Convolution Neural Networks (CNN)

Fundamentals of DL course Assignment - 2

Install

This project requires Python and the following Python libraries installed:

• NumPy
• matplotlib
• Pytorch
• TensorFlow
• OpenCV
• wandb
• wget
• mapextrackt

Code

• Convolution Neural network Class (MyNN)

The convolution neural network is built using Pytorch with conditions on batch normalization. The function fit is the contains the neural network pipeline (Forward_prog, compute_loss, back_prop, update_parameters) is used for computing the training and the validation error along with validation accuracy.

The train function consists of data-preprocessing, initializing the model and it's parameters.

The following line of code is an example to define a model using the MyNN class:

model = MyNN(drop_out = 0.2, batch_norm = 'Yes', filter_n = 128,  filter_org = 'double_down', hidden_out =196)

After defining the model, the training of the model can be done using the following command:

fit(model,train_loader,val_loader,optimizer,epoch,criterion)

• Wandb configuration

Wandb is a tool for tuning the hyper-parameters of a model. The wandb sweep requires to define a sweep configuaration with hyper-parameters in a dictionary type. The following code snippet is an example of defining the wandb sweep configuration:

sweep_config = {
    'method': 'bayes', #grid, random
    'metric': {
      'name': 'accuracy',
      'goal': 'maximize'   
    },
    'parameters': {
        'drop_out': {
            'values': [0.2, 0.3]
        },
        'batch_norm': {
            'values': ['Yes', 'No']
        },
        'filter_n': {
            'values': [64, 128]
        
        },
        'batch_size': {
            'values': [16, 32]
        },
        'filter_org': {
            'values': ['same', 'double_up', 'double_down']
        },
        'epoch': {
            'values': [5,10]
        },
        'data_aug': {
            'values': ['Yes', 'No']
        },
        'optimizer': {
            'values': ['SGD','ADAM'] 
        },
        'lr': {
            'values': [0.1,0.01] 
        },
        'hidden_out': {
            'values': [128,196] 
        },
    }
}

sweep_id = wandb.sweep(sweep_config, entity="paddy3696", project="cnn_inat")

• Train sweep function

The function train is the main function called by the wandb sweep. This function contains the wandb initialization and data pre-processing.

• Testing

The function model_test finds the accuracy of the model with test data.

• Filter visualization

The filter visualization is performed using mapextrackt library. feature extraction and visualization functions is used from this library.

• Pre - trained models

The pre-trained models are trained used from the pytorch library. These models are initialized as a hyper-parameter. The list of pre-trained models is mentioned below:

['resnet', 'alexnet', 'vgg', 'squeezenet', 'densenet', 'inception']

• Guided Backpropogation

The Guided backpropogation is performed using Tensorflow Keras because extracting features of each layer was not direct. Another significant reason for using Keras instead of Pytorch is to understand the difference between two libraries. The same CNN model is built completely from the scratch and fifth convolution layer is visualized using the guided propogat

Run

In a terminal or command window, navigate to the top-level project directory CNN_Pytorch/ (that contains this README) and run one of the following commands:

ipython notebook Inat_cnn_train.ipynb

or

jupyter notebook Inat_cnn_train.ipynb

The code for evaluating the perfomance of the custom CNN model with iNaturalist dataset is seperately uploaded and it can be run using the following command:

jupyter notebook Inat_cnn_test.ipynb

The code for guided backpropagation of the custom CNN models with iNaturalist dataset is seperately uploaded and it can be run using the following command:

jupyter notebook guided_backprop.ipynb

The code for evaluating the perfomance of the pretrained CNN models with iNaturalist dataset is seperately uploaded and it can be run using the following command:

jupyter notebook inat_cnn_pretrained.ipynb

Data

The iNaturalist datasetis downloaded directly from the downloadable link using the following the "wget" command:

wget.download('https://storage.googleapis.com/wandb_datasets/nature_12K.zip')

Data Preprocessing

• The iNaturalist dataset contains 9999 training image data and 2000 testing.
• The training data is very split with ratio of 90:10 for training and validation. This is done to avoid overfitting.
• All the test, train and validation data are imported using the data loader function in torch library.
• The transfromers function are used for resizing, cropping, normalizng the images and then convert it to tensors.

Report link

assignment2_report

Reference

• Udacity Deep Learning course
• Sentdex Pytorch tutorials
• Python Engineer Pytorch tutorials
• https://towardsdatascience.com/visualizing-convolution-neural-networks-using-pytorch-3dfa8443e74e
• mapextrackt
• https://stackoverflow.com/questions/55924331/how-to-apply-guided-backprop-in-tensorflow-2-0
• https://pytorch.org/tutorials/beginner/finetuning_torchvision_models_tutorial.html
• YoloV3
• https://www.pyimagesearch.com/2020/06/01/opencv-social-distancing-detector/