num_epochs = 10
クラスを２に指定して、二値分類のセグメンテーションを行うことが目標です。
参考にした Webページの内容より、可能性の一つはクラス数とネットワークモデルの入出力の不一致とありましたが、その方は大丈夫でした。
エラー文より、損失関数の計算の箇所で問題がありそうです。
loss = criterion(outputs, anno_class_imges.long()) / batch_multiplier
Python
1 RuntimeError Traceback ( most recent call last )
2 Input In [ 15 ] , in < cell line : 3 > ( )
3 1 num_epochs = 10
4 2 dataloaders_dict = { "train" : train_dataloader , "val" : val_dataloader }
5 - - - - > 3 train_loss_list , val_loss_list = train_model ( net , dataloaders_dict , criterion , scheduler , optimizer , num_epochs )
6
7 Input In [ 14 ] , in train_model ( net , dataloaders_dict , criterion , scheduler , optimizer , num_epochs )
8 72 with torch . set_grad_enabled ( phase == 'train' ) :
9 73 outputs = net ( imges )
10 - - - > 74 loss = criterion ( outputs , anno_class_imges . long ( ) ) / batch_multiplier
11 76 # 訓練時はバックプロパゲーション
12 77 if phase == 'train' :
13
14 File ~ \anaconda3\envs\copy38\lib\site - packages\torch\nn\modules\module . py : 1110 , in Module . _call_impl ( self , * input , ** kwargs )
15 1106 # If we don't have any hooks, we want to skip the rest of the logic in
16 1107 # this function, and just call forward.
17 1108 if not ( self . _backward_hooks or self . _forward_hooks or self . _forward_pre_hooks or _global_backward_hooks
18 1109 or _global_forward_hooks or _global_forward_pre_hooks ) :
19 - > 1110 return forward_call ( * input , ** kwargs )
20 1111 # Do not call functions when jit is used
21 1112 full_backward_hooks , non_full_backward_hooks = [ ] , [ ]
22
23 Input In [ 12 ] , in PSPLoss . forward ( self , outputs , targets )
24 8 def forward ( self , outputs , targets ) :
25 9 """
26 10 損失関数の計算。
27 11
28 (...)
29 23 損失の値
30 24 """
31 - - - > 26 loss = F . cross_entropy ( outputs [ 0 ] , targets , reduction = 'mean' )
32 27 loss_aux = F . cross_entropy ( outputs [ 1 ] , targets , reduction = 'mean' )
33 29 return loss + self . aux_weight * loss_aux
34
35 File ~ \anaconda3\envs\copy38\lib\site - packages\torch\nn\functional . py : 2996 , in cross_entropy ( input , target , weight , size_average , ignore_index , reduce , reduction , label_smoothing )
36 2994 if size_average is not None or reduce is not None :
37 2995 reduction = _Reduction . legacy_get_string ( size_average , reduce )
38 - > 2996 return torch . _C . _nn . cross_entropy_loss ( input , target , weight , _Reduction . get_enum ( reduction ) , ignore_index , label_smoothing )
39
40
41
42
43 RuntimeError : CUDA error : device - side assert triggered
44
使用したコードの一部を貼ります
Python
1 #損失関数
2 class PSPLoss ( nn . Module ) :
3
4
5 def __init__ ( self , aux_weight = 0.4 ) :
6 super ( PSPLoss , self ) . __init__ ( )
7 self . aux_weight = aux_weight # aux_lossの重み
8
9 def forward ( self , outputs , targets ) :
10
11 loss = F . cross_entropy ( outputs [ 0 ] , targets , reduction = 'mean' )
12 loss_aux = F . cross_entropy ( outputs [ 1 ] , targets , reduction = 'mean' )
13
14 return loss + self . aux_weight * loss_aux
15
16
17 criterion = PSPLoss ( aux_weight = 0.4 )
18
Python
1 def train_model ( net , dataloaders_dict , criterion , scheduler , optimizer , num_epochs ) :
2
3 train_loss_list = [ ]
4 val_loss_list = [ ]
5 # GPUが使えるかを確認
6 device = torch . device ( "cuda:0" if torch . cuda . is_available ( ) else "cpu" )
7 print ( "使用デバイス：" , device )
8
9 # ネットワークをGPUへ
10 net . to ( device )
