def forward(self, image, im_info, gt_boxes=None):
image = image - torch.tensor(config.image_mean[None, :, None, None],
dtype=image.dtype, device=image.device)
image = get_padded_tensor(image, 64)
if self.training:
return self._forward_train(image, im_info, gt_boxes)
else:
return self._forward_test(image, im_info)
def forward(self, image, im_info, gt_boxes=None):
image = (image - torch.tensor(config.image_mean[None, :, None, None]).type_as(image)) / (
torch.tensor(config.image_std[None, :, None, None]).type_as(image))
image = get_padded_tensor(image, 64)
if self.training:
return self._forward_train(image, im_info, gt_boxes)
else:
return self._forward_test(image, im_info)
def forward(self, image, im_info, gt_boxes=None):
# pre-processing the data
image = (image - torch.tensor(config.image_mean[
None, :, None, None]).type_as(image)) / (torch.tensor(
config.image_std[None, :, None, None]).type_as(image))
image = get_padded_tensor(image, 64)
# do inference
# stride: 128,64,32,16,8, p7->p3
fpn_fms = self.FPN(image)
anchors_list = self.R_Anchor(fpn_fms)
pred_cls_list, pred_reg_list = self.R_Head(fpn_fms)
# release the useless data
if self.training:
loss_dict = self.R_Criteria(pred_cls_list, pred_reg_list,
anchors_list, gt_boxes, im_info)
return loss_dict
else:
#pred_bbox = union_inference(
# anchors_list, pred_cls_list, pred_reg_list, im_info)
return anchors_list, pred_cls_list, pred_reg_list, im_info
