optas / shapegf

ShapeGF slightly adapted for building ShapeTalk's SGF-AE

ShapeGF adapted to work with ShapeTalk & ChangeIt3D

This codebase is adapted from ShapeGF's original repository

Please follow the following setup steps to enable the usage of our pre-trained ShapeGF (SFG) AE with ShapeTalk, within the ChangeIt3D codebase.

Below we assume that you have installed the changeit3d conda environment as instructed here.

The following dependencies of ShapeGF are not included in the changeit3d environment, by default, and you need to install them separately now:

• pathos
• tensorboardX

To install them in your changeit3d conda environment and use a ChangeIt3D network trained with ShapeFG. Do:

conda activate changeit3d

conda install tensorboardX
conda install -c conda-forge pathos

git clone https://github.com/optas/ShapeGF.git
cd ShapeGF

Now, (inside the ShapeGF) repo continue like this:

Download the pretrained checkpoint.

wget http://download.cs.stanford.edu/orion/changeit3d/shapeGF_ckpt.zip .
unzip shapeGF_ckpt.zip; rm -rf shapeGF_ckpt.zip

And run:

python latents_interface.py \
    configs/recon/shapenet/shapetalk_public_recon.yaml \
    --pretrained shapeGF_ckpt/epoch_1199_iters_386400.pt 

Running the above produces SGF-latent-interface-pub.pkl at the top-level directory. Now, given shape latents in an np.array (zs) you can decode them like this:

import dill as pickle
with open('SGF-latent-interface-pub.pkl', 'wb') as f: 
    sgf = pickle.load(f)

sgf.eval_z(zs, save_output=True, output_dir=OUTPUT_FOLDER)  # OPTION 1: save outputs
outputs = sgf.eval_z(zs) # OPTION 2: returns outputs