Inference API

Inference API

Introduction

There are two ways how you can infer your models:

• From Supervisely platform with the APPs like Apply NN to Images and Apply Classifier to Images.

• Right from your code.

In this tutorial, you'll learn how to infer deployed models from your code with the sly.nn.inference.Session class. This class is a convenient wrapper for a low-level API. It under the hood is just a communication with the serving app via requests.

Before starting you have to deploy your model with a Serving App (e.g. Serve YOLOv5)

Try with Colab:

Table of Contents:

• Inference API

  • Introduction

• Quick overview

  • List of all inference methods

    • Image inference methods:

    • Video inference methods:

• A Complete Tutorial

  • 1. Initialize sly.nn.inference.Session

  • 2. Get the model info

    • Session info

    • Model Meta. Classes and tags

    • Inference settings

    • Set the inference settings

  • 3. Image Inference

    • Inspecting the model prediction

    • Visualize model prediction

    • Upload prediction to the Supervisely platform

  • 4. Video Inference

    • Method 1. Inferring video with iterator

    • Method 2. Inferring video without iterator

• Advanced. Working with raw JSON output

Let's start with a quick example of how you can connect and make inference of your model!

Quick overview

(for detailed tutorial go to the next section)

Example usage: visualize prediction

import os
from dotenv import load_dotenv
import supervisely as sly


# Get your Serving App's task_id from the Supervisely platform
task_id = 27209

# init sly.Api
load_dotenv(os.path.expanduser("~/supervisely.env"))
api = sly.Api()

# Create Inference Session
session = sly.nn.inference.Session(api, task_id=task_id)

session.get_session_info()

{'app_name': 'Serve YOLOv5',
 'session_id': 27209,
 'model_files': '/sly-app-data/model/yolov5s.pt',
 'number_of_classes': 80,
 'sliding_window_support': 'advanced',
 'videos_support': True,
 'async_video_inference_support': True,
 'task type': 'object detection',
 'model_name': 'YOLOv5',
 'checkpoint_name': 'yolov5s',
 'pretrained_on_dataset': 'COCO train 2017',
 'device': 'cuda',
 'half': 'True',
 'input_size': 640}

# Inference image_id
image_id = 19386161
prediction = session.inference_image_id(image_id)  # prediction is a `sly.Annotation` object

# Download and load the image that was inferred
save_path = "demo_image.jpg"
api.image.download_path(image_id, path=save_path)
image_np = sly.image.read(save_path)

# Draw the annotation and save it to the disk
save_path_predicted = "demo_image_pred.jpg"
predicted_annotation.draw_pretty(bitmap=image_np, output_path=save_path_predicted, fill_rectangles=False, thickness=7)

# Show
from matplotlib import pyplot as plt
image_pred = sly.image.read(save_path_predicted)
plt.imshow(image_pred)
plt.axis('off')

Image with predictions of the YOLOv5 model

List of all inference methods

Image inference methods:

# Infer single image by local path
pred = session.inference_image_path("image_01.jpg")

# Infer batch of images by local paths
pred = session.inference_image_paths(["image_01.jpg", "image_02.jpg"])

# Infer image by ID
pred = session.inference_image_id(17551748)

# Infer batch of images by IDs
pred = session.inference_image_ids([17551748, 17551750])

# Infer image by url
url = "https://images.unsplash.com/photo-1674552791148-c756b0899dba?ixlib=rb-4.0.3&ixid=MnwxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8&auto=format&fit=crop&w=387&q=80"
pred = session.inference_image_url(url)

# Infer the entire Project by ID
predictions = session.inference_project_id(18635)

# Infer the entire Project by ID with iterator
for ann_info in session.inference_project_id_async(18635):
    print(ann_info)

Video inference methods:

from tqdm import tqdm

video_id = 18635803

# Infer video getting each frame as soon as it's ready
for frame_pred in tqdm(session.inference_video_id_async(video_id)):
    print(frame_pred)

# Infer video without iterator
pred = session.inference_video_id(video_id)

A Complete Tutorial

1. Initialize sly.nn.inference.Session

First serve the model you want (e.g. Serve YOLOv5) and copy the task_id from the App sessions section in the Supervisely platform:

Copy the Task ID here

Init your sly.Api:

