boulanni / theano-hf Goto Github PK

running the example with i7 4710 + 32g ram + GTX970+4G
it takes 24 hours still at 'update 3/100'
is this expected?

ubgpu@ubgpu:/github/theano-hf$ sudo python hf_examples.py
[sudo] password for ubgpu:
Using gpu device 0: GeForce GTX 970
/usr/local/lib/python2.7/dist-packages/theano/scan_module/scan_perform_ext.py:133: RuntimeWarning: numpy.ndarray size changed, may indicate binary incompatibility
from scan_perform.scan_perform import *
update 1/100, cost= [ 0.67232865 0.4154 ] lambda=0.50000, [CG iter 91, phi=-0.12633, cost=0.52522] backtracked 87/91 validation= [ 0.51908076 0.236 ] *NEW BEST
update 2/100, cost= [ 0.52221048 0.2334 ] lambda=0.33333, [CG iter 114, phi=-0.02780, cost=0.49935] backtracked 113/114 validation= [ 0.46644688 0.205 ] *NEW BEST
update 3/100, cost= [ 0.47799164 0.204 ] lambda=0.33333, [CG iter 10, phi=-0.00858, cost=0.45163]
[CG iter 38, phi=-0.01680, cost=0.45178]
[CG iter 46, phi=-0.01754, cost=0.45028]^CInterrupted by user.
ubgpu@ubgpu:/github/theano-hf$

Hi! I forked theano-hf and made some minor changes... basically plotting and writing the training progress to file and a hackish safeguard to check if the cost function stays real valued (I'm doing large scale MLE, where that can happen).

Would it be a good idea to make theano-hf compatible with Pandas? I'm not sure it would offer much gains in speed etc. but Pandas is really nice and it would be simple to sample from a Pandas dataframe...

Anyway, very nice module!

It's still the state-of-the-art on multiple music datasets per http://arxiv.org/pdf/1312.6026.pdf.