Hyper-Parameter Grid Search
Best Hyper-Parameter Combination
Hyper-Parameter Grid Search is used to find the best combination of hyper-parameters that help convnet to achieve highest accuracy. This approach is preferred as it’s difficult to configure a convnet accurately as it depends on a lot of hyper-parameters. Few examples of convnet hyper-parameters are, size of kernel, dropout, number of convolution layers, batch size, type of activation function, stride of filters, and so on.
Different combination of convnet hyper-parameters yield different accuracy of inputs. In our experiment, we have assumed dropout and kernel size to be our hyper-parameters that we are going to vary. It’s found that kernel size of 5 and dropout of 0.1 yields the best convnet model. Architecture design of the this best convnet model is as below:
We infer from our experiment that hyper-parameters play a pivotal role in choosing a model that enhance its accuracy. In our case, increase in dropout rate and reduction in filter size has been found to reduce accuracy of model. These results can’t be argued to hold good for other convnet models designed for different data-set.
Accuracy
- Training accuracy is always greater than or equal to test accuracy.
Source Code
