Stanford CS348n Homework 3

Dependencies

This codebase uses Python 3.6 and PyTorch 1.5.1.

Please install the python environments

pip install -r requirements.txt

Problem 1

Implement the two incomplete functions in data.py, and then use prob1.ipynb for visualizing the results.

Problem 2

First, go to the folder cd and follow the README there to compile the GPU implementation of Chamfer distance.

Then, download the data and unzip it under the home directory.

For problem 2 (a), implement the three pieces of code in model.py and run

python train.py

If you implement things correctly, you should be able to see promising reconstruction results at epoch 3-5. Please train the model for 10 epochs (roughly 2-3 hours) and report your curves and results.

Download the pretrained models and unzip it under the home directory.

For problem 2 (b), implement the missing code in recon.py and run

python recon.py

For problem 2 (c), implement the missing code in randgen.py and run

python randgen.py

For problem 2 (d), implement interp.py and run

python interp.py

Codebase Structure and Experiment Log

Several main files

• model.py: defines the NN;
• data.py: contains the dataset and dataloader code;
• train.py: the main trainer code;
• recon.py: for shape reconstruction;
• randgen.py: for shape free generation;
• interp.py: for shape interpolation;
• part_semantics_Chair.txt: the canonical chair hierarchy template.

Each experiment run will output a directory with name exp-vae for VAE, and it contains

• ckpts: store the model checkpoint every 1000 steps of training;
• train and val: store the tensorboard information, used for training and validation curve visualization;
• val_visu: contains the validation results for one batch every 10 epoches of training;
• conf.pth: stores all parameters for the training;
• data.py and train_ae.py: backups the training and data code.
• train_log.txt: stores the console output for the training.

Questions

Please ask TA or post on Piazza for any question. Please do not use the Github Issue for Q&As for the class.