(for more info see Basics of authentication tutorial)

import os
from dotenv import load_dotenv
import supervisely as sly

# init sly.API
load_dotenv(os.path.expanduser("~/supervisely.env"))
api = sly.Api()

Create an Inference Session, a connection to the model:

# Get your Serving App's task_id from the Supervisely platform
task_id = 27209

# create session
session = sly.nn.inference.Session(api, task_id=task_id)

(Optional) You can pass the inference settings in init:

# pass settings by dict
inference_settings = {
    "conf_thres": 0.45
}
session = sly.nn.inference.Session(api, task_id=task_id, inference_settings=inference_settings)

Or with a YAML file:

# pass settings by YAML
inference_settings_yaml = "settings.yml"
session = sly.nn.inference.Session(api, task_id=task_id, inference_settings=inference_settings_yaml)

2. Get the model info

Session info

Each app with a deployed model has its own unique task_id (or session_id which is the same), model_name, pretrained_dataset and other useful info that can be obtained with the get_session_info() method.

session.get_session_info()

{'app_name': 'Serve YOLOv5',
 'session_id': 27209,
 'model_files': '/sly-app-data/model/yolov5s.pt',
 'number_of_classes': 80,
 'sliding_window_support': 'advanced',
 'videos_support': True,
 'async_video_inference_support': True,
 'task type': 'object detection',
 'model_name': 'YOLOv5',
 'checkpoint_name': 'yolov5s',
 'pretrained_on_dataset': 'COCO train 2017',
 'device': 'cuda',
 'half': 'True',
 'input_size': 640}

Model Meta. Classes and tags

The model may be pretrained on various datasets, like a COCO, ImageNet or even your custom data. Datasets are different in classes/tags they have. Therefore each dataset has its own meta information called project_meta in Supervisely. The model also contains this information and it's called model_meta. You can get the model_meta with method get_model_meta():

model_meta = session.get_model_meta()
print("The first 10 classes of the model_meta:")
[cls.name for cls in model_meta.obj_classes][:10]

The first 10 classes of the model_meta:

['person',
 'bicycle',
 'car',
 'motorcycle',
 'airplane',
 'bus',
 'train',
 'truck',
 'boat',
 'traffic light']

The model_meta will be used later, when we will visualize model predictions.

Inference settings

Each model has its own inference settings, like a conf_thres, iou_thres and others. You can get the full list of supported settings with get_default_inference_settings():

default_settings = session.get_default_inference_settings()
default_settings

{'conf_thres': 0.25,
 'iou_thres': 0.45,
 'augment': False,
 'debug_visualization': False}

Set the inference settings

There are 3 ways to set the inference settings:

• update_inference_settings(**kwargs)

• set_inference_settings(dict)

• set_inference_settings(YAML)

Also you can pass it earlier at creating the Session.

a) Update only the parameters you need:

session.update_inference_settings(conf_thres=0.4, iou_thres=0.55)
session.inference_settings

Output:

{'conf_thres': 0.4, 'iou_thres': 0.55}

b) Set parameters with a dict:

settings = {
    "conf_thres": 0.25
}
session.set_inference_settings(settings)
session.inference_settings

Output:

{'conf_thres': 0.25}

c) Set parameters with a YAML file:

session.set_inference_settings("settings.yml")
session.inference_settings

Output:

{'conf_thres': 0.55, 'augment': False}

3. Image Inference

There are several ways how to infer an image:

• by Supervisely ID

• by local path

• by URL from the web

# Infer image by local path
pred = session.inference_image_path("image_01.jpg")

# Infer image by ID
pred = session.inference_image_id(image_id=17551748)

# Infer image by url
url = "https://images.unsplash.com/photo-1674552791148-c756b0899dba?ixlib=rb-4.0.3&ixid=MnwxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8&auto=format&fit=crop&w=387&q=80"
pred = session.inference_image_url(url)

And you can also infer a batch of images:

# Infer batch of images by local paths
pred = session.inference_image_paths(["image_01.jpg", "image_02.jpg"])

# Infer batch of images by IDs
pred = session.inference_image_ids([17551748, 17551750])

Inspecting the model prediction

The prediction is a sly.Annotation object. It contains all labels and tags for an image and can be uploaded directly to the Supervisely platform.

