Visual Programming: Compositional visual reasoning without training (CVPR 2023)

This repository contains the official code for VisProg - a neuro-symbolic system that solves complex and compositional visual tasks given natural language instructions. VisProg uses the in-context learning ability of GPT3 to generate python programs which are then executed to get both the solution and a comprehensive and interpretable rationale. Each line of the generated program may invoke one of several off-the-shelf computer vision models, image processing routines, or python functions to produce intermediate outputs that may be consumed by subsequent parts of the program.

This code base has been designed to be:

✅ easy to use (a simple ipynb per task)
✅ easy to extend with new functionality by adding new modules to VisProg
✅ easy to extend to new tasks by adding in-context examples for these tasks
✅ minimal and modular to make it easy to dig into and build upon

Install Dependencies

conda env create -f environment.yaml
conda activate visprog

Running the notebooks

Having setup and activated the conda environment, you should be all set to run the notebooks in the notebooks/ folder. If you use an editor like VSCode, openning the .ipynbs within VSCode might be the easiest way to get started.

You will find a notebook for each of the following tasks, but they are quite similar in structure:

Outside knowledge object tagging: notebooks/ok_det.ipynb
Natural language image editing: notebooks/image_editing.ipynb
NLVR: notebooks/nlvr.ipynb
GQA: notebooks/gqa.ipynb

Simply, enter your OpenAI API key in the cell that currently reads <Enter your key here> and run the notebook. The notebooks are designed to be self-contained and should run end-to-end without any additional setup.

The basic structure of the notebooks is as follows:

Setup paths
Set OPENAI_API_KEY environment variable to use GPT3
Import ProgramGenerator and ProgramInterpreter classes
Import PROMPT (a text string containing in-context examples) or create_prompt (a function that creates the prompt on the fly)
Create the ProgramGenerator and ProgramInterpreter objects
Load the image or images to perform inference on
Specify the natural language question / statement / instruction
Generate program from the specified instruction using ProgramGenerator
Interpret and execute program using ProgramInterpreter
Visualize the returned result and visual rationale (execution trace)

Example Output

We have tried to make it easy to visualize each step of the execution trace.

For instance, when running the gqa notebook for the instruction How many people or animals are in the image? on assets/camel1.png, you should see the following outputs:

Program

BOX0=LOC(image=IMAGE,object='people')
BOX1=LOC(image=IMAGE,object='animals')
ANSWER0=COUNT(box=BOX0)
ANSWER1=COUNT(box=BOX1)
ANSWER2=EVAL(expr="{ANSWER0} + {ANSWER1}")
FINAL_RESULT=RESULT(var=ANSWER2)

Visual Rationale

What if VisProg doesn't solve your task?

It is possible that the instruction you provide is not solved correctly by VisProg. This can happen for a few reasons:

The instruction is very different from in-context examples that VisProg has seen before. In this case, even though the current set of modules may be adequate for solving the task, VisProg failed because of incorrect program generation. In this case, see if you can write a program using VisProg's modules that solves the task. If you can, then you may add this program to the in-context examples and re-run the notebook to handle similar instructions.
The problem is not solvable with the current set of modules in VisProg. If this is the case, you can add new modules to VisProg to solve this task. See the next section for details.

Adding new functionality and ability to solve new tasks

Add new modules for enabling these functionalities to engine/step_interpreters.py. Don't forget to register these modules in register_step_interpreters function in the same file. Here's the step interpreter for the COUNT module. All modules have a similar structure with a parse, html, and execute function. The parse function parses the program string to extract the arguments and output variable. The html function generates the html representation for the execution trace. The execute function executes the module and returns the output and the html (if inspect=True) for the execution trace.

class CountInterpreter():
    step_name = 'COUNT'

    def __init__(self):
        print(f'Registering {self.step_name} step')

    def parse(self,prog_step):
        parse_result = parse_step(prog_step.prog_str)
        step_name = parse_result['step_name']
        box_var = parse_result['args']['box']
        output_var = parse_result['output_var']
        assert(step_name==self.step_name)
        return box_var,output_var

    def html(self,box_img,output_var,count):
        step_name = html_step_name(self.step_name)
        output_var = html_var_name(output_var)
        box_arg = html_arg_name('bbox')
        box_img = html_embed_image(box_img)
        output = html_output(count)
        return f"""<div>{output_var}={step_name}({box_arg}={box_img})={output}</div>"""

    def execute(self,prog_step,inspect=False):
        box_var,output_var = self.parse(prog_step)
        boxes = prog_step.state[box_var]
        count = len(boxes)
        prog_step.state[output_var] = count
        if inspect:
            box_img = prog_step.state[box_var+'_IMAGE']
            html_str = self.html(box_img, output_var, count)
            return count, html_str

        return count

Add your in-context examples to a new file prompts/your_task_or_dataset_name.py. Note that instead of using in-context examples to generate programs, you may experiment with different ways of prompting such as providing function signatures and docstrings without needing to change the code at all!
You can now play with examples from this dataset using a notebook similar to those in the notebooks/ folder or create a python script to run inference on a large number of examples.

Here's what VisProg can do today

A summary of currently available modules

*Note that we have replaced ViLT for VQA with a more performant model called BLIP which was recently made available on Huggingface. This shows how easy it is to swap out or upgrade modules in VisProg.

Changes since the version used in the CVPR paper

GPT3 upgraded to text-davinci-003 from text-davinci-002
VQA module upgraded from ViLT to the more performant BLIP
Changed the implementation of CLASSIFY and SELECT modules that use CLIP to use cosine similarity instead of dot product (which is the default score provided by Huggingface's CLIP model)

Citation

If you find this code useful in your research, please consider citing:

@article{Gupta2022VisProg,
  title={Visual Programming: Compositional visual reasoning without training},
  author={Tanmay Gupta and Aniruddha Kembhavi},
  journal={ArXiv},
  year={2022},
  volume={abs/2211.11559}
}

allenai/visprog

allenai

Reviews

Repository Details