Zero Shot Prediction Plugin

This plugin allows you to perform zero-shot prediction on your dataset for the following tasks:

• Image Classification
• Object Detection
• Instance Segmentation
• Semantic Segmentation

Given a list of label classes, which you can input either manually, separated by commas, or by uploading a text file, the plugin will perform zero-shot prediction on your dataset for the specified task and add the results to the dataset under a new field, which you can specify.

Updates

• 2024-03-06: Added support for YOLO-World for object detection and instance segmentation!
• 2024-01-10: Removing LAION CLIP models.
• 2024-01-05: Added support for EVA-CLIP, SigLIP, and DFN CLIP for image classification!
• 2023-11-28: Version 1.1.1 supports OpenCLIP for image classification!
• 2023-11-13: Version 1.1.0 supports calling operators from the Python SDK!
• 2023-10-27: Added support for MetaCLIP for image classification
• 2023-10-20: Added support for AltCLIP and Align for image classification and GroupViT for semantic segmentation

Requirements

• To use YOLO-World models, you must have "ultalytics>=8.1.42".

Models

Built-in Models

As a starting point, this plugin comes with at least one zero-shot model per task. These are:

Image Classification

• ALIGN
• AltCLIP
• CLIP: (OpenAI)
• CLIPA
• DFN CLIP: Data Filtering Networks
• EVA-CLIP
• MetaCLIP
• SigLIP

Object Detection

• YOLO-World
• Owl-ViT

Instance Segmentation

• Owl-ViT + Segment Anything (SAM)
• YOLO-World + Segment Anything (SAM)

Semantic Segmentation

• CLIPSeg
• GroupViT

Most of the models used are from the HuggingFace Transformers library, and CLIP and SAM models are from the FiftyOne Model Zoo

Note— For SAM you will need to have Facebook's segment-anything library installed.

Adding Your Own Models

You can see the implementations for all of these models in the following files:

• classification.py
• detection.py
• instance_segmentation.py
• semantic_segmentation.py

These models are "registered" via dictionaries in each file. In semantic_segmentation.py, for example, the dictionary is:

SEMANTIC_SEGMENTATION_MODELS = {
    "CLIPSeg": {
        "activator": CLIPSeg_activator,
        "model": CLIPSegZeroShotModel,
        "name": "CLIPSeg",
    },
    "GroupViT": {
        "activator": GroupViT_activator,
        "model": GroupViTZeroShotModel,
        "name": "GroupViT",
    },
}

The activator checks the environment to see if the model is available, and the model is a fiftyone.core.models.Model object that is instantiated with the model name and the task — or a function that instantiates such a model. The name is the name of the model that will be displayed in the dropdown menu in the plugin.

If you want to add your own model, you can add it to the dictionary in the corresponding file. For example, if you want to add a new semantic segmentation model, you can add it to the SEMANTIC_SEGMENTATION_MODELS dictionary in semantic_segmentation.py:

CLASSIFICATION_MODELS = {
    "CLIPSeg": {
        "activator": CLIPSeg_activator,
        "model": CLIPSegZeroShotModel,
        "name": "CLIPSeg",
    },
    "GroupViT": {
        "activator": GroupViT_activator,
        "model": GroupViTZeroShotModel,
        "name": "GroupViT",
    },
    ..., # other models
    "My Model": {
        "activator": my_model_activator,
        "model": my_model,
        "name": "My Model",
    }
}

💡 You need to implement the activator and model functions for your model. The activator should check the environment to see if the model is available, and the model should be a fiftyone.core.models.Model object that is instantiated with the model name and the task.

Watch On Youtube

Installation

fiftyone plugins download https://github.com/jacobmarks/zero-shot-prediction-plugin

If you want to use AltCLIP, Align, Owl-ViT, CLIPSeg, or GroupViT, you will also need to install the transformers library:

pip install transformers

If you want to use SAM, you will also need to install the segment-anything library:

pip install git+https://github.com/facebookresearch/segment-anything.git

If you want to use OpenCLIP, you will also need to install the open_clip library from PyPI:

pip install open-clip-torch

Or from source:

pip install git+https://github.com/mlfoundations/open_clip.git

If you want to use YOLO-World, you will also need to install the ultralytics library:

pip install -U ultralytics

Usage

All of the operators in this plugin can be run in delegated execution mode. This means that instead of waiting for the operator to finish, you schedule the operation to be performed separately. This is useful for long-running operations, such as performing inference on a large dataset.

Once you have pressed the Schedule button for the operator, you will be able to see the job from the command line using FiftyOne's command line interface:

fiftyone delegated list

will show you the status of all delegated operations.

To launch a service which runs the operation, as well as any other delegated operations that have been scheduled, run:

fiftyone delegated launch

Once the operation has completed, you can view the results in the App (upon refresh).

After the operation completes, you can also clean up your list of delegated operations by running:

fiftyone delegated cleanup -s COMPLETED

Operators

zero_shot_predict

• Select the task you want to perform zero-shot prediction on (image classification, object detection, instance segmentation, or semantic segmentation), and the field you want to add the results to.

zero_shot_classify

• Perform zero-shot image classification on your dataset

zero_shot_detect

• Perform zero-shot object detection on your dataset

zero_shot_instance_segment

• Perform zero-shot instance segmentation on your dataset

zero_shot_semantic_segment

• Perform zero-shot semantic segmentation on your dataset

Python SDK

You can also use the compute operators from the Python SDK!

import fiftyone as fo
import fiftyone.operators as foo
import fiftyone.zoo as foz

dataset = fo.load_dataset("quickstart")

## Access the operator via its URI (plugin name + operator name)
zsc = foo.get_operator("@jacobmarks/zero_shot_prediction/zero_shot_classify")

## Run zero-shot classification on all images in the dataset, specifying the labels with the `labels` argument
zsc(dataset, labels=["cat", "dog", "bird"])

## Run zero-shot classification on all images in the dataset, specifying the labels with a text file
zsc(dataset, labels_file="/path/to/labels.txt")

## Specify the model to use, and the field to add the results to
zsc(dataset, labels=["cat", "dog", "bird"], model="CLIP", field="predictions")

## Run zero-shot detection on a view
zsd = foo.get_operator("@jacobmarks/zero_shot_prediction/zero_shot_detect")
view = dataset.take(10)
zsd(
    view,
    labels=["license plate"],
    model="OwlViT",
    field="owlvit_license_plate",
)

All four of the task-specific zero-shot prediction operators also expose a list_models() method, which returns a list of the available models for that task.

zsss = foo.get_operator(
    "@jacobmarks/zero_shot_prediction/zero_shot_semantic_segment"
)

zsss.list_models()

## ['CLIPSeg', 'GroupViT']

Note: The zero_shot_predict operator is not yet supported in the Python SDK.

Note: With earlier versions of FiftyOne, you may have trouble running these operator executions within a Jupyter notebook. If so, try running them in a Python script, or upgrading to the latest version of FiftyOne!