if args.mode == 'ddl' or args.mode == 'ddldr':
      out = F.softmax(output, dim=1)
      norm = torch.norm(out, dim=1)
      out = out / norm[:, None]
      idx = []
      for i in range(len(out)):
          if out[i,target[i]] > .5:
              idx.append(i)
              
      if len(idx) > 0:
          target[idx] = disturb(target[idx]).to(device) 

def noise_generator(x, alpha):
    noise = torch.normal(0, 1e-8, size=(len(x), 1))
    noise[torch.randint(0, len(x), (int(len(x)*(1-alpha)),))] = 0

    return noise

def disturberror(outputs, values):
    epsilon = 1e-8
    e = values - outputs
    for i in range(len(e)):
        if (e[i] < epsilon) & (e[i] >= 0):
            values[i] = values[i] + e[i] / 4
        elif (e[i] > -epsilon) & (e[i] < 0):
            values[i] = values[i] - e[i] / 4

    return values

python main_new.py --de y --dataset "bike" --dv_annealing y --epoch 100 --T 80
python main_new.py --de y --dv y --dataset "bike" -epoch 100
python main_new.py --de y --l2 y --dataset "air" -epoch 100
python main_new.py --dv y --dv_annealing y --dataset "air" -epoch 100 #for annealing setting dv should be "y"

--dataset: 'bike', 'air', 'boston', 'housing', 'make_sklearn', 'superconduct', 'energy', 'crime', 'students'
--dropout, --dv(disturbvalue), --de(disturberror), --l2, --dv_annealing: (string) y / n
--lr: (float)
--batch_size, --epoch, --T(cos annealing T): (int)
-- default dv_annealing: alpha_min = 0.05, alpha_max = 0.12, T_i = 80