Week 3 — Identifying Book Genre and Recommendation
This week we will prepare our cover image data and build our first model using Tensorflow.
Data Preprocessing
Our dataset includes 50k images with 224x224 size. Using all of the data would take up a lot of space and as we have limited computational power and sources, we decided to decrease image sizes and use a smaller amount of examples.
We first create train and test sets including 8192, 1024 examples. Then after resizing images into 64x64 we transfer the images into a NumPy array.
In the last step of the data preprocessing, we create 2d arrays for classes (y) using one-hot encoding.
Classification
We have examples from 30 different genres and we decided to implement a Convolutional Neurol Network model. We used CNN layers with 3x3 and MaxPooling layers with a 2x2 filter. The structure has nearly 175000 parameters.
After the training, we got a really low accuracy for both train and test sets. We are thinking that having 30 classes, might be the reason for the inadequate model. Next week we are going to decrease the class number by combining similar genres.