Image classification

In this tutorial, you will learn how to run image classification in Toloka. We will use a project preset designed specifically for this type of data labeling.

Image classification is a type of data labeling task with a finite number of response options.

Use Toloka to classify a set of images into categories that you define. Tolokers look at the images and choose one of the given categories. After you collect all the labeled images, you can apply your dataset for model training.

Prerequisites

Before you begin:

Make sure you are registered in Toloka as a requester.
Top up your Toloka account. If you are unsure about the budget, you can do that later in this tutorial. Toloka will display the budget estimate for your project.

Create a project

We recommend starting with a project preset for easier configuration and better results.

Follow this link, or create a project manually:
1. Click Create a project.
2. Click Do it myself.
3. Select the Image classification preset.
Click Choose this preset in the pop-up tab.
In the General information section, add the project name and description:
- Name to show Tolokers: In 2–5 words, state the general idea of the project.
- Description for Tolokers: In a couple of sentences, explain what you expect Tolokers to do. This is just an overview. You will write instructions later.
In the Task interface section, set up what your tasks will look like. This preset has a task template with validation, keyboard shortcuts, and task layout pre-configured.

Note
This tutorial uses Template Builder, but you can use the HTML/JS/CSS editor for the same purpose.
1. Using the Visual editor, add your data to the Config section:
  - Question Tolokers will see in your task: Write a question that matches the responses Tolokers need to choose from. All tasks in a project use the same question.
  - Paste link to a sample image: This image is only used to display the task interface preview on the right.
2. Set answer options is pre-filled with sample answers. Replace the samples with your categories. Note that Other and Error are separate entities.
  - Answer option for Tolokers: This is the label that Tolokers will see. Make sure it is clear and correct.
  - Name in labeling results: This is the value you will see in the file with the labeling results.
3. Raw task data is stored in the XSLX, TSV, or JSON format. The labeling results are presented in a TSV file. The Data specification section determines which parameters these files might contain.
  
  Click Show specifications and check the values:
  - Input data: Parameters in the file with raw task data.
  - Output data: Parameters in the file with labeling results.
  Input data and Output data match the task interface you set up in Template Builder. Check that there are fields for all data types you use for your tasks, and for the ones you want to see in the results file.
In the Instructions for Tolokers editor, enter the instructions Tolokers will see when they start doing your tasks. You can add text, tables, and images to your instructions.

Check the sample text of the instructions, and update it to fit your project.

Tip
When writing instructions, remember that most Tolokers don’t know anything about your tasks beforehand. Make sure your instructions are as clear as possible, but not too wordy. For successful data labeling, try to strike a balance between covering all the essentials and keeping it short. Learn more in our knowledge base.
In the upper-right corner, click Save.

Learn more about working with the project in the Project section.

Create a pool

A pool is a set of tasks sent out to Tolokers at the same time. One project can have many pools. When creating a pool, you set up pricing, audience filters for Tolokers, and quality control.

Click Create new pool on the project page.
Select the value in the Pool type drop-down list.

Pool types

Note
If the price per task suite is zero, you must select the pool type.
Set the Pool name (visible only to you) field. Only you will see this pool name on the project page.
Specify the pool description which will be displayed instead of the project description in the task list for Tolokers. By default, Tolokers see the description from the project settings. To use a different description, uncheck the Use project description box and set Public description. If necessary, click + Private comment to add a private project description that only you will see.
Click Create.
At the Select the audience for your task step, set up filters to select Tolokers for your pool.
1. Clear My tasks may contain shocking or pornographic content if your project has none of those.
2. To select Tolokers based on their language, location, age, gender, and other parameters, click the Add filter button.
3. Use the Speed/quality balance slider to change the number of Tolokers who can see your tasks. Move the slider to the right to exclude Tolokers with lower ratings from participating in your project.
At the Setup quality control step, the Image classification preset has preconfigured quality control rules for more accurate results. In most cases, you can keep them as is:
1. In most cases, you can keep the preconfigured quality control rules as is:
  - Fast responses: This rule filters out Tolokers who complete tasks too fast. The default settings mean that Tolokers are banned from the project for 1 day if they complete tasks in 4 out of 5 task suites in less than 15 seconds.
  - Majority vote: This rule accepts the most popular response as the correct one, and allows you to filter out Tolokers who answer incorrectly. The default settings mean that Tolokers who give correct responses to less than 40% of tasks are banned from the project for 1 day. Accept as majority set to 2 means that 2 similar responses out of all responses given to a single task will be considered as the correct answer.
2. To filter out Tolokers who often make mistakes in the control tasks, click Add a quality control rule → Control tasks. Enter the following values:
  
  This means that if a Toloker completed more than three control tasks and gave incorrect answers in more than 60% of them, they will be blocked and won't be able to complete tasks on this project for 10 days.
  
