Point tracking

Step-by-step tutorial on how to integrate custom point tracking neural network into Supervisely platform on the example of PIPs.

Introduction

In this tutorial you will learn how to integrate your custom point tracking model into Supervisely by creating two simple serving apps. First, we will construct a straightforward model that only moves the original point as an illustration. The SOTA model PIPs, which already has the majority of the necessary functions implemented, will be used in the second part.

Implementation details

To integrate your model, you need to subclass sly.nn.inference.PointTracking and implement 2 methods:

• load_on_device method for downloading the weights and initializing the model on a specific device. Takes a model_dir argument, which is a directory for all model files (like configs, weights, etc). The second argument is a device - a torch.device like cuda:0, cpu.

• predict. The core implementation of model inference. It takes a list of images of np.ndarray type, inference settings and point to track as arguments, applies the model inference to the images and returns a list of predictions (both input point and predicted points are sly.nn.PredictionPoint objects).

Currently, integrating models that can track several points simultaneously is not possible due to the implementation of the sly.nn.inference.PointTracking class.

Overall structure

The overall structure of the class we will implement is looking like this:

class MyModel(sly.nn.inference.PointTracking):
    def load_on_device(
        self,
        model_dir: str,
        device: Literal["cpu", "cuda", "cuda:0", "cuda:1", "cuda:2", "cuda:3"] = "cpu",
    ):
        # preparing the model: model instantiating, downloading weights, loading it on the device.
        pass

    def predict(
        self,
        rgb_images: List[np.ndarray],
        settings: Dict[str, Any],
        start_object: PredictionPoint,
    ) -> List[PredictionPoint]:
        # the inference of a model here
        # ...
        return prediction

The superclass has a serve() method. For running the code on the Supervisely platform, m.serve() method should be executed:

if sly.is_production():
    m.serve()

And here is the beauty comes in. The method serve() internally handles everything and deploys your model as a REST API service on the Supervisely platform. It means that other applications are able to communicate with your model and get predictions from it.

So let's implement the class.

Simple model

Getting started

Step 1. Prepare ~/supervisely.env file with credentials. Learn more here.

Step 2. Create Virtual Environment and install supervisely==6.72.11 in it.

Step-by-step implementation

Defining imports and global variables

import numpy as np
from dotenv import load_dotenv
from pathlib import Path
from typing import Any, Dict, List, Literal
from typing_extensions import Literal

import supervisely as sly
import supervisely.imaging.image as sly_image
from supervisely.nn.inference import PointTracking
from supervisely.nn.prediction_dto import PredictionPoint


load_dotenv("local.env")
load_dotenv(os.path.expanduser("~/supervisely.env"))

1. load_on_device

The following code creates the model according to config model_settings.yaml. Path to .yaml config is passed during initialization. This settings can also be given as a python dictionary. Config in the form of a dictionary becomes available in self.custom_inference_settings_dict attribute. Also load_on_device will keep the model as a self.model for further use:

class MyModel(PointTracking):
    def load_on_device(
        self,
        model_dir: str,
        device: Literal["cpu", "cuda", "cuda:0", "cuda:1", "cuda:2", "cuda:3"] = "cpu",
    ):
        col_shift = self.custom_inference_settings_dict.get("col_shift", 10)
        row_shift = self.custom_inference_settings_dict.get("row_shift", 7)

        self.model = lambda point, name: PredictionPoint(
            class_name=name,
            col=point.col + col_shift,
            row=point.row + row_shift,
        )

Our settings.yaml file:

col_shift: 20
row_shift: 15

2. predict

The core method for model inference. Here we will use the defined model and make sure that predicted points are not outside of the bounds.

    def predict(
        self,
        rgb_images: List[np.ndarray],
        settings: Dict[str, Any],
        start_object: PredictionPoint,
    ) -> List[PredictionPoint]:
        name = start_object.class_name
        pred_points = []
        frame_range = len(rgb_images)
        point = start_object

        maxh, maxw, _ = rgb_images[0].shape

        for _ in range(frame_range):
            # predict next point
            new_point = self.model(point, name)

            # check bounds
            new_point.col = min(new_point.col, maxh - 1)
            new_point.row = min(new_point.row, maxw - 1)
            pred_points.append(new_point)

            # next point
            point = new_point

        return pred_points

It must return exactly a list of sly.nn.PredictionPoint objects for compatibility with Supervisely. Notice, that the first frame is not in the list.

