Composing Neural Learning and Symbolic Reasoning with an Application to Visual Discrimination

Adithya Murali, Atharva Sehgal, Paul Krogmeier, P. Madhusudan

This repository serves as an entry point into processing/running visual discrimination puzzles (VDPs). To get started:

• Read our paper here!
• Try the demo notebook (demo_vdp.ipynb) to run/make your own natural scenes VDP puzzles in <5 minutes here: !
• To replicate our results and use a wide selection of object detectors, follow the instructions in ## Installation section.

Structure

This repository is organized generally based on the generation/inference module used for our evaluation datasets. These modules are adapted versions of pre-existing tools. For attributions, we recommend looking at the ##Citations section.

.
├── analysis                # Code to reproduce the figures.
│   ├── analysis.ipynb
│   └── utils.py
├── clevr_driver.py         # Entry point into making CLEVR VDPs.
├── clevr_generate          # CLEVR Generation script.
│   └── gen_render_uconfig.py   # CLEVR rendering for custom scenes.
├── clevr_inference
│   ├── collate.py          # Collate inference results to solver's input.
│   ├── ns-vqa-readme.md    # Original instructions for installation.
│   ├── reason
│   └── scene_parse
├── data                    # Data bundle. Available at Box/Email authors.
│   ├── blender-2.78c-linux-glibc219-x86_64 # Compatible blender version
│   ├── clevr-cleaned-puzzles               # CLEVR domain puzzles (+images/IR/solver_output)
│   ├── clevr-cleaned-variants              # Variants of CLEVR puzzles              
│   ├── clevr_scene_data                    # CLEVR Generation files 
│   ├── conda-envs                          # Environment replication file
│   ├── demo_images                         # Demo attribution + CC domain images
│   ├── natscene_data                       # Data for natural scenes dataset. Please contact authors for access to this dataset!
│   ├── ns-vqa-data                         # Setup files for natural scenes inference engine.
│   ├── triplet_loss_baseline-data
│   └── yolo_data                           # Setup files for darknet.
├── demo_vdp.ipynb
├── natscene_driver.py      # Entry point to making/processing natural scenes images.
├── natscene_inference
├── protonet-baseline       # Protonet baseline
├── README.md               # This file!
├── triplet_loss_baseline   # Triplet loss baseline
├── utils                   # Miscellaneous aggregation/processing scripts.
├── vdp_solver              # Our First Order Scene Logic solver (based on Z3!)
└── vdp_website             # Data/Code to remake the hugo website!
61 directories, 89 files

Installation

Our instructions are specifically tuned to make it easy to replicate our results with minimal setup errors. A lot of these steps can be avoided in case of resource constraints.

1. Setup a Amazon AWS p2.xlarge instance. We used the Deep Learning AMI (Ubuntu 16.04) Version 38.0 (ami-067cf9e6ac97cfb53) Amazon Machine Image with 105 GiB gp2 storage volume. Running all the experiments from scratch takes about 24-30 hours of compute time (less than $30). We were able to replicate all our experiments on a different server with a different GPU. However, running the clevr_inference module requires a Ubuntu 16.04 machine.
2. Clone this git repository. The homepath should look like /home/ubuntu/vdp (this exact string has been hardcoded in a lot of places. If you prefer running this on a local machine then you'll need to group-replace this path with your the project path).
3. Install the conda environments. There should be 4 *-spec.txt files in the data/conda-envs folder. Create a conda environment from these. That is:

$ conda create --name ns-vqa --file data/conda-envs/ns-vqa-spec.txt
$ conda create --name deep-ranking --file data/conda-envs/deep-ranking-spec.txt
$ conda create --name torch --file data/conda-envs/torch-spec.txt
$ conda create --name vdp --file data/conda-envs/vdp-spec.txt

4. Download the data.zip data bundle here and merge this with the contents of vdp/data.
5. The data.zip file contains all the extra datafiles/configs/pretrained_models needed to replicate our results.
6. Setup darknet for getting bounding boxes from YOLOv4. Running $ make clean && make OR $ ./build.sh in natscene_inference/darknet should suffice. I suggest reading the readme in the darknet directory for setting up a particular GPU. The repo is setup to handle atleast the NVIDIA K80.

Running a CLEVR puzzle.

