小编Gee*_*der的帖子

尝试实现研究论文： https://ieeexplore.ieee.org/document/9479786/ 使用架构训练单调网络：

class Model(nn.Module):
  def __init__(self, q, s):
    self.layer_s_list = [nn.Linear(5, s) for _ in range(q)]
    self.inv_w, self.inv_b = self.get_layer_weights()
      
  def forward(self, x):
    # print(inv_w[0].shape, inv_b[0].shape)
    output_lst = []
    for layer in self.layer_s_list:
      v, id = torch.max(layer(x), 1)
      output_lst.append(v.detach().numpy())
    output_lst = np.array(output_lst)
    output_lst = torch.from_numpy(output_lst)
    out, _ = torch.min(output_lst, 0)
    allo_out = F.softmax(out)
    pay_out = nn.ReLU(inplace = True)(out)
    inv_out_lst = []
    
    for q_idx in range(len(self.inv_w)):
      # print(inv_w[q_idx].shape, pay_out.shape, inv_b[q_idx].shape)
      y, _ = torch.min(torch.linalg.pinv(self.inv_w[q_idx]) * (pay_out - self.inv_b[q_idx]), 0)
      inv_out_lst.append(y.detach().numpy()) …