“Guided Back Propagation for Gradient Visualization of Pytorch Video Models”
About
VideogradVis is an advanced package that encompasses cutting-edge methods for Explainable AI (XAI) in the realm of computer vision. This versatile tool is designed to facilitate the diagnosis of model predictions, serving a dual purpose in both production environments and during the model development phase. Beyond its practical applications, the project also serves as a comprehensive benchmark, offering a robust foundation for researchers to explore and evaluate novel explainability methods.
-
Compatibility Assurance: Rigorously tested on a multitude of Common Convolutional Neural Networks (CNNs).
-
Versatile Applications: Unleash the full potential of VideogradVis across a spectrum of advanced use cases. From Classification and Object Detection to Semantic Segmentation, Embedding-similarity, and beyond, this suite empowers your projects with unparalleled flexibility and performance.
Installation
python setup.py install
Usage
import torch
from VideogradVis.Vis import GradVis
odel = torch.hub.load('facebookresearch/pytorchvideo', 'x3d_m', pretrained=True)
model.blocks[5].proj = nn.Linear(in_features = 2048, out_features = args.classes, bias = True)
print(f"Loading weights: {args.checkpoint}.")
model.load_state_dict(torch.load(args.checkpoint))
vis = GradVis(model)
model_input = torch.rand((1,3,120,250,250)).requires_grad_(True) #(B,C,T,H,W)
vis.compute_grad(input_tensor=input_model, path='/home/osama/pytorch-video/output/')
Model Predictions Visualization
This document provides visualizations explaining the model's predictions on whether a headlight is ON or OFF. Each figure offers insights into the gradients and features that influenced the model's decision-making process.
View More research
vfusedSeg3d
An innovative multimodal fusion system created that combines camera and LiDAR data to significantly enhance the accuracy of 3D perception.
Learn More
VideoGrad Vis
An advanced package that encompasses cutting-edge methods for Explainable AI (XAI) in the realm of computer vision.
Learn More
Take the next Step
For product demonstration or collaboration, please contact us here.
For all other questions, please contact us here.
 brand.svg)
