Skip to content

google/ldif

Repository files navigation

Overview

alt text alt text

This is a joint codebase for LDIF (Local Deep Implicit Functions for 3D Shape) and SIF (Learning Shape Templates with Structured Implicit Functions). Note that LDIF was previously called Deep Structured Implicit Functions. It contains code to reproduce the results of those papers, convert input meshes into the LDIF and SIF representations, and visualize and extract meshes. representations.

All .py and .sh files in the top-level ldif/ directory are entry points into the code (train.py, eval.py, meshes2dataset.py, unit_test.sh, and reproduce_shapenet_autoencoder.sh). The rest of this README provides information on initial setup and basic documentation for those files. For additional documentation, please see each file.

Environment

To set up the LDIF/SIF environment, follow these steps:

1. Set up the python environment

The code was tested with python 3.6 and tensorflow 1.15 on linux. There is a requirements.txt containing all dependencies.

If you use anaconda, run the following:

conda env create --name ldif -f environment.yml
conda activate ldif

If you use a system pip installation, run pip install -r requirements.txt

After this, the python environment should be ready to go. Please activate the environment before proceeding. The build scripts include some python.

2. Build GAPS

./build_gaps.sh

GAPS is a geometry processing library used by this package to generate the data and create interactive visualizations. The script build_gaps.sh does the following. One, it installs the necessary dependencies with apt. If sudo is not available on the system, the requirements are that GAPS have include access to standard OpenGL and GLu library headers (GL/gl.h, GL/glu.h) (on both linux and macos), and that OSMesa static libraries can be linked (on linux). If these are satisfied, the sudo line can be commented out. Two, it clones the GAPS repository from GitHub, make some changes, and builds it. It also moves the qview folder into the gaps repository and modifies the makefiles. The qview executable is a C++ program written using GAPS to visualize SIF and LDIF representations. Finally, the script compiles all necessary GAPS C++ executables, which are called by the python code. If this step was successful, running ./gaps_is_installed.sh should echo Ready to go!

GAPS should compile with no warnings. Please report any warnings by opening a GitHub issue- the information would be greatly appreciated.

3. Build the inference kernel (Optional, but highly recommended)

./build_kernel.sh

If successful, there should be a binary ldif2mesh in the ldif/ldif2mesh/ subdirectory. Note that the inference kernel assumes the CUDA toolkit is installed and that a gpu supporting compute 6.1 (Pascal, so 10-series or newer) is available. The nvcc command is part of the CUDA toolkit. If you have an older gpu, you can try older compute versions for --gpu-architecture and --gpu-code, but performance may be reduced and some newer features are used, so it might not compile.

If you do not want to use the inference kernel or don't have a GPU, then you can pass --nouse_inference_kernel to eval.py, which is the only script that typically calls the kernel. It will then use pure tensorflow ops for evaluating LDIF, as is done during training (for autodiff support). However, it would be orders of magnitude slower, so it is really not recommended if more than ~20 meshes need to be evaluated.

The kernel should compile with no warnings. Please report any warnings by opening a GitHub issue- this information would be greatly appreciated.

Datasets

To run LDIF/SIF, first a dataset should be made. The input to this step is a directory of watertight meshes, and the output is a directory containing the files needed to train and evaluate LDIF/SIF.

Create an input directory somewhere on disk, with the following structure:

[path/to/root]/{train/val/test}/{class names}/{.ply files}

The properties of the dataset (# and name of classes, size of the splits, name of examples, etc.) are determined from the directory structure. If you want to reproduce the shapenet results, then see ./reproduce_shapenet_autoencoder.sh. A dataset doesn't need to have a train, test, and val split, only whichever splits you want to use. You could make a dataset with just a test split for a comparison, for example. Note that for convenience the code tries to check if the class names are wordnet synsets and will convert them to shapenet names (i.e. 02691156 -> airplane) if they are-- if it can't detect a synset it will just use the folder name as the class name.

Note that .ply files are required, but the GAPS library provides a shell utility for converting between file formats. You can do ./ldif/gaps/bin/x86_64/msh2msh mesh.obj mesh.ply as an example conversion, which will read mesh.obj and write a new file mesh.ply to disk.

