Ask HN: What pre-trained models have you used in production?

14 pointscalebkaiser3 comments

I write about ML in production—working with models like OpenAI's GPT-2, Hugging Face's DistilBERT, AllenNLP's ELMo-BiDAF, etc—at cortex.dev. What other pre-trained models should I be looking at? Are there any you've used in production?

View on HN

Inceptionv3 and Resnet50 - for detecting cracks in displays and SKUs in a supermarket.

And a Flask webservice for serving.

I wished I could use Jupyter Serve; but Flask was easier for our ops team.

How did you choose those particular NN architectures? How much data did you use to finetune them (assuming they are pretrained on Imagenet)?

Resnet50 with transfer learning

Reply on news.ycombinator.com