Hello forum,
I’m trying to train a model for RAVE and have been following the instructions here
At the preprocessing stage I keep getting the following error after specifying my data directory.
rave preprocess --input_path /path/ --output_path /path/ --channels 1
0it [00:00, ?it/s](FileNotFoundError(2, ‘No such file or directory’),)
Please excuse my ignorance, I gather it may be something to do with permissions?
Many thanks for any advice,
Erik
did you replace “/path/” in the input_path argument with the actual location of your data?
Hi sorry for not making that clear; yes I’m using the correct path.
I had the same problem, and there were crickets last time I asked. I’ll be watching this.
It’s likely the problem may lie with the Python scripts, but we don’t really see any feedback from their execution, so…
Is it on Mac? Linux? Windows? “FileNotFoundError(2, ‘No such file or directory’)” suggests that the script did not find the specified folder ; does it work if for exemple the data is in the RAVE subfolder?
It’s a m3 MacBook.ok thanks for the tip I’ll try the RAVE subfolder.
What can happen also is that the files are too short ; for example, as preprocessing takes by default a minimum of 262144 samples, file must be at min 6s approximately.
Oh I see! That is very useful advice and might very well be the issue since most of my files are shorter than 6s. I couldn’t find any information on this. The tutorial mentioned ‘short’ and ‘long’ sounds but nothing explicit on the lengths and whether it matters if the files themselves are long or short. Is it a good idea to have longer files with many short sounds inside?
Btw I tried the RAVE subfolder but I got the same error. It doesn’t seem to me that the system has a problem finding the directory since I’m able to run cd and ls commands on it.
Thanks!
You can adjust the number of chunks for preprocessing with the --num_signal arguments, followed by the number of samples for each chunk. The actual chunk size is the double of the argument, to allow randomization during training : if the --num_signal is 131072 samples (the amount of samples entering RAVE during training), the actual chunk size retrieved from the signal is 2 * 131072 (with overlapping frames). Hence, the minimum audio size is 2 * 131072 = 262144.
/!\ if you reduce the amount with --num_signal, also specify the keyword argument when launching the training with --n_signal.
It is true that this should be specified somewhere, we will do a small update soon with slight updates and an FAQ.
Thanks that’s super helpful. Does it take longer to train with smaller chunks then? And does the synthesis behave differently if it’s trained on shorter chunks?
Convergence may be slower, and yes synthesis may behave a bit differently ; but nothing really sure here 
fair enough, I’ll give it a whirl and see