Kaggle Digit Recognizer別人分享的tricks
自己在利用CNN做了kaggle上的入門題——Digit Recognizer,簡單的利用了CNN的網絡架構就實現了0.98+的準確率，但還是不夠啊，所以看了上面大佬的分享，找到一篇不錯的分享，這裏記錄下其中我當時沒用過的tricks。加油，AI learning！
- 數據增強（可以增加訓練的數據量，The idea is to alter the training data with small transformations to reproduce the variations occuring when someone is writing a digit.
For example, the number is not centered The scale is not the same (some who write with big/small numbers) The image is rotated…
Approaches that alter the training data in ways that change the array representation while keeping the label the same are known as data augmentation techniques. Some popular augmentations people use are grayscales, horizontal flips, vertical flips, random crops, color jitters, translations, rotations, and much more.
- 學習率的動態調整，一開始可以高一點，使模型學習速率快點，然後到了比較高的準確率時，把學習率降低，可以繼續學習下去，而且可以儘量避免最優值是在局部極值點。
- Confusion Matrix Confusion matrix can be very helpfull to see your model drawbacks.
主要是這三點，然後對於CNN 的網絡層次結構，這篇Notebook也介紹的比較好，記錄以下可以回去回顧。
鏈接是：Introduction to CNN Keras - Acc 0.997 (top 8%)
我只在自己運用的cnn中添加了數據增強的部分，準確率達到了0.99，效果還是挺好的。