aeaswar81/cs6910_deeplearning_assignment_2

This project uses Object oriented programming for modularity. There are three parts namely

1. Training CNN from Scratch(trainingFromScratch)
2. Transfer Learning(transferLearning)
3. Application using YoloV3(yolov3App)

Part A - Training CNN from Scratch

For the purpose of part A, we have created a class CNNClassifier using tensorflow and keras. Its constructor can take the following parameters:

1. input_shape - Dimension of the input
2. filters - List of number of filters in every layer
3. kernel_size - List of size of kernel in every layer
4. strides - Stride size to be used in every layer. By default stride size is (1,1)
5. conv_activation - Activation function to be used in convolution layers
6. pool_size - Pooling size. By default, pool size is (2,2)
7. dense_layer_size - List of size of dense layers
8. dense_activation - Activation function to be used at the dense layers
9. output_layer_size - Size of the output layer. This should be set equal to the number of classes.
10. output_activation - Activation function to be used at the output layer. By default, it is set as 'softmax'
11. batch_normalization - Whether batch normalization has to be used or not. Boolean value. If true, batch normalization is done at every layer.
12. dropout - The value of dropout to be used at the dense layers.

This class has the following methods:

fit(optimizer, loss, metrics, trainX, trainY, epochs, validation_data)

To train the CNNClassifier.

optimizer - optimizer to be used while training loss - loss function to be used for training metrics - metrics to be monitored while training trainX - training data trainY - training labels epochs - number of epochs until which the model will be trained validation_data - validation data to be used

fit_generator(optimizer, loss, metrics, train_generator, epochs, validation_generator, callbacks)

This is same as fit except that it uses keras data generator as input instead of arrays. It also has a callback parameter.

evaluate(testX, testY)

To evaluate the model on test data

evaluate_generator(test_generator)

To evaluate the model using a test data generator

summary()

To print the summary of the model

guidedbackprop(image)

This method generates plot that shows which part of the image excites various neurons in the fifth convolution layer.

This part also the following python scripts:

1. runCNN.py - loads cifar10 dataset, builds a simple CNNClassifier, trains the model, plots accuracy vs epoch, and evaluates the model on test data
2. inatClassification.py - loads the inaturalist dataset, builds a CNNClassifier, trains the model, plot the guided backpropogation, and evaluates the model on test data
3. Sweep.py - Runs wandb sweeps for hyperparameter tuning
4. bestCNNModel.py - Trains and evaluates the best CNN configuration obtained from wandb sweeps
5. preprocess_crop.py - To add support for center cropping of images as a preprocessing step
6. visualizeFilters.py - To visualize the filters and feature maps of a layer
7. plotTestPredictions.py - To plot samples of test data and the predictions made by a model

Part B - Transfer Learning

The python notebook has code for loading various pretrained models and fine tuning them. This also can run sweeps for the same.

Part C - Application using YoloV3

This part has a separate readme file.