[WEEK 7-Country Classification Using House Photos]
Team Members: Meltem Tokgöz , Enes Furkan Çiğdem , Asma Aiouez
This week, we used a different CNN architecture, AlexNet method, which happens to work well on image classification problems. The purpose is that we wanted to compare our results to the ones obtained from last week when using ResNet18 architecture.
ResNet18 architecture contains eighteen layers whereas AlexNet one has eight layers; the first five are convolutional layers, some of them followed by max-pooling layers, and the last three are fully connected layers. Since the latter has fewer number of layers, we expected it to perform not quite as good as the former architecture. The following is a summary of the result we obtained when applying AlexNet model:
As we said before we tested alexnet and results are like below.
The accuracy graph of alexnet architecture with 30 epochs and 0.01 learning rate.While the epochs are increasing to 10,the accuracies also increasing. However after about 10th epoch there is no big change.
The matrix in the above is confusion matrix of Alexnet architecture. When we look at the the diagonal line of the matrix,we can say we are quiet succesfull. Particularly, we got the best results in Australia with 100% and for all dataset we got 81% accuracy rate.
However, the france predictions are bad since there is no enough training data and there are mislabeled noisy data in france dataset.
Some prediction examples :
In addition to that, this week we have created a presentation video of our project and shared it on YouTube.
The only thing left is to finalize our project report and work on the presentation of our project that we will present this upcoming week.
When we decided to do this project, we did not know that collecting the dataset is going to be this hard. We think that the hardest part that we have faced is collecting the dataset because there isn’t enough data for some countries and for some countries contains so much noise in it.
We tried to get different continental countries and different architectures.
For example hanok type houses in South Korea or Stone houses of Ireland.
Trying and experiencing with CNN architectures is also hard; for that reason, we decided to use transfer learning and tested some pretrained models like the ones below. We also used the versions of these models which are not pretrained.
Resnet18 : We used this pretrained model in our project.This is bad for time because training with 30 epoch takes 1 hour even in GPU. Although this model is time consuming, it gave us best accuracy (89%)
Alexnet: This is better compared to resnet18 model in terms of the time. On the other hand, it gave us lower accuracies than resnet18 accuracies.