It is very important that the input meshes be watertight at training time. GAPS provides a program msh2df that can do the conversion, if you are not interested in exactly replicating the OccNet experiment's process. Here is an example command that will make a unit-cube sized mesh watertight:

./ldif/gaps/bin/x86_64/msh2df input.ply tmp.grd -estimate_sign -spacing 0.002 -v
./ldif/gaps/bin/x86_64/grd2msh tmp.grd output.ply
rm tmp.grd

Msh2df outputs an SDF voxel grid, while grd2msh runs marching cubes to extract a mesh from the generated SDF grid. The msh2df algorithm rasterizes the mesh to a voxel grid and then floodfills at a resolution determined by the -spacing parameter in order to determine the sign. The smaller the value, the higher the resolution, the smaller the smallest allowable hole in the mesh, and the slower the algorithm. The bigger the value, the lower the resolution, the bigger the smallest allowable hole in the mesh, and the faster the algorithm. The run time of both msh2df and of the rest of the dataset creation pipeline will vary greatly depending on the -spacing parameter. The default value of 0.002 is quite high resolution for a mesh the size of a unit cube.

While msh2df is provided as a utility, it was not used to generate the data for the trained LDIF+SIF models. For reproducing the shapenet results, please use the TSDF fusion package used by the OccNet repository, not msh2df.

To actually make a dataset once watertight meshes are available, run:

python meshes2dataset.py --mesh_directory [path/to/dataset_root] \
  --dataset_directory [path/to/nonexistent_output_directory]

Please see meshes2dataset.py for more flags and documentation. To avoid excess disk usage (and avoid having to pass in the input directory path to all subsequent scripts), symlinks are created during this process that point to the meshes in the input directory. Please do not delete or move the input directory after dataset creation, or the code won't have access to the ground truth meshes for evaluation.

The dataset generation code writes 7-9mb of data per mesh (about 330GB for shapenet-13).

Training

To train a SIF or LDIF, run the following:

python train.py --dataset_directory [path/to/dataset_root] \
  --experiment_name [name] --model_type {ldif, sif, or sif++}

The dataset directory should be whatever it was set to when running meshes2dataset.py. The experiment name can be arbitrary, it is a tag used to load the model during inference/eval/interactive sessions. The model_type determines what hyperparameters to use. ldif will train a 32x32 LDIF with 16 symmetric and 16 asymmetric elements. sif will replicate the SIF representation proposed in the SIF paper. sif++ will train an improved version of SIF using the loss and network from LDIF, as well as gaussians that support rotation, but without any latent codes per element. By default trained models are stored under {root}/trained_models/, but this can be changed with the --model_directory flag. For more flags and documentation, please see train.py.

It is also possible to make model types besides the paper version of LDIF/SIF. For details, please see ldif/model/hparams.py. Both LDIF and SIF are stored as specific hparam combos. Adding a new combo and/or new hyperparameters would be the easiest way to evaluate how a modification to LDIF/SIF would change the performance. It would also be how to turn off partial symmetry, or adjust the number of shape elements or size of the latent codes. The only special hyperparameter is batch size, which is read directly by the train.py script, and always set to 1 during inference.

While training, the model write tensorboard summaries. If you don't have tensorboard, you can install it with conda install tensorboard or pip install tensorboard. Then you can run

tensorboard --logdir [ldif_root]/trained_models/sif-transcoder-[experiment_name]/log

assuming that --model_root was set to the default ldif_root]/trained_models/

Warning: Training an LDIF from scratch takes a long time. SIF also takes a while, though not nearly as long. The expected performance with a V100 and a batch size of 24 is 3.5 steps per second for LDIF, 6 steps per second for SIF. LDIF takes about 3.5M steps to fully converge on ShapeNet, while SIF takes about 700K. So that is about 10 days to train an LDIF from scratch, and about 32 hours for SIF. Note that LDIF performance is pretty reasonable after 3-4 days, so depending on your uses it may not be necessary to wait the whole time. The plan is to 1) add pretrained checkpoints (the most pressing TODO) and 2) add multi-gpu support, later on, to help mitigate this issue. Another practical option might be switching out the encoder for a smaller one, because most of the training time is the forward+backward pass on the ResNet50.

