search

pkmital / time-domain-neural-audio-style-transfer

NIPS2017 "Time Domain Neural Audio Style Transfer" code repository
Apache License 2.0
140 stars 32 forks source link

out of memory #3

Closed ahihi closed 4 years ago

ahihi commented 5 years ago

hello,

i am trying to run timedomain.py on a system with 16GB RAM and running out of memory in compute_features(). is there any simple way of running the computations in smaller batches? if not, how much RAM is required?

2019-09-11 13:38:51.235211: W tensorflow/core/framework/allocator.cc:107] Allocation of 1509949440 exceeds 10% of system memory.
LLVM ERROR: out of memory
Aborted (core dumped)
pkmital commented 4 years ago

Make sure you're not trying a super long audio file. Try with for example 3 seconds of audio only to being. If you are still running into issues, try reducing the parameters, e.g. the number of filters from 4096 to 512. You can also increase the hop_length to reduce the memory footprint. 16 GB should be a lot though! Feel free to reopen if any issues.