Readme.md

CNN1q/Readme.md

Summary

• VGG16-like deep models are tried.

Training conditions and Result of score

No	Conditions	Min of val_loss	Max of val_accuracy	Score
Ref	filters=256	0.02154 (epochs=60)	0.99512 (epochs=58)	0.99532 (epochs=60)
Ref	CNN1p/00			0.99553 (soft)
00	n_layers=2, filters=128			0.99435 (epochs=30)
01	n_layers=2, filters=128	0.02382 (epochs=46)	0.99488 (epochs=38)	0.99285 (epochs=46)

Detail

00 ; n_layers=2, filters=128

01 ; n_laysers=2, filters=128 (Same as 00)

• Same condition as 00, but watch loss to save data.

02 ; condition is same as 01 + ensamble training (epochs=5), saved as Ver.18 on Kaggle

• Same condition as 01, and ensamble training is used.

Results

• 00
  • epochs=30 ; 0.99435
• 01
  • epochs=46 ; 0.99285
• 02
  • soft ensamble ; 0.99510

graphs

Ref ; filters=256

00 ; n_laysers=2, filters=128

• val_loss is not stable. So it seems better to check loss instead of val_loss.

01 ; n_laysers=2, filters=128 (Same as 00)

• When epochs > 40, seems over-fit.