Featured Report

This report is a saved snapshot of Roberto's research. He has published this example so you can see how to use W&B to visualize training and keep track of your work. Feel free to add a visualization, click on graphs and data, and play with features. Your edits will not overwrite his work.

Project Description

Roberto developed a deep learning model that can identify the genre of a YouTube video. Using PyTorch & Keras models to aggregate spatial strings (pixels) and sequential strings (audio). Later, he concatenates the model onto a fully connected network to output the video label genre: Games, Art & Entertainment, etc. He wrote a more thorough description of his work in his thesis paper, and the code for this project is available on GitHub.

Section 1

Featured Report

This report is a saved snapshot of Roberto's research. He's published this example so you can see how to use W&B to visualize training and keep track of your work. Feel free to add a visualization, click on graphs and data, and play with features. Your edits won't overwrite his work.

Project Description

Roberto developed a deep learning model that can identify the genre of a YouTube video. Using PyTorch & Keras models to aggregate spatial strings (pixels) and sequential strings (audio). Later, he concatenates the model onto a fully connected network to output the video label genre: Games, Art & Entertainment, etc. He wrote a more thorough description of his work in his thesis paper and the code for this project is available on GitHub.

Comparing models and frameworks

I coded eight deep learning models: 4 in Keras and the same models in PyTorch. I chose to compare not only models but frameworks as well based on research I did on algorithms crashing in production. Tensorflow (Keras) has the largest user base and most traction currently. I want to compare it with PyTorch which is an imperative framework that performs computation as you type it. Tensorflow (Keras) uses symbolic programming: only computing your code at the end of each graph session. Tensorflow is evolving to become more “PyTorch” like with eager execution, however that’s still in alpha and didn’t exist at the time time my project started. The value of comparing frameworks was also recognized by google engineers as they were working along a similar path in parallel to my work. More information about the difference between each framework and their performance is written in my paper. My experiments reveal surprising results.

Read more details →

Comparing Models and Frameworks

I coded eight deep learning models: 4 in Keras and the same models in PyTorch. I chose to compare not only models but frameworks as well based on research I did on algorithms crashing in production. Tensorflow (Keras) has the most extensive user base and most traction currently. I want to compare it with PyTorch, an imperative framework that performs computation as you type it. Tensorflow (Keras) uses symbolic programming: only computing your code at the end of each graph session. Tensorflow is evolving to become more “PyTorch” like with eager execution; however, that is still in alpha and did not exist when my project started. Google engineers also recognized the value of comparing frameworks as they worked along a similar path parallel to my work. More information about the difference between each framework and its performance is written in my paper. My experiments reveal surprising results.

Read more details →