Readme.md

CNN1n/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 various method.

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	factor = 0.631	0.02099 (epochs=61)	0.99595 (epochs=65)	0.99475 (epochs=61)
01	00 + doubled channels	0.02247 (epochs=54)	0.99536 (epochs=54)	0.99482 (epochs=54)

Detail

00 ; learning rate reducing factor ; 0.47 -> 0.631

Factor of learning rate reducing is changed from 0.47 to 0.631

01 ; 00 + doubled channels of Conv2D (as in CNN1l/07)

Base is CNN1n/00 above, and channels of Cond2D are doubled (256 - 1024 - 1024).

Results

• 00

  • epochs=61 ; 0.99475
  • epochs=60 ; 0.99450

• 01

  • epochs=54 ; 0.99482
  • epochs=53 ; 0.99439

Graphs

Reference (CNN1l/06)

00 ; learning rate reducing factor ; 0.47 -> 0.631

• Slightly improved comparing to CNN1l/06.

Reference (CNN1l/07)

01 ; 00 + doubled channels of Conv2D (as in CNN1l/07)

• Comparing to CNN1l/07
  • val_accuracy is improved, others seems the same
• Comparing to 00
  • accuracy and loss are improved, but val_accuracy and val_loss seems the same