Scraping Pinterest for thousands of labelled images sounds like fun?
Welcome to my pinterest scraper DNN practice repository.
The idea is pretty vague at the moment, I’ve written a pinterest scraper to
scrape images from pinterest. Part of the scraper scrapes the recommended pins,
I intend to use these as labels in order to fine-tune a DNN and eventually extract
embeddings from the fine-tuned DNN.
Once I’ve got embeddings, running umap
will allow me to visualise the image dataset in 3-dimensions, which could be
pretty cool. I know style transfer is possible but I’ll need to look into it a
bit more, but I intend to play around with that also.
Finally, I can package this all up as a pipeline that can produce a
3d-visualisation and image search by providing user-queries for pinterest.
For example you could provide “dogs” and “cats” as labels to the pipeline.
We can then scrape 1000+ images per label to quickly run through and fine-tune
a DNN to finally produce embeddings and enable visualisation.
I’m sure I’ll think of more functionality once I’ve got the initial pipeline
setup. Enjoy!