Not sure if this is the correct forum, but I have been working with a large (non-image) dataset that will eventually be used to train a neural network. I have been puzzling over how to manage wide data sets. For this application "wide" is maybe 10,000 or 20,000 points wide. It is not really possible to store this as a row in a conventional RDMS (which are usually limited to several hundred columns). Is it better to use a huge CSV file or maybe a no-sql technology like Cassandra (the data is originally in JSON format)?
Training with images makes use of pictures that are very large, like 7360×4912 (36 mega pixels). For image recognition at any angle and size, an image is rotated (about 60 times) and resized (depends on the original size, possibly 20 times) many times. This means you end up with a ton of data. In my example here, you would get some 60 × 36 × 2 = 4Gb of data for that one image. Although you really only save the data of interest (noise in the light signal) so in the end the AI does not use as much, but it would still be a quite large data set... especially when you know that a well trained image recognition system uses a good 2,000 images.
So 10 to 20 thousands data points isn't much at all. I guess it depends how many of those you have (one per second would be 86,400 a day...)
I would consider using a system like Cassandra because the database can grow as big as you'd like, if that is one of the problem you're facing. Also you can save the data in binary and the number of columns can vary on each row. However, Cassandra expects a key for each row. This is an important point as it makes use of that key to distribute the data over multiple computers (for faster access later and distribution of the data set over all computers.) It also includes replication "for free" so if a computer fails, you can just replace it (no need to restore it).
What is not a very good idea with Cassandra is creating many indexes. Those use rows (i.e. one index = one row) and thus the few computers handling that one row will work extra-hard on queries on that index... But I would imagine that for AI you would not really need such.
A really large CSV will be difficult to work with because you'll be responsible for accessing the rows (unless you have to access all the rows each time anyway?) Also the file format is Text and the file size will be limited by your OS (which should be really large, but be careful as many I/O calls in higher level libraries miss the fact that the size could be 64 bits.)
If you are looking at a longer/more permanent solution, I would definitely look into Cassandra, especially if you're using Java as a language (also I use it with C++ and it works just fine with C++ too.) If you're just writing a throw away app. anyway, then CSV is probably much easier than handling a whole Cassandra cluster! That's a learning curve...