def get_hard_samples(self, hm_pred, wh_pred, reg_pred, gt_hm, gt_inds,
gt_bboxes, num_levels, down_ratios):
#op last layer of HG
hm_pred = hm_pred[num_levels]
wh_pred = wh_pred[num_levels]
reg_pred = reg_pred[num_levels]
gt_hm = gt_hm[num_levels]
gt_inds = gt_inds[num_levels]
down_ratio = down_ratios[num_levels]
batch, cat, height, width = hm_pred.size()
hard_bboxes_pred, hard_gt_bbox = [], []
hard_hm_pred, hard_gt_hm = [], []
for i in range(batch):
_gt_hm = gt_hm[i].unsqueeze(0)
_hm_pred = hm_pred[i].unsqueeze(0)
_gt_bbox = gt_bboxes[i] / down_ratio[0] ## WARNING: hard code
_reg_pred = reg_pred[i].unsqueeze(0)
_wh_pred = wh_pred[i].unsqueeze(0)
K = gt_bboxes[i].shape[0] #number of gt
gt_box_inds = gt_inds[i][:K].unsqueeze(0)
ys = (gt_box_inds / width).int().float()
xs = (gt_box_inds % width).int().float()
# decode pred_boxes according to the gt_hm
_reg_pred = transpose_and_gather_feat(_reg_pred, gt_box_inds)
_reg_pred = _reg_pred.view(1, K, 2)
xs = xs.view(1, K, 1) + _reg_pred[:, :, 0:1]
ys = ys.view(1, K, 1) + _reg_pred[:, :, 1:2]
_wh_pred = transpose_and_gather_feat(_wh_pred, gt_box_inds)
_wh_pred = _wh_pred.view(1, K, 2)
bboxes_pred = torch.cat([
xs - _wh_pred[..., 0:1] / 2, ys - _wh_pred[..., 1:2] / 2,
xs + _wh_pred[..., 0:1] / 2, ys + _wh_pred[..., 1:2] / 2
],
dim=2)
bboxes_pred = bboxes_pred.squeeze(0)
hard_gt_bbox.append(_gt_bbox)
hard_bboxes_pred.append(bboxes_pred)
# hm dynamic_training
_gt_hm = transpose_and_gather_feat(_gt_hm,
gt_box_inds) #([1, K, cat])
_hm_pred = transpose_and_gather_feat(_hm_pred,
gt_box_inds) #([1, K, cat])
hard_gt_hm.append(_gt_hm)
hard_hm_pred.append(_hm_pred)
return hard_bboxes_pred, hard_gt_bbox, hard_hm_pred, hard_gt_hm
def forward(self, output, mask, ind, target):
pred = transpose_and_gather_feat(output, ind)
mask = mask.unsqueeze(2).expand_as(pred).float()
loss = F.l1_loss(pred * mask,
target * mask,
reduction='elementwise_mean')
return loss
def forward(self, output, mask, ind, target):
pred = transpose_and_gather_feat(output, ind)
mask = mask.float()
# loss = F.l1_loss(pred * mask, target * mask, reduction='elementwise_mean')
loss = F.l1_loss(pred * mask, target * mask, size_average=False)
loss = loss / (mask.sum() + 1e-4)
return loss
def forward(self, output, mask, ind, target):
pred = transpose_and_gather_feat(output, ind)
mask = mask.unsqueeze(2).expand_as(pred).float()
# loss = F.l1_loss(pred * mask, target * mask, reduction='elementwise_mean')
pred = pred / (target + 1e-4)
target = target * 0 + 1
loss = F.l1_loss(pred * mask, target * mask, size_average=False)
loss = loss / (mask.sum() + 1e-4)
return loss
def forward(self, output, mask, ind, rotbin, rotres):
pred = transpose_and_gather_feat(output, ind)
loss = compute_rot_loss(pred, rotbin, rotres, mask)
return loss
def forward(self, output, mask, ind, target):
pred = transpose_and_gather_feat(output, ind)
loss = _reg_loss(pred, target, mask)
return loss