  The rule will work if you specify the correct answers for the control tasks. You will do that later in this tutorial.
At the Set the task price and overlap step, set up how much a single task will cost for you.
1. In Price per task suite, $, set the amount of money to pay per task suite done by one Toloker. A task suite is a page with a number of tasks. It can contain one or several tasks. If the tasks are simple, you can add 10–20 tasks per suite.
2. In the Overlap field, define how many Tolokers must do each task.
  
  The default value (3) means that each task will have 3 responses.
At the Prepare and upload data step, upload your task data.
1. Attach a prepared dataset or media files.
  1. To download a template, click one of the buttons:
    - Template in XLSX
    - Template in TSV
    - Template in JSON
    For this type of project, the file with tasks must have one parameter. Its name equals INPUT:image, and the values are links to the images.
    INPUT:image https://tlk.s3.yandex.net/dataset/cats_vs_dogs/dogs/041adb571f4342e7a42e47712f492101.jpg https://tlk.s3.yandex.net/dataset/cats_vs_dogs/dogs/048e5760fc5a46faa434922b2447a527.jpg https://tlk.s3.yandex.net/dataset/cats_vs_dogs/dogs/05334365c060421ab25264166bbb4fd1.jpg
  2. Open the downloaded file, and replace the sample links with links to your images.
  3. Click Select prepared dataset, and upload the file you’ve just made.
2. Click Continue.
3. Tasks are shown to Tolokers in suites. A suite is a single page with multiple tasks. Define how many tasks to include per suite:
  - General tasks: These are tasks for Tolokers to label.
  - Control tasks: These are tasks with predefined answers used to control the quality of responses. You will create them in the next step.
  - Training tasks: These are tasks with predefined answers and explanations for Tolokers. Normally you use training tasks in separate training pools. You don’t have to include them.
  For example, you can add 9 general tasks and 1 control task per suite:
4. Click Combine tasks into suites.
Create control tasks at the Add control tasks for checking performance step. To do it, add correct answers to some of your tasks.
1. Check the result checkbox, and select the correct answer for a task. Then, click the Save and go to next button. Add several control tasks this way.
  
  Tip
  For large pools (over 1000 tasks), we recommend adding at least 1% of control tasks to the pool. For small pools (around 100 tasks), you need 10% control tasks.
2. Note the Distribution of correct responses for control tasks graph on the right side of the page. It shows how many control tasks of each type you have. We recommend adding an equal quantity of each correct response.
At the Double-check your project and try out tasks step, check how the task will look from the Toloker's point of view.

Note
This step will be enabled after you complete the previous steps. You can skip this step by clicking Do it later.
At the Add optional pool settings step, set up advanced pool settings.

Note
This step will be enabled after you complete the previous steps. You can skip this step by clicking Use default settings.

After all the steps, you'll see the Set up is finished and your pool is ready for labeling tip on the pool page.

Start labeling

Make sure you have topped up your account.
To send the tasks to Tolokers and begin the labeling process, click Start labeling.
In the pop-up panel, review the budget and click Launch.

See the results

You can see the labeling progress on the pool page. Wait until the labeling is completed.
When the labeling is complete, click the arrow next to the Download results button and choose Run Dawid-Skene model from the drop-down menu. Click Yes in the pop-up window.
Open the same drop-down menu again, and click View aggregations list.
Wait until the aggregation is complete, and click Download. You will get the TSV file with the labeling results:
- INPUT: The data you uploaded for labeling.
- OUTPUT: The results of labeling (category picked by Tolokers).
- CONFIDENCE: The response significance according to the Dawid-Skene model.