Help us understand the problem. What is going on with this article?

[Kaggle]大腸癌を分類[fine tuning]

はじめに

つくりながら学ぶ!PyTorchによる発展ディープラーニングという本の1-5のファインチューニングで細胞の分類をしてみました。(筆者GitHubで全てのコードが見られます)
データはKaggleのColorectal Histology MNISTを使いました。

開発環境

  • Google Colaboratory

やったこと

Kather_texture_2016_image_tiles_5000フォルダ内に8種類に画像が分類されているのでそれを見分けます。

実行結果

使用デバイス cuda:0
  0%|          | 0/47 [00:00<?, ?it/s]Epoch 1/100
-------------
100%|██████████| 47/47 [07:26<00:00,  9.49s/it]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 2.1278 Acc: 0.1060
Epoch 2/100
-------------
100%|██████████| 110/110 [17:17<00:00,  9.43s/it]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.8146 Acc: 0.7206
100%|██████████| 47/47 [00:12<00:00,  3.76it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.4196 Acc: 0.8547
Epoch 3/100
-------------
100%|██████████| 110/110 [01:04<00:00,  1.71it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.3953 Acc: 0.8597
100%|██████████| 47/47 [00:12<00:00,  3.79it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.3262 Acc: 0.8853
Epoch 4/100
-------------
100%|██████████| 110/110 [01:04<00:00,  1.71it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.3165 Acc: 0.8894
100%|██████████| 47/47 [00:12<00:00,  3.84it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.2910 Acc: 0.8973
Epoch 5/100
-------------
100%|██████████| 110/110 [01:04<00:00,  1.71it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.2828 Acc: 0.8971
100%|██████████| 47/47 [00:12<00:00,  3.81it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.2194 Acc: 0.9247
Epoch 6/100
-------------
100%|██████████| 110/110 [01:04<00:00,  1.71it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.2596 Acc: 0.9097
100%|██████████| 47/47 [00:12<00:00,  3.83it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.2573 Acc: 0.9087
Epoch 7/100
-------------
100%|██████████| 110/110 [01:04<00:00,  1.71it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.2405 Acc: 0.9171
100%|██████████| 47/47 [00:12<00:00,  3.84it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.2294 Acc: 0.9240
Epoch 8/100
-------------
100%|██████████| 110/110 [01:04<00:00,  1.71it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.2199 Acc: 0.9223
100%|██████████| 47/47 [00:12<00:00,  3.88it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.2053 Acc: 0.9267
Epoch 9/100
-------------
100%|██████████| 110/110 [01:04<00:00,  1.71it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.1993 Acc: 0.9309
100%|██████████| 47/47 [00:12<00:00,  3.85it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.2009 Acc: 0.9293
Epoch 10/100
-------------
100%|██████████| 110/110 [01:03<00:00,  1.72it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.2097 Acc: 0.9280
100%|██████████| 47/47 [00:12<00:00,  3.85it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.1770 Acc: 0.9400
Epoch 11/100
-------------
100%|██████████| 110/110 [01:03<00:00,  1.72it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.1860 Acc: 0.9363
100%|██████████| 47/47 [00:12<00:00,  3.90it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.1753 Acc: 0.9400
Epoch 12/100
-------------
100%|██████████| 110/110 [01:03<00:00,  1.74it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.1751 Acc: 0.9429
100%|██████████| 47/47 [00:11<00:00,  3.95it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.2092 Acc: 0.9260
Epoch 13/100
-------------
100%|██████████| 110/110 [01:03<00:00,  1.73it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.1595 Acc: 0.9466
100%|██████████| 47/47 [00:11<00:00,  3.92it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.2082 Acc: 0.9307
Epoch 14/100
-------------
100%|██████████| 110/110 [01:03<00:00,  1.73it/s]
  0%|          | 0/47 [00:00<?, ?it/s]train Loss: 0.1653 Acc: 0.9431
100%|██████████| 47/47 [00:11<00:00,  3.94it/s]
  0%|          | 0/110 [00:00<?, ?it/s]val Loss: 0.1639 Acc: 0.9500

大体この辺でval lossが定常状態に達したので載せるのはここまでにしておきます。
精度95%はすごくないですか。

考察・終わりに

病理組織画像も見分けられるというのは驚き桃の木山椒の木。

tamahassam
ディープラーニングを勉強中の医大生です。
Why not register and get more from Qiita?
  1. We will deliver articles that match you
    By following users and tags, you can catch up information on technical fields that you are interested in as a whole
  2. you can read useful information later efficiently
    By "stocking" the articles you like, you can search right away