(see more in SDK reference)

image_id = 19386163
prediction = session.inference_image_id(image_id)
prediction.to_json()

{'annotation': {'description': '',
  'size': {'height': 800, 'width': 1200},
  'tags': [],
  'objects': [{'classTitle': 'umbrella',
    'description': '',
    'tags': [{'name': 'confidence', 'value': 0.85693359375}],
    'points': {'exterior': [[540, 363], [694, 468]], 'interior': []},
    'geometryType': 'rectangle',
    'shape': 'rectangle'},
   {'classTitle': 'car',
    'description': '',
    'tags': [{'name': 'confidence', 'value': 0.86376953125}],
    'points': {'exterior': [[724, 380], [1198, 708]], 'interior': []},
    'geometryType': 'rectangle',
    'shape': 'rectangle'},
   {'classTitle': 'person',
    'description': '',
    'tags': [{'name': 'confidence', 'value': 0.8740234375}],
    'points': {'exterior': [[562, 442], [661, 685]], 'interior': []},
    'geometryType': 'rectangle',
    'shape': 'rectangle'},
   {'classTitle': 'car',
    'description': '',
    'tags': [{'name': 'confidence', 'value': 0.89501953125}],
    'points': {'exterior': [[4, 408], [509, 699]], 'interior': []},
    'geometryType': 'rectangle',
    'shape': 'rectangle'}],
  'customBigData': {}},
 'data': {}}

Visualize model prediction

sly.Annotation has a draw_pretty() method for convenient visualization routines:

# Draw the annotation and save it to disk
save_path_predicted = "demo_image_pred.jpg"
prediction.draw_pretty(bitmap=image_np, output_path=save_path_predicted, fill_rectangles=False, thickness=7)

# Show
from matplotlib import pyplot as plt
image_pred = sly.image.read(save_path_predicted)
plt.imshow(image_pred)
plt.axis('off');

Image with predictions of the YOLOv5 model

Upload prediction to the Supervisely platform

Now you can upload the image with predictions to the Supervisely platform:

workspace_id = 662

# Create new project and dataset
project_info = api.project.create(workspace_id, "My model predictions", change_name_if_conflict=True)
dataset_info = api.dataset.create(project_info.id, "First dataset")

# Update project meta with model's classes
api.project.update_meta(project_info.id, model_meta)
api.project.pull_meta_ids(project_info.id, model_meta)

# Upload the image
image_name = os.path.basename(image_path)
img_info = api.image.upload_path(dataset_info.id, name=image_name, path=image_path)

# Upload model predictions to Supervisely
api.annotation.upload_ann(img_info.id, prediction)

Note: when you update a project_meta with api.project.update_meta() the server generates ids for the classes and tags that have pushed for the first time and you have to update the model_meta too for the further uploading a prediction. This is where api.project.pull_meta_ids() method is helpful. It assigns the ids directly to the model_meta object. Because of all predictions have a reference to the model_meta, without this step we can't upload the predictions to the platform as predictions' ProjectMeta will not have the ids.

Result on the Supervisely platform:

Result in the Supervisely Labeling Tool

4. Video Inference

Method 1. Inferring video with iterator

The video inference is simple too.

The first way is to infer the video with inference_video_id_async method. It returns an iterator, which can be useful in processing predictions frame by frame. As soon as the model done with a one frame it will be yielded by the iterator:

from tqdm import tqdm

video_id = 18635803

pred_frames = []
for frame_ann in tqdm(session.inference_video_id_async(video_id)):
    pred_frames.append(frame_ann)

There are some parameters can be passed to the video inference:

• start_frame_index: the first frame to start

• frames_count: total frames to infer

• frames_direction: video playback direction, either "forward" or "backward"

Getting more information about the inference process:

video_id = 18635803
video_info = api.video.get_info_by_id(video_id)

frame_iterator = session.inference_video_id_async(video_id)
total_frames = video_info.frames_count
for i, frame_ann in enumerate(frame_iterator):
    labels = frame_ann.labels
    predicted_classes = [x.obj_class.name for x in labels]
    print(f"Frame {i+1}/{total_frames} done. Predicted classes = {predicted_classes}")

{"message": "The video is preparing on the server, this may take a while...", "timestamp": "2023-02-13T16:10:03.827Z", "level": "info"}
{"message": "Inference has started:", "progress": {"current": 0, "total": 10}, "is_inferring": true, "cancel_inference": false, "result": null, "pending_results": [], "timestamp": "2023-02-13T16:10:13.014Z", "level": "info"}


