Project Introduction

As part of my AI coursework (CS231n) at Stanford, my team attempted to develop an encoder-decoder generative AI model to produce a customizable CAD file from a 2D isometric sketch. The goal was to expedite CAD workflows that can be resource and time intensive.

Sketch 2 CAD

Generative AI CAD Step Generation from Isometric Sketch Inputs

Category: Generative AI
Client: Stanford CS 231N
Project date: 03/2023 - 06/2023
Project Report: View

Contributions & Outcomes

I contributed to data collection from the DeepCAD subset of the ABC dataset. I contributed to model development of our four encoder-decoder models using PyTorch. I was responsible for training models on an AWS EC2 instance. I was also responsible for evaluating models with saliency plots and 3D renders of predictive CAD. The resulting optimal model was a CNN encoder with a pretrained transformer decoder developed as part of an autoencoder. Notably, the model featured command and parameter test accuracies of 93.50 and 68.30 % and resulted in saliency images that clearly focus on the input sketch lines.

Technical Skills

Computer Vision
AWS Cloud Computing
Python
Data Science
PyTorch
Research
Machine Learning
Deep Learning