This is improved version of my CarND-Traffic-Sign-Classifier-Project. The dataset is the German Traffic Signs as well.
I remodeled the architecture of convolutional neural networks from scratch using simple inception module, which is used in Google's ConvNets.
Based on GoogLeNet, I made the new neural networks architecture as follows:
The new architecture using inception module
The hyperparameters I used are:
| Name | Value | Description |
|---|---|---|
mu |
0 | For initilazing Wights with normal distribution |
sigma |
0.01 | For initilazing Wights with normal distribution |
learning_rate |
0.0005 | For training neural networks |
epochs |
350 | Number of training times |
BATCH_SIZE |
256 | Number of images feeding to the model at one time |
After training, the validation accuracy of new architecture is:
The validation accuracy of new architecture
That shows the new convolutional neural networks using inception module improved the max validation accuracy from 97.3% to 98.3%.