Usage of our class

Once the class is created, here we initialize it and get one test prediction for debugging.

settings = "model_settings.yaml"
# or just use dict 
# settings = {"col_shift": 20, "row_shift": 15}

images_path = Path("demo_images")

m = MyModel(model_dir="", custom_inference_settings=settings)

if sly.is_production():
    # this code block is running on Supervisely platform in production
    # just ignore it during development
    m.serve()
else:
    # for local debugging
    settings = m.custom_inference_settings_dict
    img_names = sorted(os.listdir(images_path))
    frames = []

    # load frames
    for name in img_names:
        pth = images_path / name
        frames.append(sly_image.read(str(pth)))

    # make predictions
    start_point = PredictionPoint("", col=100, row=100)
    pred_points = m.predict(frames, settings, start_point)

    # save frames with predicted points
    m.visualize(
        pred_points,
        frames[1:],  # skip first frame
        vis_path="predictions",  # folder to save images
        thickness=10,
    )

Here are the output predictions of our simple model:

Example of Simple model work

PIPs tracking model

Let's now implement the class for pre-trained model. The majority of the code used to load and run the model is taken directly from the original repository.

Getting started

Step 1. Prepare ~/supervisely.env file with credentials. Learn more here.

Step 2. Clone repository with source code and create Virtual Environment.

git clone https://github.com/supervisely-ecosystem/pips
cd pips
source .venv/bin/activate
pip3 install -r requirements.txt

It's feasible to run the present model on the CPU, thus installing CUDA requirements is not required.

Step 3. Load model weights.

./get_reference_model.sh

Step 4. Open the repository directory in Visual Studio Code.

code -r .

Step 5. Run debug for script src/main.py

Python script

The integration script is simple:

1. Initialize model.

2. Runs inference on a demo images.

3. Predictions adds and new frames saves in chronological order.

Step-by-step implementation

Defining imports and global variables

import os
import numpy as np
import saverloader
import torch
from dotenv import load_dotenv
from pathlib import Path
from typing import Any, Dict, List, Literal
from typing_extensions import Literal
from nets.pips import Pips

import sly_functions
import supervisely as sly
import supervisely.imaging.image as sly_image
from supervisely.nn.inference import PointTracking
from supervisely.nn.prediction_dto import PredictionPoint


root = (Path(__file__).parent / ".." / ".." / "..").resolve().absolute()

load_dotenv(root / "local.env")
load_dotenv(os.path.expanduser("~/supervisely.env"))

1. load_on_device

The following code creates the model according to config supervisely/serve/model_settings.yaml. Path to .yaml config is passed during initialization. The saverloader.load function provided by the creator of the original repository loads the model state dict from model_dir. Also load_on_device will keep the model as a self.model and the device as self.device for further use:

class PipsTracker(PointTracking):
    def load_on_device(
        self,
        model_dir: str,
        device: Literal["cpu", "cuda", "cuda:0", "cuda:1", "cuda:2", "cuda:3"] = "cpu",
    ):
        frames_per_iter = self.custom_inference_settings_dict.get("frames_per_iter", 8)
        stride = self.custom_inference_settings_dict.get("stride", 4)

        self.model = Pips(S=frames_per_iter, stride=stride).to(torch.device(device))
        if model_dir:
            _ = saverloader.load(str(model_dir), self.model, device=device)
        self.model.eval()
        self.device = device

Here we are downloading the model weights from local storage, but it can be also downloaded by path in Supervisely Team Files.

