The data was split into 2 folders of train and test and we were given one csv files with names from the test folder.
The files in the train folder was named as 'cat.x' and 'dog.x' with a total of 25000 files
This code was compiled in Kaggle.
We first create a dataframe from the train and test directories and label them according to if it was named cat or dog in file.
We then split the data set into train and valid dataset.
We then create two ImageDataGenerators based on if it is a train or valid dataset.
Also we use 2 types of Data generators on the basis of whether it is a Vanilla model or a ResNet model.
If it was a ResNet model then we use a preprocess_input
Now we create a CNN model using Sequential API with 2 blocks of two conv2d layers followed by 2 blocks of three conv2d layers.
This is followed by 4 Dense layers with a BatchNormalization and Drop out Layer in between.
We fit the model and get a validation_set accuracy of 92%.
We then create a Resnet50 model with non trainable layers and then we add 2 Dense layers as head of the model. We get a validation accuracy of 99%
Now for the test dataset we couldn't use DataGenerator for this because of some reason which I probably think is due to the architecture of the test folder and according to flow_from_directory we need to have a structure of Folder->SubFolder->images.
Instead we create a function to get the filenames from the test directory and then use load_img to create an array and then predict it using the model and kept a threshold of 0.5 where if it is greater than 0.5 then predict = 'dog' else 'cat'
We then create a 4x3 subplot with text which gives predicted result of both the Vanilla Model and also the Resnet model.