概要

以下のgithubに、いま注目されているViT（Vision Transformer）の実装が示されていたので、動かしてみた。

(1)
https://github.com/kamalkraj/Vision-Transformer

対象は、cifar10でした。jupyter notebookの形で、使い方が示されていた。

(2)
https://github.com/emla2805/vision-transformer

こちらは、対象は、imagenet_resized/32x32でした。

(3)
https://github.com/tuvovan/Vision_Transformer_Keras
も試しました。これは、cifar10。

(4)
https://github.com/kentaroy47/vision-transformers-cifar10
こちらも、cifar10。80％超えましたっ！

動かしてみた結果(まとめ）

	URL	Cifar10/imagenet32x32	結果(%)	備考
1	https://github.com/kamalkraj/Vision-Transformer	cifar10	48.9
2	https://github.com/emla2805/vision-transformer	imagenet32x32	9.2(途中）	途中ですが、見込みなしと判断
3	https://github.com/tuvovan/Vision_Transformer_Keras	cifar10	63%
4	https://github.com/kentaroy47/vision-transformers-cifar10	cifar10	★80.9

cifar10で、80.2%でた！すごい。
↑どこが優れていいるのか、調べていません。

動かしてみた結果（https://github.com/kamalkraj/Vision-Transformer)

cifar10。エポック数は10になっていたので、300に延ばした。

出だしは、以下のような感じ。悪くない。計算も速かった。

epoch 1: train loss 1.79971. train accuracy 0.34166
epoch 1: test loss 1.56127. test accuracy 0.43050
epoch 2: train loss 1.51326. train accuracy 0.44944
epoch 2: test loss 1.45918. test accuracy 0.45840
epoch 3: train loss 1.41494. train accuracy 0.48768
epoch 3: test loss 1.39249. test accuracy 0.49690

300エポックで、以下のような感じ。
(非力なGPUで半日程度で完了しました。）

epoch 295: train loss 0.11632. train accuracy 0.96030
epoch 295: test loss 5.43269. test accuracy 0.49070
epoch 296: train loss 0.10554. train accuracy 0.96354
epoch 296: test loss 5.39661. test accuracy 0.49290
epoch 297: train loss 0.10795. train accuracy 0.96264
epoch 297: test loss 5.40721. test accuracy 0.49250
epoch 298: train loss 0.11270. train accuracy 0.96154
epoch 298: test loss 5.41567. test accuracy 0.48540
epoch 299: train loss 0.10613. train accuracy 0.96350
epoch 299: test loss 5.40492. test accuracy 0.48590
epoch 300: train loss 0.10614. train accuracy 0.96536
epoch 300: test loss 5.57218. test accuracy 0.48910

test accuracy 0.48910 と、滅茶ワル！
train accuracy 0.96536　は、普通？

⇒　実装は、augmentationとかは入っていなかった。
それにしても、test accuracy 0.48910 はナイはずなので、、、
(←augmentationがない世間の実行例をみても70%とかにはなっていました。当然、モデルにも依存しますが。）
なんでしょう？？？？

動かしてみた結果（https://github.com/emla2805/vision-transformer)

バッチサイズは、メモリ不足が出て実行できなかったので、デフォルトの4096を1024に変更した。
imagenet_resized/32x32。

Epoch 1/300
1252/1252 [==============================] - 3325s 3s/step - loss: 6.3762 - accuracy: 0.0150 - val_loss: 6.0516 - val_accuracy: 0.0278
Epoch 2/300
1252/1252 [==============================] - 3686s 3s/step - loss: 5.9006 - accuracy: 0.0350 - val_loss: 5.7768 - val_accuracy: 0.0434
Epoch 3/300
1252/1252 [==============================] - 3687s 3s/step - loss: 5.7059 - accuracy: 0.0458 - val_loss: 5.6090 - val_accuracy: 0.0531
Epoch 4/300
1252/1252 [==============================] - 3688s 3s/step - loss: 5.5572 - accuracy: 0.0555 - val_loss: 5.4895 - val_accuracy: 0.0627
Epoch 5/300
1252/1252 [==============================] - 3689s 3s/step - loss: 5.4528 - accuracy: 0.0631 - val_loss: 5.3938 - val_accuracy: 0.0689
Epoch 6/300
1252/1252 [==============================] - 3685s 3s/step - loss: 5.3737 - accuracy: 0.0691 - val_loss: 5.3331 - val_accuracy: 0.0732
Epoch 7/300
1252/1252 [==============================] - 3686s 3s/step - loss: 5.3049 - accuracy: 0.0746 - val_loss: 5.2774 - val_accuracy: 0.0774
Epoch 8/300
1252/1252 [==============================] - 3685s 3s/step - loss: 5.2486 - accuracy: 0.0789 - val_loss: 5.2161 - val_accuracy: 0.0831
Epoch 9/300
1252/1252 [==============================] - 3683s 3s/step - loss: 5.1994 - accuracy: 0.0830 - val_loss: 5.1670 - val_accuracy: 0.0867
Epoch 10/300
1252/1252 [==============================] - 3684s 3s/step - loss: 5.1580 - accuracy: 0.0866 - val_loss: 5.1153 - val_accuracy: 0.0912
Epoch 11/300
1252/1252 [==============================] - 3086s 2s/step - loss: 5.1240 - accuracy: 0.0893 - val_loss: 5.0699 - val_accuracy: 0.0937
Epoch 12/300
1252/1252 [==============================] - 2549s 2s/step - loss: 5.0951 - accuracy: 0.0917 - val_loss: 5.0980 - val_accuracy: 0.0932

