rvk007/deepnet

A python library for computer vision applications

Deepnet is an open-source library that can be used for solving problems of Computer vision in Deep Learning.

NOTE: This documentation applies to the MASTER version of DeepNet only.

Install Dependencies

Install the required packages
pip install -r requirements.txt

Features

DeepNet currently supports the following features:

Models

Models	Description
ResNet	ResNet-18
ResModNet	A modified version of ResNet-18
CustomNet	A modified version of ResNet-18
MaskNet3	A model to predict the Segmentation mask of the given image.
DepthMaskNet8	A model to predict the Monocular Depth Maps of the given image.

Training and Validation

Functionality	Description
Train	Training and Validation of the model
Model	Handles all the function for training a model
Dataset	Contains classes to handle data for training the model

Metrics

• Mean Absolute Error
• Root Mean Squared Error
• Mean Absolute Relative Error
• Intersection Over Union Error
• Root Mean Square Error

Losses

Loss	Description
Dice	ResNet-18
SSIM	A modified version of ResNet-18
MSE	Mean squared error (squared L2 norm) between each element in the input and target
BCE	Binary Cross Entropy between the target and the output
BCEWithLogitsLoss	Combination of Sigmoid layer and the BCE in one single class
RMSE	Root mean squared error (squared L2 norm) between each element in the input and target

Weighted Combination of loss functions

• BCE-RMSE
• SSIM-RMSE
• BCE-SSIM
• RMSE-SSIM
• SSIM-DICE
• RMSE-DICE
• BCE-DICE
• RMSE-BCE-DICE

Scheduler

• StepLR
• ReduceLROnPlateau
• OneCycleLR

Data Augmentation

• Resize
• Padding
• Random Crop
• Horizontal Flip
• Vertical Flip
• Gaussian Blur
• Random Rotation
• CutOut

Utilities

Utility	Description
GRADCAM	Calculates GradCAM(Gradient-weighted Class Activation Map) saliency map
GradCAMpp	Calculate GradCAM++ salinecy map using heatmap and image
LRFinder	Range test to calculate optimal Learning Rate
Checkpoint	Loading and saving checkpoints
ProgressBar	Display Progress bar
Tensorboard	Creates Tensorboard visualization
Summary	Display model summary
Plot	Plot the graph of a metric, prediction image and class accuracy

Dependencies

DeepNet has the following third-party dependencies

• numpy
• torch
• torchvision
• torchsummary
• tqdm
• matplotlib
• albumentations
• opencv-python

For a demo on how to use these modules, refer to the notebooks present in the examples directory.

Contact/Getting Help

If you need any help or want to report a bug, raise an issue in the repo.

deepnet's Issues

Correct, Incorrect Predictionsand Class Accuracy

Add file to display sample images with correct and incorrect predictions.
Add file to display test accuracy per class.

Data Augmentation with Albumentation

Add data augmentations using albumentation package.
Following augmentations will be added:

• Horizontal Flip
• Vertical Flip
• Rotate
• Cutout
• Gaussian Blur

Update metrics

Accuracy doesn't work for depth estimation, so incorporate evaluation metrics.

Get the images of same batch

To get an image gif, fetch the output of same batch.

Add Examples

Add a Example folder that contains colab files that can referred to understand the package.

Loss incorporation and learning rate change

Update for getting learning rate changes and Update for loss functions.

GradCAM

Add GradCAM functionality to the project

• GradCAM
• GradCAM++

Documention and Examples