Converting a theano model built on GPU to CPU?

问题:

I have some pickle files of deep learning models built on gpu. I'm trying to use them in production. But when i try to unpickle them on the server, i'm getting the following error.

Traceback (most recent call last):
File "score.py", line 30, in
model = (cPickle.load(file))
File "/usr/local/python2.7/lib/python2.7/site-packages/Theano-0.6.0-py2.7.egg/theano/sandbox/cuda/type.py", line 485, in CudaNdarray_unpickler
return cuda.CudaNdarray(npa)
AttributeError: ("'NoneType' object has no attribute 'CudaNdarray'", , (array([[ 0.011515 , 0.01171047, 0.10408644, ..., -0.0343636 ,
0.04944979, -0.06583775],
[-0.03771918, 0.080524 , -0.10609912, ..., 0.11019105,
-0.0570752 , 0.02100536],
[-0.03628891, -0.07109226, -0.00932018, ..., 0.04316209,
0.02817888, 0.05785328],
...,
[ 0.0703947 , -0.00172865, -0.05942701, ..., -0.00999349,
0.01624184, 0.09832744],
[-0.09029484, -0.11509365, -0.07193922, ..., 0.10658887,
0.17730837, 0.01104965],
[ 0.06659461, -0.02492988, 0.02271739, ..., -0.0646857 ,
0.03879852, 0.08779807]], dtype=float32),))

I checked for that cudaNdarray package in my local machine and it is not installed, but still i am able to unpickle them. But in the server, i am unable to. How do i make them to run on a server which doesnt have a GPU?

回答1:

There is a script in pylearn2 which may do what you need:

pylearn2/scripts/gpu_pkl_to_cpu_pkl.py

回答2:

The related Theano code is here.

From there, it looks like there is an option config.experimental.unpickle_gpu_on_cpu which you could set and which would make CudaNdarray_unpickler return the underlying raw Numpy array.

回答3:

This works for me. Note: this doesn't work unless the following environment variable is set: export THEANO_FLAGS='device=cpu'

import os
from pylearn2.utils import serial
import pylearn2.config.yaml_parse as yaml_parse

if __name__=="__main__":

_, in_path, out_path = sys.argv
os.environ['THEANO_FLAGS']="device=cpu"

model = serial.load(in_path)

model2 = yaml_parse.load(model.yaml_src)
model2.set_param_values(model.get_param_values())

serial.save(out_path, model2)

回答4:

I solved this problem by just saving the parameters W & b, but not the whole model. You can save the parameters use this:http://deeplearning.net/software/theano/tutorial/loading_and_saving.html?highlight=saving%20load#robust-serialization This can save the CudaNdarray to numpy array. Then you need to read the params by numpy.load(), and finally convert the numpy array to tensorSharedVariable use theano.shared().