LeNet is the first CNN network proposed by LeCun.
I used different CNN training tricks to show you how to train your model efficiently.
LeNet_keras.py is the baseline of LeNet,
LeNet_dp_keras.py used the Data Prepossessing [DP],
LeNet_dp_da_keras.py used both DP and the Data Augmentation[DA],
LeNet_dp_da_wd_keras.py used DP, DA and Weight Decay [WD]
| network | GPU | DP | DA | WD | training time | accuracy(%) |
|---|---|---|---|---|---|---|
| LeNet_keras | GTX1080TI | - | - | - | 5 min | 58.48 |
| LeNet_dp_keras | GTX1080TI | √ | - | - | 5 min | 60.41 |
| LeNet_dp_da_keras | GTX1080TI | √ | √ | - | 26 min | 75.06 |
| LeNet_dp_da_wd_keras | GTX1080TI | √ | √ | √ | 26 min | 76.23 |
PS: try to change learning rate setting and retrain the network to see different results.