Readme.md

CNN1o/Readme.md

In the CNN1l, best conditions (06) are as follows;

• 1st Cond2D ; filter is 7x7
• Use learning rate reducing, starting lr=0.001, factor=0.47
• Use dropout (0.4) after each Cond2D
• Channels are doubled in each Cond2D

Here, based on the condition of CNN1l/06, try "test time augumentation".

Training conditions and Result of score

Common conditions

• Batch size ; 32
• Dropout after Cond2D ; Yes (0.4)
• BatchNormalization after Cond2D ; No

Training conditions

No	Conditions	Min of val_loss	Max of val_accuracy	Score
Ref	CNN1l/06	0.02138 (epochs=65)	0.99512 (epochs=68)	0.99507 (epochs=62)
00	TTA (epochs=5)	0.02195 (epochs=64)	0.99512 (epochs=66)	0.99467 (TTA)
01	TTA (epochs=5)	0.02154 (epochs=60)	0.99512 (epochs=58)	0.99532 (epochs=60)
02	filters=512	0.02273 (epochs=70)	0.99476 (epochs=57)	0.99478 (epochs=70)

Detail

00 ; Test Time Augumentation (epochs=5), Saved as Ver.15 on Kaggle

• Test time augumentation is used. Epochs = 5.
• Accuracy score of train data: 0.9972380952380953

01 ; Test Time Augumentation (epochs=5), filters=256

• filters=256
• Accuracy score of train data: 0.9973095238095238

02 ; filters=512

• TTA is not used
• filters=512
• Accuracy score of train data: 0.9973095238095238

Results

• 00
  • digit-recognizer_CNN1o_tta.csv ; 0.99467
  • epochs=64 ; 0.99407
  • epochs=59 ; 0.99425
  • not better than CNN1l/06
• 01
  • digit-recognizer_CNN1o_tta.csv ; 0.99500
  • epochs=60 ; 0.99532 (257 / 2254 = 0.1140)
• 02
  • epochs=70 ; 0.99478
  • epochs=69 ; 0.99460

Graphs

Reference (CNN1l/06)

00 ; Test Time Augumentation (epochs=5)

01 ; Test Time Augumentation (epochs=5), filters=256

• Prediction without TTA is better than Reference.

02 ; filters=512

• not so good