長いエポック数は、実行していないが、すでに、勢いが止まっているので、まともな率にはならないと見切りました。
単純には、高い率は、でない気がします。

動かしてみた結果（https://github.com/tuvovan/Vision_Transformer_Keras)

cifar10。63%。

Epoch 1/50
782/782 [==============================] - 71s 91ms/step - loss: 1.9103 - accuracy: 0.2675 - val_loss: 1.6849 - val_accuracy: 0.3607
Epoch 2/50
782/782 [==============================] - 74s 95ms/step - loss: 1.6341 - accuracy: 0.3881 - val_loss: 1.5447 - val_accuracy: 0.4328
Epoch 3/50
782/782 [==============================] - 181s 231ms/step - loss: 1.4721 - accuracy: 0.4582 - val_loss: 1.3912 - val_accuracy: 0.4886
Epoch 4/50
782/782 [==============================] - 336s 430ms/step - loss: 1.3570 - accuracy: 0.5040 - val_loss: 1.3044 - val_accuracy: 0.5215
Epoch 5/50
782/782 [==============================] - 337s 430ms/step - loss: 1.2924 - accuracy: 0.5304 - val_loss: 1.2297 - val_accuracy: 0.5506
Epoch 6/50
782/782 [==============================] - 336s 430ms/step - loss: 1.2307 - accuracy: 0.5537 - val_loss: 1.2331 - val_accuracy: 0.5489
Epoch 7/50
782/782 [==============================] - 337s 431ms/step - loss: 1.1847 - accuracy: 0.5724 - val_loss: 1.1611 - val_accuracy: 0.5777
Epoch 8/50
782/782 [==============================] - 337s 431ms/step - loss: 1.1462 - accuracy: 0.5853 - val_loss: 1.1819 - val_accuracy: 0.5754
Epoch 9/50
782/782 [==============================] - 338s 432ms/step - loss: 1.1151 - accuracy: 0.5981 - val_loss: 1.1267 - val_accuracy: 0.5923
Epoch 10/50
782/782 [==============================] - 338s 432ms/step - loss: 1.0888 - accuracy: 0.6074 - val_loss: 1.0718 - val_accuracy: 0.6131
Epoch 11/50
782/782 [==============================] - 338s 432ms/step - loss: 1.0692 - accuracy: 0.6155 - val_loss: 1.0739 - val_accuracy: 0.6144
Epoch 12/50
782/782 [==============================] - 337s 431ms/step - loss: 1.0522 - accuracy: 0.6235 - val_loss: 1.1166 - val_accuracy: 0.5950
Epoch 13/50
782/782 [==============================] - 344s 440ms/step - loss: 1.0327 - accuracy: 0.6312 - val_loss: 1.1115 - val_accuracy: 0.6009
Epoch 14/50
782/782 [==============================] - 337s 431ms/step - loss: 0.9722 - accuracy: 0.6583 - val_loss: 1.0198 - val_accuracy: ★0.6392
Epoch 15/50
782/782 [==============================] - 337s 432ms/step - loss: 1.0219 - accuracy: 0.6455 - val_loss: 1.0559 - val_accuracy: 0.6287
Epoch 16/50
782/782 [==============================] - 337s 431ms/step - loss: 1.0714 - accuracy: 0.6281 - val_loss: 1.1004 - val_accuracy: 0.6127
Epoch 17/50
782/782 [==============================] - 337s 431ms/step - loss: 1.1214 - accuracy: 0.6117 - val_loss: 1.1289 - val_accuracy: 0.6028
Epoch 18/50
782/782 [==============================] - 337s 431ms/step - loss: 1.2367 - accuracy: 0.5923 - val_loss: 1.3054 - val_accuracy: 0.5793
Epoch 19/50
782/782 [==============================] - 337s 431ms/step - loss: 1.4621 - accuracy: 0.5401 - val_loss: 1.5011 - val_accuracy: 0.5335
Epoch 20/50
782/782 [==============================] - 337s 430ms/step - loss: 1.6503 - accuracy: 0.4775 - val_loss: 1.6644 - val_accuracy: 0.4793
Epoch 21/50
782/782 [==============================] - 337s 430ms/step - loss: 1.8397 - accuracy: 0.4129 - val_loss: 1.9011 - val_accuracy: 0.3987
Epoch 22/50
782/782 [==============================] - 337s 431ms/step - loss: 2.0167 - accuracy: 0.3414 - val_loss: 2.0578 - val_accuracy: 0.3361
Epoch 23/50
782/782 [==============================] - 337s 430ms/step - loss: 2.1249 - accuracy: 0.2904 - val_loss: 2.1507 - val_accuracy: 0.2885
Epoch 24/50
782/782 [==============================] - 337s 431ms/step - loss: 2.1957 - accuracy: 0.2472 - val_loss: 2.2183 - val_accuracy: 0.2379