2. predict

The core method for model inference. Here we are preparing images and getting an inference of the model. The function sly_functions.run_model is borrowed from the original repository. However, there are a few changes that can be made to improve quality: preserve the aspect ratio, apply padding before resizing and make sure that predicted points are not outside of the bounds. Then we wrap model predictions into sly.nn.PredictionPoint.

    def predict(
        self,
        rgb_images: List[np.ndarray],
        settings: Dict[str, Any],
        start_object: PredictionPoint,
    ) -> List[PredictionPoint]:
        class_name = start_object.class_name
        h_resized = settings.get("h_resized", 360)
        w_resized = settings.get("w_resized", 640)
        frames_per_iter = settings.get("frames_per_iter", 8)

        rgbs = [torch.from_numpy(rgb_img).permute(2, 0, 1) for rgb_img in rgb_images]
        rgbs = torch.stack(rgbs, dim=0).unsqueeze(0)
        point = torch.tensor([[start_object.col, start_object.row]], dtype=float)

        with torch.no_grad():
            traj = sly_functions.run_model(
                self.model,
                rgbs,
                point,
                (h_resized, w_resized),
                frames_per_iter,
                device=self.device,
            )

        pred_points = [PredictionPoint(class_name, col=p[0], row=p[1]) for p in traj[1:]]
        return pred_points

It must return exactly a list of sly.nn.PredictionPoint objects for compatibility with Supervisely.

Usage of our class

Once the class is created, here we initialize it and get one test prediction for debugging.

In the code below a custom_inference_settings is used. It allows us to provide custom settings that could be used in predict() (See more in Customized Inference Tutorial)

settings = root / "supervisely" / "serve" / "model_settings.yaml"

if sly.is_debug_with_sly_net() or not sly.is_production():
    model_dir = root / "reference_model"  # local debug
else:
    model_dir = Path("/weights")  # path in Docker

pips = PipsTracker(model_dir=str(model_dir), custom_inference_settings=str(settings))

if sly.is_production():
    # this code block is running on Supervisely platform in production
    # just ignore it during development
    pips.serve()
else:
    pth = Path("demo_images")
    img_names = sorted(os.listdir(pth))
    imgs = []

    for name in img_names:
        if "jpg" in name:
            imgs.append(sly_image.read(str(pth / name)))

    traj = pips.predict(
        imgs,
        pips.custom_inference_settings_dict,
        PredictionPoint("", 448, 98),
    )
    pips.visualize(
        traj,
        imgs[1:],
        vis_path="preds",
        thickness=7,
    )

Here are the output predictions of our PIPs model:

Example of PIPs model work

Run and debug

The beauty of this class is that you can easily debug your code locally in your favorite IDE.

For now, we recommend using Visual Studio Code IDE, because our repositories have prepared settings for convenient debugging in VSCode. It is the easiest way to start.

Local debug

You can run the code locally for debugging. For Visual Studio Code we've created a launch.json config file that can be selected:

Local debug

Debug in Supervisely platform

Once the code seems working locally, it's time to test the code right in the Supervisely platform as a debugging app. For that:

1. If you develop in a Docker container, you should run the container with --cap_add=NET_ADMIN option.

2. Install sudo apt-get install wireguard iproute2.

3. Define your TEAM_ID in the local.env file. Actually there are other env variables that is needed, but they are already provided in ./vscode/launch.json for you.

4. Switch the launch.json config to the Advanced debug in Supervisely platform:

Advanced Debug in Supervisely

5. Run the code.

✅ It will deploy the model in the Supervisely platform as a regular serving App that is able to communicate with all other apps in the platform:

Develop and Debug

Now you can use apps like Apply NN to Images, Apply NN to videos with your deployed model.

Or get the model inference via Python API with the help of sly.nn.inference.Session class just in one line of code. See Inference API Tutorial.

Release your code as a Supervisely App.

Once you've tested the code, it's time to release it into the platform. It can be released as an App that is shared with the all Supervisely community, or as your own private App.

Refer to How to Release your App for all releasing details. For a private app check also Private App Tutorial.