Evaluation and Inference

To evaluate a fully trained LDIF or SIF network, run the following:

python eval.py --dataset_directory [path/to/dataset_root] \
  --experiment_name [name] --split {test, val, or train}

This will compute metrics over the dataset and then print out the result to the terminal. By default, it will print out a table of results and LaTeX code. In addition, there are flags --save_results, --save_meshes, and --save_ldifs, which can be set to true. If they are set, the code will also write

  1. pandas-readable CSV files containing the metrics for each mesh and class,
  2. a directory of result meshes generated by the algorithm, and/or
  3. a directory of txt files containing the actual LDIF/SIF representation (parseable by qview, ldif2mesh, and the Decoder class in the ipynb). If these flags are set, then --result_directory must also be provided indicating where they should be written.

Interactive Sessions

You can run the code interactively in an Jupyter notebook. To do so, open the provided file ldif_example_inference.ipynb with Jupyter and attach it to a python 3.6 kernel with the requirements.txt installed. Then follow the prompts in the notebook. The notebook has a demo of loading a mesh, creating an example, running inference, visualizing the underlying SIF elements, extracting a mesh, and computing metrics. There is additional documentation in the .ipynb.

Unit Tests

There is a script unit_test.sh. If you want to check whether the code is installed correctly and works, run it with no arguments. It will make a small dataset using open source models and train/evaluate an LDIF network. Note that it doesn't train to convergence so that it doesn't take very long to run. As a result, the final outputs don't look very good. You can fix this by setting the step count in unit_test.sh higher (around 50K steps is definitely sufficient).

The code also has some unit tests for various pieces of functionality. To run a test, cd into the directory of the *_test.py file, and run it with no arguments. Please be aware that not all of the code is tested, and that the unit tests aren't well documented. The easiest way to check if the code still works is by running ./unit_test.sh.

Other code and PyTorch

In addition to the scripts described above, there are also model definitions and beam pipelines provided for generating datasets and running inference on a larger scale. To use these scripts, it would be necessary to hook up your own beam backend.

There is also very limited PyTorch support in the ldif/torch directory. This code is a basic implementation of SIF that can't train new SIF models, but can load and evaluate SIFs generated by the tensorflow training+evaluation code. It is mainly intended for using SIF correspondences as a building block of another unrelated project in PyTorch. Note that PyTorch is not included in the requirements.txt, and the torch/ subdirectory is independent from the rest of the code base (it interacts only through the .txt files written by the tensorflow code and takes no dependencies on the rest of this codebase). To use it, it is probably easiest to just download the torch/ folder and import the sif.py file as a module.

Updates to the code

  • The code now supports tfrecords dataset generation and usage. This reduces the IO workload done during training. It is enabled by default. Existing users can git pull, rerun meshes2dataset.py with --optimize and --optimize_only, and then resume training where they left off with the new dataset improvements. If you currently experience less than 100% GPU utilization, it is highly recommended. Note it increases dataset size by 3mb per example (and can be disabled with --nooptimize).

  • Support for the inference kernel on Volta, Turing and CC 6.0 Pascal cards should now work as intended. If you had trouble with the inference kernel, please git pull and rerun ./build_kernel.sh.

TODOS

This is a preliminary release of the code, and there are a few steps left:

  • Pretrained model checkpoints are on the way. In the mean-time, please see reproduce_shapenet_autoencoder.sh for shapenet results.
  • This code base does not yet support training a single-view network. In the mean-time, the single-view network architecture has been provided (see ldif/model/hparams.py for additional information).
  • While the eval code is fast enough for shapenet, the post-kernel eval code is written in numpy and is an unnecessary bottleneck. So inference at 256^3 takes a few seconds per mesh, even though the kernel completes in ~300ms.
  • Pressing the 'f' key in a qview visualization session will extract a mesh and show it alongside the SIF elements; however it only considers the analytic parameters. Therefore, it shows a confusing result for LDIF representations, which also have neural features.
  • To make setup easier, we would like to provide a docker container that is ready to go.