AdditiveParts

UC Berkeley EECS282 Project

Overview

This project has 4 components, each finished by a different team member. The parts are: baseline, 3D CNN, Mesh CNN, and pointcloud encoder. We have streamlined some portions of our process for ease of replicating our results, to test our code for yourself, please check the run folder.

Additional hyperparameter tuning and discussion can be found at the following links:

3D CNN Regression

3D CNN Classification

PCE Regression

PCE Classification

File Structure

Please download the files from this folder: https://drive.google.com/drive/folders/1jCzhg4bwJk7lqaEBTcfoWZch0BIdxavM?usp=sharing and arrange into the following structure

BASE_DIR
├── AdditiveParts
├── rawcloud
├── rawcloud.json
├── rawnorm
├── rawnorm.json
├── repaired_files
├── sanitized_dict.json

Baseline

Stable Pose

1. Run the run/stable_pose_baseline.ipynb notebook

Optimized Normal

1. Run the run/normalBaseline.ipynb notebook

Point Cloud Encoder

1. Run the run/trainPCE.ipynb notebook, it walks through how to run our experiments.

MeshCNN

Downloading the Data:

Dataset: https://drive.google.com/drive/folders/1C0MGixYalkqlBkXeAsyGjXVHUfst113t?

Running the Network

0. cd into AdditiveParts/additive_parts
1. Unzip the dataset. Make sure the unzipped folder is named "sdata"
2. Move the dataset into the ./meshcnn/dataset/
3. The filepath to the data should look like ./meshcnn/dataset/sdata
4. from the root folder of the repo do cd ./meshcnn/MeshCNN/
5. Create a virtual environment using conda create -n meshcnn python=3.6.8.
6. Activate the virtual environment and run pip install -r requirements.txt inside of the MeshCNN folder.
7. Run bash scripts/keene/test.sh to test the pretrained neural network.