From a first check at least several thousands of the image URLs are not accessible .
Since at least several thousands of the image URLs are not accessible it causes any loading and embedding script to hang out for several seconds each time - this makes the embedding process of a couple of million vectors dataset an unacceptable time.
Any idea/solution for that issue?
change the http timeout values.
change the http timeout values.
Of course, I had already done that - however, it is not always helpful for connection issues such as "Name or service not known" etc'.
maybe just grep for the domain names, resolve all of them at initialization, collect those that do not resolve, and add them to a blacklist.
maybe just grep for the domain names, resolve all of them at initialization, collect those that do not resolve, and add them to a blacklist.
It's possible, yes. Although it also consumes time on the fly or before that.
In any case, it seems odd to me to make hardware efforts to use a GPU to accelerate the embedding process whereas most of the time is lost during so many URL download connection timeouts and hanging.
That's only if you are designing your system naively, and if you setup a queue, and async / multiprocessing scraping, then the issues would likely largely cease to exist, as you will have a retrieval rate in excess of the consumption rate.
The download process hanging on dead links shouldn't be a major issue if you've setup a reasonable pipeline with queues, multiprocessing, etc and tuned it for current dead link %. It's expected at this point for there to be a lot of link rot, probably ~20% of the dataset will be gone at this point, possibly more.
You also need a non-typical DNS resolution setup to download this at max speed. You'll overload a typical resolver. See https://github.com/rom1504/img2dataset#setting-up-a-high-performance-dns-resolver ... i usually use 3-4 knot instances on the download machine. Do be careful though, messing around with the resolver can impact some network services.