def compute_loss(pred_encoded, gt_encoded, loss_weights=[1., 1., 1., 1.]):
# Expand tuples
score, pos_offsets, dim_offsets, ang_offsets = pred_encoded
heatmaps, gt_pos_offsets, gt_dim_offsets, gt_ang_offsets, mask = gt_encoded
score_weight, pos_weight, dim_weight, ang_weight = loss_weights
# Compute losses
score_loss = huber_loss(score, heatmaps)
pos_loss = huber_loss(pos_offsets, gt_pos_offsets, mask.unsqueeze(2))
dim_loss = huber_loss(dim_offsets, gt_dim_offsets, mask.unsqueeze(2))
ang_loss = huber_loss(ang_offsets, gt_ang_offsets, mask.unsqueeze(2))
# Combine loss
total_loss = score_loss * score_weight + pos_loss * pos_weight \
+ dim_loss * dim_weight + ang_loss * ang_weight
# Store scalar losses in a dictionary
loss_dict = {
'score': float(score_loss),
'position': float(pos_loss),
'dimension': float(dim_loss),
'angle': float(ang_loss),
'total': float(total_loss)
}
total_loss = score_loss
return total_loss, loss_dict
def compute_loss(pred_encoded, gt_encoded, loss_weights=[1., 1., 1., 1.]):
# Expand tuples
score, pos_offsets, dim_offsets, ang_offsets = pred_encoded
labels, sqr_dists, gt_pos_offsets, gt_dim_offsets, gt_ang_offsets = gt_encoded
score_weight, pos_weight, dim_weight, ang_weight = loss_weights
# Compute losses
score_loss = oft.model.loss.balanced_cross_entropy_loss(score, labels)
pos_loss = huber_loss(pos_offsets, gt_pos_offsets, labels.unsqueeze(2))
dim_loss = huber_loss(dim_offsets, gt_dim_offsets, labels.unsqueeze(2))
ang_loss = huber_loss(ang_offsets, gt_ang_offsets, labels.unsqueeze(2))
# Combine loss
total_loss = score_loss * score_weight + pos_loss * pos_weight \
+ dim_loss * dim_weight + ang_loss * ang_weight
# Store scalar losses in a dictionary
loss_dict = {
'score': float(score_loss),
'position': float(pos_loss),
'dimension': float(dim_loss),
'angle': float(ang_loss),
'total': float(total_loss)
}
total_loss = score_loss
return total_loss, loss_dict