def __call__(self, pred_cls_list, pred_reg_list, anchors_list, gt_boxes,
im_info):
all_anchors = torch.cat(anchors_list, axis=0)
all_pred_cls = torch.cat(pred_cls_list,
axis=1).reshape(-1, config.num_classes - 1)
all_pred_cls = torch.sigmoid(all_pred_cls)
all_pred_reg = torch.cat(pred_reg_list, axis=1).reshape(-1, 4)
# get ground truth
labels, bbox_target = retina_anchor_target(all_anchors,
gt_boxes,
im_info,
top_k=1)
# regression loss
fg_mask = (labels > 0).flatten()
valid_mask = (labels >= 0).flatten()
loss_reg = smooth_l1_loss(all_pred_reg[fg_mask], bbox_target[fg_mask],
config.smooth_l1_beta)
loss_cls = focal_loss(all_pred_cls[valid_mask], labels[valid_mask],
config.focal_loss_alpha, config.focal_loss_gamma)
num_pos_anchors = fg_mask.sum().item()
self.loss_normalizer = self.loss_normalizer_momentum * self.loss_normalizer + (
1 - self.loss_normalizer_momentum) * max(num_pos_anchors, 1)
loss_reg = loss_reg.sum() / self.loss_normalizer
loss_cls = loss_cls.sum() / self.loss_normalizer
loss_dict = {}
loss_dict['retina_focal_loss'] = loss_cls
loss_dict['retina_smooth_l1'] = loss_reg
return loss_dict
def forward(self, features, im_info, boxes=None):
# prediction
pred_cls_score_list = []
pred_bbox_offsets_list = []
batch=features[0].shape[0]
for x in features:
t = F.relu(self.rpn_conv(x))
pred_cls_score_list.append(self.rpn_cls_score(t))
pred_bbox_offsets_list.append(self.rpn_bbox_offsets(t))
# get anchors
all_anchors_list = []
# stride: 64,32,16,8,4 p6->p2
base_stride = 4
off_stride = 2**(len(features)-1) # 16
for fm in features:
layer_anchors = self.anchors_generator(fm, base_stride, off_stride)
off_stride = off_stride // 2
all_anchors_list.append(layer_anchors)
# sample from the predictions
rpn_rois,shapes = find_top_rpn_proposals(
self.training, pred_bbox_offsets_list, pred_cls_score_list,
all_anchors_list, im_info,batch)
rpn_rois = rpn_rois.type_as(features[0])
if self.training:
rpn_labels, rpn_bbox_targets = fpn_anchor_target(
boxes, im_info, all_anchors_list)
#rpn_labels = rpn_labels.astype(np.int32)
pred_cls_score, pred_bbox_offsets = fpn_rpn_reshape(
pred_cls_score_list, pred_bbox_offsets_list)
# rpn loss
valid_masks = rpn_labels >= 0
objectness_loss = softmax_loss(
pred_cls_score[valid_masks],
rpn_labels[valid_masks])
pos_masks = rpn_labels > 0
localization_loss = smooth_l1_loss(
pred_bbox_offsets[pos_masks],
rpn_bbox_targets[pos_masks],
config.rpn_smooth_l1_beta)
normalizer = 1 / valid_masks.sum().item()
loss_rpn_cls = objectness_loss.sum() * normalizer
loss_rpn_loc = localization_loss.sum() * normalizer
loss_dict = {}
loss_dict['loss_rpn_cls'] = loss_rpn_cls
loss_dict['loss_rpn_loc'] = loss_rpn_loc
return rpn_rois, loss_dict
else:
return rpn_rois,shapes
def forward(self, fpn_fms, proposals, labels=None, bbox_targets=None):
# input p2-p5
fpn_fms = fpn_fms[1:][::-1]
stride = [4, 8, 16, 32]
pool_features = roi_pooler(fpn_fms, proposals, stride, (7, 7),
"ROIAlignV2")
flatten_feature = torch.flatten(pool_features, start_dim=1)
flatten_feature = F.relu_(self.fc1(flatten_feature))
