In short, this is a program that is shown a group of images of declared tags, then it can draw a prediction for what tag(of the ones previously introduced) a brand new image belongs in.
- Preprocessing, training, and prediction functions run without crashing.
- Upon running the prediction function for the model, the model favors one image tag/class over the other three available tags.
- Symptoms point to an overfitted model:
- Diverging pattern begins to take shape upon finishing the 10th epoch for training loss and validation loss.
- Diverging pattern begins to take shape upon finishing the 10th epoch for training accuracy and validation accuracy. (Anticipated Solution) Apply more data augmentation techniques to the data set, tweak the number of epochs, experiment with different batch sizes, and make sure any corrupt image data is completley taken care of.
The VGG-16 Convolutional Neural Network(VGG-16 CNN) architecture looks at image features as it moves through the layers of the network. Using square filter/kernel structures, it selects edges and other simple shapes on the image. Moving further into the model, more abstract features are dealt with such as a mouth, an ear, a furry tail, a wheel, or a beak. Finally, the final layers are the ones that look at the gathered image scraps to accurately identify the appropriate image class.
model.py -> The function here takes an input image of size "56pi x 56pi x 3channels". The image is dismembered into filters that are then passed into a series of convolutional layers, pooling layers, and fully connected layers. The final layer genereates a probability distribution value of the four available image labels. The highest value of these four tells which class the input image most likely belongs to. From beginning to end, the model uses the ReLU activation function which is how the model extracts features from pixels. For backpropagation, gradient descent is used which just helps minimize error between guessed image labels and the true labels of an image data set.
imgPrep.py -> The function here preprocesses images before they are fed to the model. Image labels are reliant on folder hierarchy, so all dog images are inside the folder "Dog", all car images are inside the folder "Car", etc. All these folders would them be inside one root folder like(in this case) "Images". A traversal is made through all of these folders and any corrupt images are deleted. Images are all resized to "56pi x 56pi", and image batches are made for both training and validation. Batches are marked for data augmentation, and the batches have their data type changed to match the desired input of the model.