strides = torch.cat(flatten_strides) flatten_bboxes = self.bbox_coder.decode(flatten_anchors, flatten_bbox_preds, flatten_strides.unsqueeze(-1)) if with_nms and (flatten_objectness.size(0) == 0): return torch.zeros((0, 5)), torch.zeros((0, )) if rescale: flatten_bboxes /= flatten_bboxes.new_tensor( scale_factors).unsqueeze(1) padding = flatten_bboxes.new_zeros(num_imgs, flatten_bboxes.shape[1], 1) flatten_cls_scores = torch.cat([flatten_cls_scores, padding], dim=-1) det_results = [] for (bboxes, scores, objectness) in zip(flatten_bboxes, flatten_cls_scores, flatten_objectness): # Filtering out all predictions with conf < conf_thr conf_thr = cfg.get('conf_thr', -1) if conf_thr > 0: conf_inds = objectness >= conf_thr bboxes = bboxes[conf_inds, :] scores = scores[conf_inds, :] objectness = objectness[conf_inds] det_bboxes, det_labels = multiclass_nms( bboxes, scores, cfg.score_thr, cfg.nms, cfg.max_per_img, score_factors=objectness) det_results.append(tuple([det_bboxes, det_labels])) return det_results @force_fp32(apply_to=('pred_maps', )) def loss(self, pred_maps, gt_bboxes, gt_labels, img_metas, gt_bboxes_ignore=None): """Compute loss of the head. Args: pred_maps (list[Tensor]): Prediction map for each scale level, shape (N, num_anchors * num_attrib, H, W) gt_bboxes (list[Tensor]): Ground truth bboxes for each image with shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format. gt_labels (list[Tensor]): class indices corresponding to each box img_metas (list[dict]): Meta information of each image, e.g., image size, scaling factor, etc. gt_bboxes_ignore (None | list[Tensor]): specify which bounding boxes can be ignored when computing the loss. Returns: dict[str, Tensor]: A dictionary of loss components. """ num_imgs = len(img_metas) device = pred_maps[0][0].device featmap_sizes = [ pred_maps[i].shape[-2:] for i in range(self.num_levels) ] mlvl_anchors = self.prior_generator.grid_priors( featmap_sizes, device=device) anchor_list = [mlvl_anchors for _ in range(num_imgs)] responsible_flag_list = [] for img_id in range(len(img_metas)): responsible_flag_list.append( self.prior_generator.responsible_flags(featmap_sizes, gt_bboxes[img_id], device)) target_maps_list, neg_maps_list = self.get_targets( anchor_list, responsible_flag_list, gt_bboxes, gt_labels) losses_cls, losses_conf, losses_xy, losses_wh = multi_apply( self.loss_single, pred_maps, target_maps_list, neg_maps_list) return dict( loss_cls=losses_cls, loss_conf=losses_conf, loss_xy=losses_xy, loss_wh=losses_wh) def loss_single(self, pred_map, target_map, neg_map): """Compute loss of a single image from a batch. Args: pred_map (Tensor): Raw predictions for a single level. target_map (Tensor): The Ground-Truth target for a single level. neg_map (Tensor): The negative masks for a single level. Returns: tuple: loss_cls (Tensor): Classification loss. loss_conf (Tensor): Confidence loss. loss_xy (Tensor): Regression loss of x, y coordinate. loss_wh (Tensor): Regression loss of w, h coordinate. """ num_imgs = len(pred_map) pred_map = pred_map.permute(0, 2, 3, 1).reshape(num_imgs, -1, self.num_attrib) neg_mask = neg_map.float() pos_mask = target_map[..., 4] pos_and_neg_mask = neg_mask + pos_mask pos_mask = pos_mask.unsqueeze(dim=-1) if torch.max(pos_and_neg_mask) > 1.: warnings.warn('There is overlap between pos and neg sample.') pos_and_neg_mask = pos_and_neg_mask.clamp(min=0., max=1.) pred_xy = pred_map[..., :2] pred_wh = pred_map[..., 2:4] pred_conf = pred_map[..., 4] pred_label = pred_map[..., 5:] target_xy = target_map[..., :2] target_wh = target_map[..., 2:4] target_conf = target_map[..., 4] target_label = target_map[..., 5:] loss_cls = self.loss_cls(pred_label, target_label, weight=pos_mask) loss_conf = self.loss_conf( pred_conf, target_conf, weight=pos_and_neg_mask) loss_xy = self.loss_xy(pred_xy, target_xy, weight=pos_mask) loss_wh = self.loss_wh(pred_wh, target_wh, weight=pos_mask) return loss_cls, loss_conf, loss_xy, loss_wh def get_targets(self, anchor_list, responsible_flag_list, gt_bboxes_list, gt_labels_list): """Compute target maps for anchors in multiple images. Args: anchor_list (list[list[Tensor]]): Multi level anchors of each image. The outer list indicates images, and the inner list corresponds to feature levels of the image. Each element of the inner list is a tensor of shape (num_total_anchors, 4). responsible_flag_list (list[list[Tensor]]): Multi level responsible flags of each image. Each element is a tensor of shape (num_total_anchors, ) gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image. gt_labels_list (list[Tensor]): Ground truth labels of each box. Returns: tuple: Usually returns a tuple containing learning targets. - target_map_list (list[Tensor]): Target map of each level. - neg_map_list (list[Tensor]): Negative map of each level. """ num_imgs = len(anchor_list) # anchor number of multi levels num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]] results = multi_apply(self._get_targets_single, anchor_list, responsible_flag_list, gt_bboxes_list, gt_labels_list) all_target_maps, all_neg_maps = results assert num_imgs == len(all_target_maps) == len(all_neg_maps) target_maps_list = images_to_levels(all_target_maps, num_level_anchors) neg_maps_list = images_to_levels(all_neg_maps, num_level_anchors) return target_maps_list, neg_maps_list def _get_targets_single(self, anchors, responsible_flags, gt_bboxes, gt_labels): """Generate matching bounding box prior and converted GT. Args: anchors (list[Tensor]): Multi-level anchors of the image. responsible_flags (list[Tensor]): Multi-level responsible flags of anchors gt_bboxes (Tensor): Ground truth bboxes of single image. gt_labels (Tensor): Ground truth labels of single image. Returns: tuple: target_map (Tensor): Predication target map of each scale level, shape (num_total_anchors, 5+num_classes) neg_map (Tensor): Negative map of each scale level, shape (num_total_anchors,) """ anchor_strides = [] for i in range(len(anchors)): anchor_strides.append( torch.tensor(self.featmap_strides[i], device=gt_bboxes.device).repeat(len(anchors[i]))) concat_anchors = torch.cat(anchors) concat_responsible_flags = torch.cat(responsible_flags) anchor_strides = torch.cat(anchor_strides) assert len(anchor_strides) == len(concat_anchors) == \ len(concat_responsible_flags) assign_result = self.assigner.assign(concat_anchors, concat_responsible_flags, gt_bboxes) sampling_result = self.sampler.sample(assign_result, concat_anchors, gt_bboxes) target_map = concat_anchors.new_zeros( concat_anchors.size(0), self.num_attrib) target_map[sampling_result.pos_inds, :4] = self.bbox_coder.encode( sampling_result.pos_bboxes, sampling_result.pos_gt_bboxes, anchor_strides[sampling_result.pos_inds]) target_map[sampling_result.pos_inds, 4] = 1 gt_labels_one_hot = F.one_hot( gt_labels, num_classes=self.num_classes).float() if self.one_hot_smoother != 0: # label smooth gt_labels_one_hot = gt_labels_one_hot * ( 1 - self.one_hot_smoother ) + self.one_hot_smoother / self.num_classes target_map[sampling_result.pos_inds, 5:] = gt_labels_one_hot[ sampling_result.pos_assigned_gt_inds] neg_map = concat_anchors.new_zeros( concat_anchors.size(0), dtype=torch.uint8) neg_map[sampling_result.neg_inds] = 1 return target_map, neg_map def aug_test(self, feats, img_metas, rescale=False): """Test function with test time augmentation. Args: feats (list[Tensor]): the outer list indicates test-time augmentations and inner Tensor should have a shape NxCxHxW, which contains features for all images in the batch. img_metas (list[list[dict]]): the outer list indicates test-time augs (multiscale, flip, etc.) and the inner list indicates images in a batch. each dict has image information. rescale (bool, optional): Whether to rescale the results. Defaults to