Start by making a directory data/clevr-cleaned-puzzles/test_puzzle. Make a file scene_file.json that follows this format:

[
    { # Image 0   
        "objects": [
            {
                "x": -2.0,          # [-3, 3]
                "y": -2.0,          # [-3, 3]
                "z": 0.0,           # Ideally use 0. The inference engine isn't trained for >0.
                "size": "small",    # ["small", "large"]
                "shape": "sphere",  # ["cube", "sphere", "cylinder"]
                "color": "yellow",  # ["gray", "red", "blue", "green", "brown", "purple", "cyan", "yellow"]
                "orientation": 0.0, # [0, 360)
                "material": "rubber" # ["rubber", "metal"]
            },
            ...
        ]

    },
    ...
]

-->

To run the CLEVR puzzle:

$ python clevr_driver.py --puzzle_dir data/clevr-cleaned-puzzles/test_puzzle --examples "3 4 5 6" --candidates "0 1 2" --vdp_flags "- 2 --autotune" --use_gpu

More information about vdp_flags can be found by running python vdp_solver/scripts/vdpsolve.py --help. Also, make sure you don't have any environment activated with conda deactivate before running this!

Running a Natural scenes puzzle.

1. Assume we want to process a puzzle defined by eg1.png, eg2.png, eg3.png, cand1.png, cand2.png, cand3.png.
2. Put the images in data/natscene_data/images/.
3. Create a directory data/natscene_data/puzzles/test_puzzle. Make a file config.json with the follow format:

   {
       "name": "test_puzzle",
       "train": [
           "data/natscene_data/images/eg1.png",
           "data/natscene_data/images/eg2.png",
           "data/natscene_data/images/eg3.png"
       ],
       "test": [
           "data/natscene_data/images/cand1.png",
           "data/natscene_data/images/cand2.png",
           "data/natscene_data/images/cand3.png"
       ]
   }
   

4. Run the natural scenes puzzles with the default YOLO bounding box detector:

   $ python natscene_driver.py --config "data/natscene_data/puzzles/test_puzzle/config.json" --vdp_flags "- 3 --autotune" --regenerate
   

5. The output of the solver will be written to data/natscene_data/processed/test_puzzle.out.

Discriminator Synthesis Tutorial

The command to use our solver is

$ python vdp_solver/scripts/vdpsolve.py /path/to/puzzle/IR  - <arguments>

Say your puzzle IR generated after the inference step is at /path/to/puzzle/IR and the target discriminator is $$\exists, person_1.,\exists, person_2., \exists, ball., left(person_1,,ball) \land right(person_2,, ball)$$ which says that there is a ball in between two people. Then, run the following command to see what our tool finds

$ python vdp_solver/scripts/vdpsolve.py /content/processed  - 3 2

where the 3 is the number of quantifiers and the 2 is the number of conjuncts. This tells the solver to look for a discriminator of the specified shape/size. The second argument is optional, so you can just specify the number of quantifiers if you don't know the number of conjuncts. However, if you don't specify the number of conjuncts the tool is not guaranteed to find the formula with the smallest possible number of conjuncts.

To find the smallest formula, use

$ python vdp_solver/scripts/vdpsolve.py /path/to/puzzle/IR  - 3 --autotune

This command will treat the number of quantifiers as an upper bound and return the smallest formula within the specified parameters. The order it uses is a lexicographic one with the following hierarchy: formulas are smaller if they have (1) lesser number of quantifiers (2) existentially quantified rather than universally (3) outer quantifiers are existential (so $\exists\forall$ is smaller than $\forall\exists$$) (4) lesser conjuncts, if the quantification pattern is the same.

Some other helpful arguments:

• -E find a discriminator that is purely existentially quantified
• -A find a discriminator that is purely universally quantified

All these options can also be used with --autotune. Parameters are usually treated as an exact description of the target discriminator, but they are interpreted as an upper bound if --autotune is used.

To see all these options and more, see the help menu by running

python vdp_solver/scripts/vdpsolve.py .  - -h

Some notes:

• Make sure that you are in the project directory (.../vdp) when you run these commands to avoid any path resolution issues.
• Notice the - in the middle of the command!
• The discriminators are assumed to be in prenex form. The options above specify the size and quantifier pattern only for such formulas. If your target discriminator is not in prenex form, convert the formula to prenex form first and then calculate what options are needed.
• We usually use a timeout of 1 to 2 minutes. If our solver can't find a discriminator by then, it either terminates saying that there is no discriminator in the given search space or runs out of resources.
• The solver works fastest when the number of quantifiers is 3 or lower. When there are three quantifiers, we encourage using -E first before using --autotune since existential-only discriminators are easier to find if they exist, and also easier to report if they don't exist.