動かしてみた結果（https://github.com/kentaroy47/vision-transformers-cifar10)

cifar10。80.9%。

Epoch: 44
 [===============================>]  Step: 791ms | Tot: 782/782  Loss: 0.273 | Acc: 90.354% (45177/50000)4)2)
 [===============================>]  Step: 1s536ms | To 100/100 | Loss: 0.646 | Acc: 80.570% (8057/10000)0)
Epoch: 45
 [===============================>]  Step: 768ms | Tot: 782/782  Loss: 0.268 | Acc: 90.502% (45251/50000)))6)
 [===============================>]  Step: 1s653ms | To 100/100 | Loss: 0.646 | Acc: ★80.920% (8092/10000)0)
Epoch: 46
 [===============================>]  Step: 767ms | Tot 782/782  Loss: 0.255 | Acc: 90.958% (45479/50000)4)0))
 [===============================>]  Step: 1s583ms | To 100/100 | Loss: 0.645 | Acc: 80.810% (8081/10000)))
Epoch: 47
 [===============================>]  Step: 838ms | Tot: 782/782  Loss: 0.256 | Acc: 90.874% (45437/50000)4)6)
 [===============================>]  Step: 1s552ms | To 100/100 | Loss: 0.646 | Acc: 80.730% (8073/10000)))
Epoch: 48
 [===============================>]  Step: 756ms | Tot: 3 782/782  Loss: 0.257 | Acc: 90.864% (45432/50000)))
 [===============================>]  Step: 1s689ms | To 100/100 | Loss: 0.647 | Acc: 80.700% (8070/10000)))
Epoch: 49
 [===============================>]  Step: 813ms | Tot: 782/782  Loss: 0.253 | Acc: 91.026% (45513/50000)4)8)
 [===============================>]  Step: 1s529ms | To 100/100 | Loss: 0.648 | Acc: 80.750% (8075/10000)0)



・・・・
・・・・


Epoch: 80
 [===============================>]  Step: 967ms | Tot: 782/782  Loss: 0.254 | Acc: 90.834% (45417/50000)))2)
 [===============================>]  Step: 1s663ms | To 100/100 | Loss: 0.648 | Acc: 80.720% (8072/10000)))
Epoch: 81
 [===============================>]  Step: 815ms | Tot: 782/782  Loss: 0.249 | Acc: 91.138% (45569/50000)0)4)
 [===============================>]  Step: 1s426ms | To 100/100 | Loss: 0.648 | Acc: 80.720% (8072/10000)))
Epoch: 82
 [===============================>]  Step: 916ms | Tot: 782/782  Loss: 0.251 | Acc: 91.200% (45600/50000)4)4)
 [===============================>]  Step: 1s659ms | To 100/100 | Loss: 0.648 | Acc: 80.730% (8073/10000)))

試行錯誤

Transformerに期待し？、Verticalのaugmentationを有効にしてみたが、、、何もいいことは起きなかった！！

まとめ

上記のとおり、**なんでしょう？？？？**終わりです。
コメントなどあれば、お願いします。

関連

ViT(Vision Transformer)を動かしてみた、4つほど、うっ、Cifar10で80%超えた、ぼちぼちくるかー。【番外編】