[WEEK 6-Country Classification Using House Photos]
Team Members: Meltem Tokgöz , Enes Furkan Çiğdem , Asma Aiouez
Last week, we have trained our model using a single layer neural network. This week however, as we mentioned in our last blog post, we have applied a multi-layer neural network. Changing the learning rate parameter, we have obtained several results that we’ll show next.
As a CNN model, using the machine learning library PyTorch that we have installed we have decided to use the pre-trained model Residual Network ( ResNet18 ) as a start for the training. The model includes five convolutional stages as shown the table below.
Changing the value of the learning rate with epoch values less than 30, and combining all our dataset which now has 15 classes, we obtained the following two confusion matrices. Moreover, we drew their respective graphs representing the training and validation accuracies where the blue curve is for the training set whereas the red one belongs to the validation set.
From the confusion matrices, we notice that the accuracy results differ from a country to another where France registers the lowest accuracy rate which increased a little bit when increasing the learning rate parameter. Turkey in both tests obtained the highest accuracy rate. We see from the following table that the overall performance of the model is quiet good and the results are summarized below.
In the coming week, we’ll use a different CNN architecture and compare the results we are going to obtain with this week’s outcomes hence we’ll be able to better see and understand the effect of deep neural networks.
