One of the main bottlenecks of data labeling is managing multiple data annotation projects at scale. With various data sources and data pipelines, ML teams are spending valuable time on project management with restrictive, subpar tools that slow down innovation. 

The hard way - For example, consider a use case that involves simple steps like image classification, object detection, image segregation for instances, and lastly, full-pixel segmentation. To implement this simple workflow, firstly, you will require a dedicated engineering team to build the workflow management infrastructure. You will also need efficient systems to generate tasks, logically allocate tasks to annotators, route outputs from one step to further steps, and tools for quality check at every step. This project will most probably involve two to three months of dedicated engineering effort from a team of at least four/five engineers. This project will also come with overhead maintenance costs. Any iterations will require significant engineering effort due to the complexities of the workflow systems. The building cost is not justifiable, and needless to say, the platform will not be scalable in the long run.  

The easy way - To ease the load of additional engineering efforts and costs, Playment has engineered a built-in no-code Workflow Builder in GT Studio to help ML teams organise, configure, and optimise their workflows seamlessly. We have observed hundreds of workflows and have automated standard flows where anyone and configure workflows in a few minutes. Our no-code platform eliminates the need for dedicated engineering effort and makes the process simpler and faster. 

Effective Pipeline Setup With Workflow Builder

The Workflow Builder helps ML teams play around with pre-configured labeling steps to build creative pipelines for their labeling requirements. Once you upload the data on GT Studio, you can use the Workflow Builder for all further actions like:

• Divide complex projects into a simple logical sequence. 
• Automatically direct the flow of data/annotation jobs into different steps.
• Select particular user groups for the steps. 
• Configure projects in a few clicks.
• Monitor overall project status via task updates. 
  ‍
  

‍

The platform allows teams to create complex, customised workflows for a wide range of labeling tasks. Here’s a quick overview of the different workflow steps you can use: 

• Start Step: This automates the routing of specific jobs created on the platform to other steps in just one click. 
   
• Maker Step: The jobs then automatically flows to assigned makers who can create annotations using our unique annotation tools on the labeling interface. All annotation specifications, including attributes, sub-attributes, rules, availability of tools and features, deadlines, etc. can be configured easily with this step.
  
  
• Editor Step: You can send the annotated data to the editor for review using this step. The editors accept, edit, or reject submissions, and the tasks get routed to respective steps in the workflow. 
  

• Python Code Step: This step provides a python environment to run python code on raw/annotated data. This step is useful while writing heuristics checks/rules, data format conversions, call any external APIs, data validation checks, etc.
  

• AI Step: This step allows you to configure Playment’s proprietary ML models for helping annotators create annotations faster. 
  

• Empty Step: This step can easily configure complex workflow routes. The default route allows data flow without any other conditional steps. The Boolean and String Equal routes enable you to set up custom conditions for different data flows in the pipeline. 
  

• Manual Step: This step enables teams to manually collect jobs and push it for other actions after the data is pushed through one or more steps in the workflow. 
  

• End Step: Usually, the last step in the pipeline. As one or more jobs reach the end step, you can perform further quality checks or export the annotated data in desired formats.

Advanced Tooling For Workflow Customisation 

There’s no one-size-fits-all approach when it comes to innovative ML initiatives. That’s why GT Studio’s workflow interface has a wide variety of functions, offering maximum flexibility to accommodate nuanced workflow and configuration logic setups. Here are a few essential customisation functions that help teams tailor the platform for their unique requirements. 

Python Code Compatibility

The platform’s python environment allows you to write custom python scripts or functions using the Python Code Step. This function can help in multiple ways. Here are a few examples:

• Data format conversions to relative coordinates.
• PNG mask extraction for semantic segmentation tasks.
• Export data as JSON supported in different internal or popular dataset formats.
• Saving data in cloud storage (AWS buckets, Google Cloud, or Azure)
• Using Python Code step to call external APIs 
• Upload pre-annotated data for use in tasks for faster annotations
• Call open-source APIs to fetch or predict annotations 
• Set up data validation checks
  ‍

Easy Conditional Data Query

The Empty Step allows easy task route configuration for conditional data queries. ML engineers often manually curate data for their problem statements. Conditional data query through string equal routes helps them automate the process of data curation for further labeling. 

For example, to train a self-driving car to differentiate between sunny and rainy weather conditions, previously an engineer would manually curate training datasets with sunny and rainy weather scenarios and isolating all other datasets. 

With the Workflow Builder, the engineer can first create a classification task using the Python code step and using the string equal route configuration to curate rainy, sunny, and snowy scenarios seamlessly in half the time.

Pre-Annotated Data + Human In The Loop Functions

Consider this workflow example: Input Data → Pre-label high-precision classes → Manual Correction Maker Task → Python-Automated Validations → Manual Sample Quality Check → Output (refer image)

‍
You can upload pre-annotated data and create specific jobs for improving model accuracies and data performance. The Workflow Builder’s easy data visualisation and curation via conditional data query make the workflow setup process completely frictionless. You can pre-label high-precision classes using ML proposals and other automation features. 

Create manual correction tasks using the Maker Step to fine-tune data as per different use case requirements or create a second review task using the Python Code Step to further improve the accuracy rates. You could then enable python-automated validations and use the Manual Step to configure a QC task for the pre-annotated data to weave human and machine functions suitable for your modelling process.

The output that surpasses pre-defined accuracy thresholds can be immediately exported via the customer dashboard. The rest are looped back to previous steps to improve accuracy rates. 

Enriched Workflows For Better Model Performance  

Playment’s Workflow Builder provides unlimited possibilities and is the perfect addition to an ML engineer’s toolkit for accelerated innovation. With the Workflow Builder, you can:

• Create and configure complex workflows with just a few clicks. 
• Better task visibility and faster execution via methodical project division into a logical sequence for better project execution. 
• Lower coding workloads with the drag and drop interface for quicker project configurations and setups. 
• Access advanced customisation options by writing custom scripts in the Python environment and integrating it with the workflow using the Python Code step.
• Easy data visualisation and curation via effective conditional data query and task routing to improve model performance. 
• Leverage proprietary ML models for increasing dataset production, annotation accuracy, and overall model performance. 
• Obtain high-quality output by setting accuracy thresholds and criteria at different levels of the pipeline using pre-configured steps.
  

We’re committed to accelerating the AI age by focusing our efforts on bringing data labeling closer to automation with advanced tech infrastructures that help companies scale their ML initiatives without any hassle. 
For more information, visit our website or signup for a live demo of our labeling platform.