False. Returns: list[ndarray]: bbox results of each class """ return self.aug_test_bboxes(feats, img_metas, rescale=rescale) @force_fp32(apply_to=('pred_maps')) def onnx_export(self, pred_maps, img_metas, with_nms=True): num_levels = len(pred_maps) pred_maps_list = [pred_maps[i].detach() for i in range(num_levels)] cfg = self.test_cfg assert len(pred_maps_list) == self.num_levels device = pred_maps_list[0].device batch_size = pred_maps_list[0].shape[0] featmap_sizes = [ pred_maps_list[i].shape[-2:] for i in range(self.num_levels) ] mlvl_anchors = self.prior_generator.grid_priors( featmap_sizes, device=device) # convert to tensor to keep tracing nms_pre_tensor = torch.tensor( cfg.get('nms_pre', -1), device=device, dtype=torch.long) multi_lvl_bboxes = [] multi_lvl_cls_scores = [] multi_lvl_conf_scores = [] for i in range(self.num_levels): # get some key info for current scale pred_map = pred_maps_list[i] stride = self.featmap_strides[i] # (b,h, w, num_anchors*num_attrib) -> # (b,h*w*num_anchors, num_attrib) pred_map = pred_map.permute(0, 2, 3, 1).reshape(batch_size, -1, self.num_attrib) # Inplace operation like # ```pred_map[..., :2] = \torch.sigmoid(pred_map[..., :2])``` # would create constant tensor when exporting to onnx pred_map_conf = torch.sigmoid(pred_map[..., :2]) pred_map_rest = pred_map[..., 2:] pred_map = torch.cat([pred_map_conf, pred_map_rest], dim=-1) pred_map_boxes = pred_map[..., :4] multi_lvl_anchor = mlvl_anchors[i] multi_lvl_anchor = multi_lvl_anchor.expand_as(pred_map_boxes) bbox_pred = self.bbox_coder.decode(multi_lvl_anchor, pred_map_boxes, stride) # conf and cls conf_pred = torch.sigmoid(pred_map[..., 4]) cls_pred = torch.sigmoid(pred_map[..., 5:]).view( batch_size, -1, self.num_classes) # Cls pred one-hot. # Get top-k prediction from mmdet.core.export import get_k_for_topk nms_pre = get_k_for_topk(nms_pre_tensor, bbox_pred.shape[1]) if nms_pre > 0: _, topk_inds = conf_pred.topk(nms_pre) batch_inds = torch.arange(batch_size).view( -1, 1).expand_as(topk_inds).long() # Avoid onnx2tensorrt issue in https://github.com/NVIDIA/TensorRT/issues/1134 # noqa: E501 transformed_inds = ( bbox_pred.shape[1] * batch_inds + topk_inds) bbox_pred = bbox_pred.reshape(-1, 4)[transformed_inds, :].reshape( batch_size, -1, 4) cls_pred = cls_pred.reshape( -1, self.num_classes)[transformed_inds, :].reshape( batch_size, -1, self.num_classes) conf_pred = conf_pred.reshape(-1, 1)[transformed_inds].reshape( batch_size, -1) # Save the result of current scale multi_lvl_bboxes.append(bbox_pred) multi_lvl_cls_scores.append(cls_pred) multi_lvl_conf_scores.append(conf_pred) # Merge the results of different scales together batch_mlvl_bboxes = torch.cat(multi_lvl_bboxes, dim=1) batch_mlvl_scores = torch.cat(multi_lvl_cls_scores, dim=1) batch_mlvl_conf_scores = torch.cat(multi_lvl_conf_scores, dim=1) # Replace multiclass_nms with ONNX::NonMaxSuppression in deployment from mmdet.core.export import add_dummy_nms_for_onnx conf_thr = cfg.get('conf_thr', -1) score_thr = cfg.get('score_thr', -1) # follow original pipeline of YOLOv3 if conf_thr > 0: mask = (batch_mlvl_conf_scores >= conf_thr).float() batch_mlvl_conf_scores *= mask if score_thr > 0: mask = (batch_mlvl_scores > score_thr).float() batch_mlvl_scores *= mask batch_mlvl_conf_scores = batch_mlvl_conf_scores.unsqueeze(2).expand_as( batch_mlvl_scores) batch_mlvl_scores = batch_mlvl_scores * batch_mlvl_conf_scores if with_nms: max_output_boxes_per_class = cfg.nms.get( 'max_output_boxes_per_class', 200) iou_threshold = cfg.nms.get('iou_threshold', 0.5) # keep aligned with original pipeline, improve # mAP by 1% for YOLOv3 in ONNX score_threshold = 0 nms_pre = cfg.get('deploy_nms_pre', -1) return add_dummy_nms_for_onnx( batch_mlvl_bboxes, batch_mlvl_scores, max_output_boxes_per_class, iou_threshold, score_threshold, nms_pre, cfg.max_per_img, ) else: return batch_mlvl_bboxes, batch_mlvl_scores