Fatal Python error: (pygame parachute) Segmentation Fault, and Keras exiting without errors about composer HOT 3 CLOSED

Keras's load_model usually only works when the model was saved with the same version of Keras as the version loading it. But it should still print an error if something is wrong, not exit unexpectedly...

Sounds more like an issue with either pygame or keras than with the composer. For reference, here's what I'm using at the moment, it doesn't look too different:

numpy==1.14.5
Keras==2.2.0
pydot==1.2.3
Theano==0.9.0
PyAudio==0.2.11
pygame==1.9.3

from composer.

After a few more days of wrangling with it, I finally got it to work. I'll post some advice for what I used to get it going again.

Firstly, I'd recommend Miniconda for setup if you haven't setup python yet. As for the required classes, I'm providing a pip freeze this time since apparently pipreqs doesn't get classes that are required by classes called in scripts:

asn1crypto==0.24.0
backports-abc==0.5
backports.functools-lru-cache==1.5
certifi==2018.4.16
cffi==1.11.5
chardet==3.0.4
conda==4.5.9
cryptography==2.2.2
cycler==0.10.0
docopt==0.6.2
enum34==1.1.6
functools32==3.2.3.post2
futures==3.2.0
h5py==2.8.0
idna==2.6
ipaddress==1.0.22
Keras==2.2.0
Keras-Applications==1.0.2
Keras-Preprocessing==1.0.1
kiwisolver==1.0.1
mako==1.0.7
MarkupSafe==1.0
matplotlib==2.2.2
menuinst==1.4.14
mido==1.2.8
mkl-fft==1.0.5
numpy==1.14.5
pipreqs==0.4.9
PyAudio==0.2.11
pycosat==0.6.3
pycparser==2.18
pydot==1.2.3
pygame==1.9.3
pygpu==0.7.6
pyOpenSSL==18.0.0
pyparsing==2.2.0
PySocks==1.6.8
python-dateutil==2.7.3
pytz==2018.5
pywin32==223
PyYAML==3.13
requests==2.18.4
ruamel-yaml==0.15.37
scipy==1.1.0
singledispatch==3.4.0.3
six==1.11.0
Theano==1.0.2+2.gc449c8699
tornado==5.1
urllib3==1.22
win-inet-pton==1.0.1
wincertstore==0.2
yarg==0.1.9

Lastly, make sure you don't have any other CUDA versions on the system (or at the very least, cudnn). I was initially confused by the gpu.theanorc file, and thought I'd need 8.0 for this and 9.1 for deep doodle, but 9.1 will work fine for this (you'll just need to shift the directory in the gpu file).

For those curious, I have a GTX 970M as my graphics card, and I think due to it possibly being older, ran into 'out of memory' problems if the computer hasn't restarted. I've seemed to fix them by adding this line in the gpu file:

[gpuarray]

preallocate = 2400

Try and get the number close to what it is when it works naturally (for me, 2.4GB, which translates to 2400MB).

The only error I've seen with this is when holding down left click on the command window (Windows 10+ I think), it'll bring up this error, which probably has to do with that stalling command line applications:

700/942 [=====================>........] - ETA: 0s - loss: 0.0091D:\python\Miniconda2\lib\site-packages\keras\callbacks.py:120: UserWarning: Method on_batch_end() is slow compared to the batch update (8.145000). Check your callbacks.

Also, slightly off topic, but how many songs did you have for your network?

from composer.

I have about ~4500 unique songs. With transposition it's double that.

from composer.