Building complicated deep convolutional networks using PyTorch
SAURABH KUMAR 14BCE1033 (~saurabh_kumar21) |
With the plethora of research that has been done on convolutional neural networks, building a new network for your own problem or dataset is as simple as copying the design philosophies of powerful networks that have proven their mettle against the mammoth of ImageNet data. However replicating their work using modern deep learning frameworks may not be so obvious. The fine details are usually missed out or are not explained sufficiently for reproduction of the work.
Objective of my talk is
- Building 3 of the most popular ImageNet networks - AlexNet, GoogleLeNet, ResNet, from scratch using PyTorch.
- Testing these networks for design flaws.
- Provide some basic design principle to make it easier to build these networks.
- Examples of how to feed data to the network and perform end to end training using PyTorch.
- Visualizing the training.
By the end of this talk, you will be able to
- Replicate these design philosophies for your own networks
- Efficiently test and train your networks
A basic understanding of deep learning (Backpropagation) would be helpful but not needed. You need no prior knowledge about PyTorch as I will be going into depth as I explain and build the networks. An interest towards deep learning and familiarity with other frameworks would help enjoying the talk.
Presentation (WIP): https://docs.google.com/presentation/d/1SOpWz9HzUaM9rxpPtu4WRkgC5a1DTZsg3zFZczlOy10/edit?usp=sharing
I am passionate about deep learning and its amazing advancements in the field of computer vision and NLP. I spend my days reading blogs and papers and trying to implement them myself. I am an incoming software developer at PayPal, currently in my senior year pursuing CS from VIT, Chennai Campus. Apart from that I love playing games and reading books.