Data

The data/ folder contains all data used by the tool chain. This includes pretrained models, csv files, json configuration files, etc. The data directory can be downloaded here and is subject to a fair use in research agreement (8.6 GB): https://utexas.box.com/s/rxd0fqoqya721bo3uqaaikpp9a7uk8v8. This does not include the natural scenes images because we weren't able to track the licences for some of the files. Please email us directly about this.

Contact Us

The fastest way to reach the authors for any debugging help is by opening a GitHub issue. This also helps us document and address potential "gotchas." For data related issues, please email the authors directly at adithya5@illinois.edu or atharvas@utexas.edu.

Common Gotcha's (Known Issues)

• The triplet_loss_baseline is written in Tensorflow/Keras. Keras documents default behavior in ~/.keras/keras.json. The default behavior of image_dim_ordering is different for the triplet loss baseline so if you run into batch dimension errors, messing with the image_dim_ordering will help.

Citations

The following external tools were used for this project:

• GQA Dataset

  @article{hudson2018gqa,
      title={GQA: A New Dataset for Real-World Visual Reasoning 
      and Compositional Question Answering},
      author={Hudson, Drew A and Manning, Christopher D},
      journal={Conference on Computer Vision and Pattern Recognition (CVPR)},
      year={2019}
  }
  

• NS-VQA inference pipeline

  Link: https://github.com/kexinyi/ns-vqa

  @inproceedings{yi2018neural,
  title={Neural-symbolic vqa: Disentangling reasoning from vision and language understanding},
  author={Yi, Kexin and Wu, Jiajun and Gan, Chuang and Torralba, Antonio and Kohli, Pushmeet and Tenenbaum, Joshua B.},
  booktitle={Advances in Neural Information Processing Systems},
  pages={1039--1050},
  year={2018}
  }
  

• CLEVR dataset generator

  Link: https://github.com/facebookresearch/clevr-dataset-gen

  @inproceedings{johnson2017clevr,
    title={CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning},
    author={Johnson, Justin and Hariharan, Bharath and van der Maaten, Laurens
            and Fei-Fei, Li and Zitnick, C Lawrence and Girshick, Ross},
    booktitle={CVPR},
    year={2017}
  }
  

• Deep Ranking Similarity Baseline

  Link: https://github.com/USCDataScience/Image-Similarity-Deep-Ranking

  • Code

    Dr. Chris Mattmann, JPL.
    Purvak Lapsiya, USC.
    

  • Paper

    @article{DBLP:journals/corr/WangSLRWPCW14,
      author    = {Jiang Wang and Yang Song and Thomas Leung and Chuck Rosenberg and Jingbin Wang and James Philbin and Bo Chen and Ying Wu},
      title     = {Learning Fine-grained Image Similarity with Deep Ranking},
      journal   = {CoRR},
      volume    = {abs/1404.4661},
      year      = {2014},
      url       = {http://arxiv.org/abs/1404.4661},
      archivePrefix = {arXiv},
      eprint    = {1404.4661},
      timestamp = {Tue, 08 Sep 2020 16:29:29 +0200},
      biburl    = {https://dblp.org/rec/journals/corr/WangSLRWPCW14.bib},
      bibsource = {dblp computer science bibliography, https://dblp.org}
    }
    

• YOLOv4

  Link: https://github.com/pjreddie/darknet

  @article{DBLP:journals/corr/abs-2011-08036,
    author    = {Chien{-}Yao Wang and
                 Alexey Bochkovskiy and
                 Hong{-}Yuan Mark Liao},
    title     = {Scaled-YOLOv4: Scaling Cross Stage Partial Network},
    journal   = {CoRR},
    volume    = {abs/2011.08036},
    year      = {2020},
    url       = {https://arxiv.org/abs/2011.08036},
    archivePrefix = {arXiv},
    eprint    = {2011.08036},
    timestamp = {Wed, 18 Nov 2020 16:48:35 +0100},
    biburl    = {https://dblp.org/rec/journals/corr/abs-2011-08036.bib},
    bibsource = {dblp computer science bibliography, https://dblp.org}
  }