11
12 # ネットワークがある程度固定であれば、高速化させる
13 torch . backends . cudnn . benchmark = True
14
15 # 画像の枚数
16 num_train_imgs = 100
17 num_val_imgs = 30
18 batch_size = 4
19
20 # イテレーションカウンタをセット
21 iteration = 1
22 logs = [ ]
23
24 # multiple minibatch
25 batch_multiplier = 3
26
27 # epochのループ
28 for epoch in range ( num_epochs ) :
29
30 # 開始時刻を保存
31 t_epoch_start = time . time ( )
32 t_iter_start = time . time ( )
33 epoch_train_loss = 0.0 # epochの損失和
34 epoch_val_loss = 0.0 # epochの損失和
35
36 print ( '-------------' )
37 print ( 'Epoch {}/{}' . format ( epoch + 1 , num_epochs ) )
38 print ( '-------------' )
39
40 # epochごとの訓練と検証のループ
41 for phase in [ 'train' , 'val' ] :
42 if phase == 'train' :
43 net . train ( ) # モデルを訓練モードに
44 scheduler . step ( ) # 最適化schedulerの更新
45 optimizer . zero_grad ( )
46 print ( '（train）' )
47
48 else :
49 net . eval ( ) # モデルを検証モードに
50 print ( '-------------' )
51 print ( '（val）' )
52
53 # データローダーからminibatchずつ取り出すループ
54 count = 0 # multiple minibatch
55 for imges , anno_class_imges in dataloaders_dict [ phase ] :
56 # ミニバッチがサイズが1だと、バッチノーマライゼーションでエラーになるのでさける
57 if imges . size ( ) [ 0 ] == 1 :
58 continue
59
60 # GPUが使えるならGPUにデータを送る
61 imges = imges . to ( device , dtype = torch . float )
62 anno_class_imges = anno_class_imges . to ( device , dtype = torch . float )
63
64
65 # multiple minibatchでのパラメータの更新
66 if ( phase == 'train' ) and ( count == 0 ) :
67 optimizer . step ( )
68 optimizer . zero_grad ( )
69 count = batch_multiplier
70
71 # 順伝搬（forward）計算
72 with torch . set_grad_enabled ( phase == 'train' ) :
73 outputs = net ( imges )
74 loss = criterion ( outputs , anno_class_imges . long ( ) ) / batch_multiplier
75
76 # 訓練時はバックプロパゲーション
77 if phase == 'train' :
78 loss . backward ( ) # 勾配の計算
79 count -= 1 # multiple minibatch
80
81 if ( iteration % 10 == 0 ) : # 10iterに1度、lossを表示
82 t_iter_finish = time . time ( )
83 duration = t_iter_finish - t_iter_start
84 print ( 'Iteration {} || Loss: {:.4f} || 10iter: {:.4f} sec.' . format (
85 iteration , loss . item ( ) / batch_size * batch_multiplier , duration ) )
86 t_iter_start = time . time ( )
87
88 epoch_train_loss += loss . item ( ) * batch_multiplier
89 iteration += 1
90
91 # 検証時
92 else :
93 epoch_val_loss += loss . item ( ) * batch_multiplier
94
95 # epochのphaseごとのlossと正解率
96 t_epoch_finish = time . time ( )
97 print ( '-------------' )
98 print ( 'epoch {} || Epoch_TRAIN_Loss:{:.4f} ||Epoch_VAL_Loss:{:.4f}' . format (
99 epoch + 1 , epoch_train_loss / num_train_imgs , epoch_val_loss / num_val_imgs ) )
100 print ( 'timer: {:.4f} sec.' . format ( t_epoch_finish - t_epoch_start ) )
101 t_epoch_start = time . time ( )
102
103 #lossの可視化のリスト
104 train_loss_list . append ( epoch_train_loss / num_train_imgs )
105 val_loss_list . append ( epoch_val_loss / num_val_imgs )
106 if ( ( epoch + 1 ) % 5 == 0 ) :
107 torch . save ( net . state_dict ( ) , 'weights/pspnet50_' + str ( epoch + 1 ) + '.pth' )
108 # ログを保存
109
110 log_epoch = { 'epoch' : epoch + 1 , 'train_loss' : epoch_train_loss /
111 num_train_imgs , 'val_loss' : epoch_val_loss / num_val_imgs }
112 logs . append ( log_epoch )
113 df = pd . DataFrame ( logs )
114 df . to_csv ( "log_output.csv" )
115
116 # 最後のネットワークを保存する
117 torch . save ( net . state_dict ( ) , 'weights/pspnet50_' +
118 str ( epoch + 1 ) + '.pth' )
119 return train_loss_list , val_loss_list
Q&A