Frame 1/10 done. Predicted classes = ['car']
Frame 2/10 done. Predicted classes = ['car', 'car']
Frame 3/10 done. Predicted classes = ['car', 'car']
Frame 4/10 done. Predicted classes = ['car', 'car']
Frame 5/10 done. Predicted classes = ['car', 'car']
Frame 6/10 done. Predicted classes = ['car', 'car']
Frame 7/10 done. Predicted classes = ['car', 'car']
Frame 8/10 done. Predicted classes = ['car', 'car']
Frame 9/10 done. Predicted classes = ['car']
Frame 10/10 done. Predicted classes = ['car']

Stop video inference

If you need to stop the inference, use session.stop_async_inference():

from tqdm import tqdm

video_id = 18635803

for i, frame_ann in enumerate(tqdm(session.inference_video_id_async(video_id))):
    if i == 2:
        session.stop_async_inference()

{"message": "The video is preparing on the server, this may take a while...", "timestamp": "2023-02-09T23:15:47.232Z", "level": "info"}
{"message": "Inference has started:", "progress": {"current": 0, "total": 10}, "is_inferring": true, "cancel_inference": false, "result": null, "pending_results": [], "timestamp": "2023-02-09T23:15:55.878Z", "level": "info"}
 20%|██        | 2/10 [00:03<00:13,  1.63s/it]{"message": "Inference will be stopped on the server", "timestamp": "2023-02-09T23:16:01.559Z", "level": "info"}
 30%|███       | 3/10 [00:05<00:13,  1.88s/it]

Method 2. Inferring video without iterator

If you don't need to iterate every frame, you can use the inference_video_id method:

video_id = 18635803

predictions_list = session.inference_video_id(
    video_id, start_frame_index=5, frames_count=15, frames_direction="forward"
)

Note: it is recommended to use this method for very small videos, because the code will wait until the whole video has been inferred and you even can't to track the progress.

5. Project Inference

Method 1. Inferring project with iterator

from tqdm import tqdm

project_id = 18635

pred_ann_infos = []
for ann_info in tqdm(session.inference_project_id_async(project_id)):
    pred_ann_infos.append(ann_info)

There is extra parameter that can be passed to the project inference:

• dest_project_id: destination project id. If not passed, iterator will return annotation infos. If dest_project_id is equal to project_id, iterator will upload annotations to images in the project. If it is different from project_id, iterator will copy images and upload annotations to the new project.

Method 2. Inferring project without iterator

If you don't need to iterate every image, you can use the inference_project_id method:

project_id = 18635

predictions_list = session.inference_project_id(project_id)

Advanced. Working with raw JSON output

SessionJSON

There is a sly.nn.inference.SessionJSON class which is useful when it needed to work with raw json outputs.

The class has all the same methods as Session, it just returns a raw JSONs.

The prediction is a dict with the following fields:

• "annotation": contains a predicted annotation, that can be easily converted to sly.Annotation.

• "data": additional metadata of the prediction. In most cases you won't need this.

session = sly.nn.inference.SessionJSON(api, task_id=task_id)

prediction_json = session.inference_image_path("img/image_01.jpg")
prediction_json

{'annotation': {'description': '',
  'size': {'height': 1600, 'width': 1280},
  'tags': [],
  'objects': [{'classTitle': 'sheep',
    'description': '',
    'tags': [{'name': 'confidence', 'value': 0.255615234375}],
    'points': {'exterior': [[308, 1049], [501, 1410]], 'interior': []},
    'geometryType': 'rectangle',
    'shape': 'rectangle'},
  {'classTitle': 'person',
    'description': '',
    'tags': [{'name': 'confidence', 'value': 0.869140625}],
    'points': {'exterior': [[764, 272], [1062, 1002]], 'interior': []},
    'geometryType': 'rectangle',
    'shape': 'rectangle'},
  {'classTitle': 'horse',
    'description': '',
    'tags': [{'name': 'confidence', 'value': 0.87109375}],
    'points': {'exterior': [[393, 412], [1274, 1435]], 'interior': []},
    'geometryType': 'rectangle',
    'shape': 'rectangle'}],
  'customBigData': {}},
'data': {}}