Repository structure

The structure of our GitHub repository is the following:

.
├── LICENSE
├── README.md
├── badjadataset.py
├── chain_demo.py
├── crohddataset.py
├── demo.py
├── demo_images
│   ├── 000100.jpg
|   |   ...
│   └── extract_frames.sh
├── filter_trajs.py
├── flyingthingsdataset.py
├── get_reference_model.sh
├── local.env
├── make_occlusions.py
├── make_trajs.py
├── nets
│   ├── pips.py
│   ├── raft_core
│   │   ├── __init__.py
│   │   ├── corr.py
│   │   ├── datasets.py
│   │   ├── extractor.py
│   │   ├── raft.py
│   │   ├── update.py
│   │   └── util.py
│   └── raftnet.py
├── reference_model
│   └── model-000200000.pth
├── requirements.txt
├── saverloader.py
├── supervisely
│   ├── docker
│   │   └── Dockerfile
│   └── serve
│       ├── README.md
│       ├── config.json
│       ├── local.env
│       ├── model_settings.yaml
│       └── src
│           ├── main.py
│           └── sly_functions.py
├── test_on_badja.py
├── test_on_crohd.py
├── test_on_davis.py
├── test_on_flt.py
├── train.py
└── utils
    ├── basic.py
    ├── bremm.png
    ├── improc.py
    ├── misc.py
    ├── samp.py
    └── test.py

Explanation:

• supervisely/serve/src/main.py - main inference script

• supervisely/serve/src/sly_functions.py - functions to run the PIPs model based on the original repository code

• reference_model - directory with model weights; will be created automatically in get_reference_model.sh

• demo_images - directory with demo frames for inference

• supervisely/serve/README.md - readme of your application, it is the main page of an application in Ecosystem with some images, videos, and how-to-use guides

• supervisely/serve/config.json - configuration of the Supervisely application, which defines the name and description of the app, its context menu, icon, poster, and running settings

• requirements.txt - all packages needed for debugging

• local.env - file with variables used for debugging

• supervisely/serve/docker - directory with the custom Dockerfile for this application and the script that builds it and publishes it to the docker registry

App configuration

App configuration is stored in config.json file. A detailed explanation of all possible fields is covered in this Configuration Tutorial. Let's check the config for our current app:

{
  "name": "PIPs object tracking",
  "type": "app",
  "version": "2.0.0",
  "poster": "https://user-images.githubusercontent.com/115161827/233959102-9c48949f-b353-4a4b-ab7d-c1da99dfd914.jpg",
  "icon_cover": true,
  "icon": "https://user-images.githubusercontent.com/115161827/233959116-9d0922c6-e6fc-4f3a-958d-430742533f3a.jpg",
  "categories": [
    "neural network",
    "videos",
    "detection & tracking",
    "serve"
  ],
  "description": "serve and use in videos annotator",
  "docker_image": "supervisely/pips:1.0.0",
  "entrypoint": "python -m uvicorn main:pips.app --app-dir ./supervisely/serve/src --host 0.0.0.0 --port 8000 --ws websockets",
  "port": 8000,
  "task_location": "application_sessions",
  "headless": true,
  "need_gpu": true,
  "restart_policy": "on_error",
  "session_tags": [
    "sly_video_tracking"
  ],
  "community_agent": false,
  "allowed_shapes": [
    "point",
    "polygon",
    "graph"
  ]
}

Here is the explanation for the fields:

• type - type of the module in Supervisely Ecosystem

• version - version of Supervisely App Engine. Just keep it by default

• name - the name of the application

• description - the description of the application

• categories - these tags are used to place the application in the correct category in Ecosystem.

• session_tags - these tags will be assigned to every running session of the application. They can be used by other apps to find and filter all running sessions

• "need_gpu": true - should be true if you want to use any cuda devices.

• "community_agent": false - this means that this app can not be run on the agents started by Supervisely team, so users have to connect their own computers and run the app only on their own agents. Only applicable in Community Edition. Enterprise customers use their private instances so they can ignore the current option

• docker_image - Docker container will be started from the defined Docker image, github repository will be downloaded and mounted inside the container.

• entrypoint - the command that starts our application in a container

• port - port inside the container

• "headless": true means that the app has no User Interface

• allowed_shapes - shapes can be tracked with this model. Сonversion of figures to a set of points and vice versa is implemented in the base class, so you can keep this field default.