flatten_feature = F.relu_(self.fc2(flatten_feature))
pred_cls = self.pred_cls(flatten_feature)
pred_delta = self.pred_delta(flatten_feature)
if self.training:
labels = labels.long().flatten()
fg_masks = labels > 0
valid_masks = labels >= 0
# loss for regression
localization_loss = smooth_l1_loss(pred_delta[fg_masks],
bbox_targets[fg_masks],
config.rcnn_smooth_l1_beta)
# loss for classification
objectness_loss = softmax_loss(pred_cls, labels)
objectness_loss = objectness_loss * valid_masks
normalizer = 1.0 / valid_masks.sum().item()
loss_rcnn_loc = localization_loss.sum() * normalizer
loss_rcnn_cls = objectness_loss.sum() * normalizer
loss_dict = {}
loss_dict[self.stage_name + '_loc'] = loss_rcnn_loc
loss_dict[self.stage_name + '_cls'] = loss_rcnn_cls
# proposals
with torch.no_grad():
pred_bbox = restore_bbox(proposals[:, 1:5], pred_delta,
True).detach()
pred_proposals = torch.cat(
[proposals[:, 0].reshape(-1, 1), pred_bbox], axis=1)
#pred_proposals = batch_clip_boxes_opr(pred_proposals, im_info)
return pred_proposals, loss_dict
else:
pred_bbox = restore_bbox(proposals[:, 1:5], pred_delta,
True).detach()
pred_proposals = torch.cat(
[proposals[:, 0].reshape(-1, 1), pred_bbox], axis=1)
pred_scores = F.softmax(pred_cls, dim=-1)[:, 1].reshape(-1, 1)
return pred_proposals, pred_scores
def forward(self, fpn_fms, rcnn_rois, labels=None, bbox_targets=None):
# input p2-p5
fpn_fms = fpn_fms[1:][::-1]
stride = [4, 8, 16, 32]
pool_features = roi_pooler(fpn_fms, rcnn_rois, stride, (7, 7),
"ROIAlignV2")
flatten_feature = torch.flatten(pool_features, start_dim=1)
flatten_feature = F.relu_(self.fc1(flatten_feature))
flatten_feature = F.relu_(self.fc2(flatten_feature))
pred_cls = self.pred_cls(flatten_feature)
pred_delta = self.pred_delta(flatten_feature)
if self.training:
# loss for regression
labels = labels.long().flatten()
fg_masks = labels > 0
valid_masks = labels >= 0
# multi class
pred_delta = pred_delta.reshape(-1, config.num_classes, 4)
fg_gt_classes = labels[fg_masks]
pred_delta = pred_delta[fg_masks, fg_gt_classes, :]
localization_loss = smooth_l1_loss(pred_delta,
bbox_targets[fg_masks],
config.rcnn_smooth_l1_beta)
# loss for classification
objectness_loss = softmax_loss(pred_cls, labels)
objectness_loss = objectness_loss * valid_masks
normalizer = 1.0 / valid_masks.sum().item()
loss_rcnn_loc = localization_loss.sum() * normalizer
loss_rcnn_cls = objectness_loss.sum() * normalizer
loss_dict = {}
loss_dict['loss_rcnn_loc'] = loss_rcnn_loc
loss_dict['loss_rcnn_cls'] = loss_rcnn_cls
return loss_dict
else:
class_num = pred_cls.shape[-1] - 1
tag = torch.arange(class_num).type_as(pred_cls) + 1
tag = tag.repeat(pred_cls.shape[0], 1).reshape(-1, 1)
pred_scores = F.softmax(pred_cls, dim=-1)[:, 1:].reshape(-1, 1)
pred_delta = pred_delta[:, 4:].reshape(-1, 4)
base_rois = rcnn_rois[:, 1:5].repeat(1, class_num).reshape(-1, 4)
pred_bbox = restore_bbox(base_rois, pred_delta, True)
pred_bbox = torch.cat([pred_bbox, pred_scores, tag], axis=1)
return pred_bbox
