confidence_lossがnanになる

[kosuke55]

confidence_loss 0.056517358869314194
depth_loss 0.004850194789469242
rotation_loss 0.3008980453014374
epoch 4, 334/1266,train loss is confidence:21.170931436121464 rotation:91.45359233021736 depth:10.529736892785877
train epoch=4:  26%|███████████████▌                                           | 335/1266 [28:35<1:31:47,  5.92s/it]/home/kosuke55/catkin_ws/src/hanging_points_cnn/hanging_points_cnn/utils/rois_tools.py:176: RuntimeWarning: invalid value encountered in greater
  confidence_mask[confidence_mask > 1] = 1
/home/kosuke55/catkin_ws/src/hanging_points_cnn/hanging_points_cnn/utils/rois_tools.py:177: RuntimeWarning: invalid value encountered in less
  confidence_mask[confidence_mask < 0] = 0
train_hpnet.py:159: RuntimeWarning: invalid value encountered in greater_equal
  confidence_np[confidence_np >= 1] = 1.
train_hpnet.py:160: RuntimeWarning: invalid value encountered in less_equal
  confidence_np[confidence_np <= 0] = 0.
confidence_loss nan
depth_loss 0.0
rotation_loss 0.0

[kosuke55]

/media/kosuke/SANDISK/hanging_points_net/checkpoints/gray/hpnet_latestmodel_20200812_2224.pt

[kosuke55]

ネットワークの出力自体がnan

ipdb> hp_data
tensor([[[[0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          ...,
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000]],

         [[0.3686, 0.3686, 0.3765,  ..., 0.4275, 0.4275, 0.4275],
          [0.3686, 0.3686, 0.3765,  ..., 0.4314, 0.4314, 0.4314],
          [0.3725, 0.3725, 0.3804,  ..., 0.4392, 0.4392, 0.4392],
          ...,
          [0.4000, 0.4000, 0.3961,  ..., 0.4196, 0.4196, 0.4196],
          [0.4039, 0.4039, 0.4000,  ..., 0.4196, 0.4196, 0.4196],
          [0.4039, 0.4039, 0.4000,  ..., 0.4196, 0.4196, 0.4196]]]],
       device='cuda:0')
ipdb> self.model(hp_data)
(tensor([[[[nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          ...,
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan]]]], device='cuda:0',
       grad_fn=<ReluBackward1>), tensor([[nan, nan, nan, nan, nan]], device='cuda:0', grad_fn=<AddmmBackward>))

[kosuke55]

ひとつ前のmodelとoptimizerを保存する？？
https://qiita.com/syoamakase/items/a9b3146e09f9fcafbb66

            if torch.isnan(loss):
                print('loss is nan!!')
                self.model = self.prev_model
                self.optimizer = torch.optim.Adam(
                    self.prev_model.parameters(), lr=args.lr,  betas=(0.9, 0.999),
                    eps=1e-10, weight_decay=0, amsgrad=False)
                self.optimizer.load_state_dict(
                    self.prev_optimizer.state_dict())
                continue
            else:
                self.prev_model = copy.deepcopy(self.model)
                self.prev_optimizer = copy.deepcopy(self.optimizer)

[kosuke55]

どこかの勾配が大きくなっている。これで大きいところをクリップする
https://pytorch.org/docs/master/generated/torch.nn.utils.clip_grad_norm_.html

torch.nn.utils.clip_grad_norm_(parameters, max_norm, norm_type=2)

https://discuss.pytorch.org/t/how-to-check-norm-of-gradients/13795/2
より

Q: How do we choose the hyperparameter c?
A: We can train our neural networks for some epochs and look at the statistics of the gradient norms. The average value of gradient norms is a good initial trial.

clipは数エポックのgradient normの平均にするとよい。

求め方
https://towardsdatascience.com/what-is-gradient-clipping-b8e815cdfb48

In [76]:         for p in self.model.parameters():
    ...:             param_norm = p.grad.data.norm(2)
    ...:             total_norm += param_norm.item() ** 2
    ...:         total_norm = total_norm ** (1. / 2)

[kosuke55]

clip_grad_normしてもnanになってしまう。。
#1 (comment)
でバッチサイズを下げれば(64->16)とりあえずはnanで学習が止まらることはない。

[kosuke55]

v_predが小さくなっている？

In [122]: v_pred = torch.Tensor([1e-30, 0, 0])
     ...: print(torch.norm(v_pred))
     ...: v_pred_n = v_pred / torch.norm(v_pred )
     ...: print(v_pred_n)
     ...:
tensor(0.)
tensor([inf, nan, nan])