| | |
|---|
| | | # Regression loss coefficient |
|---|
| | | alpha = args.alpha |
|---|
| | | |
|---|
| | | idx_tensor = [idx for idx in xrange(66)] |
|---|
| | | idx_tensor = Variable(torch.FloatTensor(idx_tensor)).cuda(gpu) |
|---|
| | | |
|---|
| | | optimizer = torch.optim.Adam([{'params': get_ignored_params(model), 'lr': 0}, |
|---|
| | | {'params': get_non_ignored_params(model), 'lr': args.lr}, |
|---|
| | | {'params': get_fc_params(model), 'lr': args.lr * 5}], |
|---|
| | |
|---|
| | | label_roll_cont = Variable(cont_labels[:,2]).cuda(gpu) |
|---|
| | | |
|---|
| | | # Forward pass |
|---|
| | | yaw, pitch, roll, angles = model(images) |
|---|
| | | pre_yaw, pre_pitch, pre_roll = model(images) |
|---|
| | | |
|---|
| | | # Cross entropy loss |
|---|
| | | loss_yaw = criterion(yaw, label_yaw) |
|---|
| | | loss_pitch = criterion(pitch, label_pitch) |
|---|
| | | loss_roll = criterion(roll, label_roll) |
|---|
| | | loss_yaw = criterion(pre_yaw, label_yaw) |
|---|
| | | loss_pitch = criterion(pre_pitch, label_pitch) |
|---|
| | | loss_roll = criterion(pre_roll, label_roll) |
|---|
| | | |
|---|
| | | # MSE loss |
|---|
| | | yaw_predicted = angles[:,0] |
|---|
| | | pitch_predicted = angles[:,1] |
|---|
| | | roll_predicted = angles[:,2] |
|---|
| | | yaw_predicted = softmax(pre_yaw) |
|---|
| | | pitch_predicted = softmax(pre_pitch) |
|---|
| | | roll_predicted = softmax(pre_roll) |
|---|
| | | |
|---|
| | | yaw_predicted = torch.sum(yaw_predicted * idx_tensor, 1) * 3 - 99 |
|---|
| | | pitch_predicted = torch.sum(pitch_predicted * idx_tensor, 1) * 3 - 99 |
|---|
| | | roll_predicted = torch.sum(roll_predicted * idx_tensor, 1) * 3 - 99 |
|---|
| | | |
|---|
| | | loss_reg_yaw = reg_criterion(yaw_predicted, label_yaw_cont) |
|---|
| | | loss_reg_pitch = reg_criterion(pitch_predicted, label_pitch_cont) |
|---|
| | |
|---|
| | | loss_roll += alpha * loss_reg_roll |
|---|
| | | |
|---|
| | | loss_seq = [loss_yaw, loss_pitch, loss_roll] |
|---|
| | | grad_seq = [torch.Tensor(1).cuda(gpu) for _ in range(len(loss_seq))] |
|---|
| | | optimizer.zero_grad() |
|---|
| | | grad_seq = [torch.ones(1).cuda(gpu) for _ in range(len(loss_seq))] |
|---|
| | | torch.autograd.backward(loss_seq, grad_seq) |
|---|
| | | optimizer.step() |
|---|
| | | |
|---|
