Ask HN: How do you collect data for training your ML models?

If you're building and ML models, what is your process for collecting data?

For Example:

1/ Do you have dedicated hardware pipelines that send images/video/audio/text to a database in real-time or do you collect data in batches and then offload later?

2/ Do you collect mobile sensor data (iOS/Android) or do you have custom hardware?

3/ Do you rely on open source data sets like ImageNet / Google Open Images / Kaggle / Fishnet or are you collecting images for a use case that those data sets don't solve?