因为自己在上Coursera的Advanced Machine Learning,里面第四周的Assignment要用到PYMC3，然后这个似乎是基于theano后端的。然而cpu版TMD太慢了，跑个马尔科夫蒙特卡洛要10个小时，简直不能忍了。所以妥妥换gpu版。

为了不把环境搞坏，我在Anaconda里面新建了一个环境。(关于Anaconda，可以看我之前翻译的文章)

Conda Create -n theano-gpu python=3.4

（theano GPU版貌似不支持最新版，保险起见装了旧版）

conda install theano pygpu

这里面会涉及很多依赖，应该conda会给你搞好，缺什么的话自己按官方文档去装。

然后至于Cuda和Cudnn的安装，可以看我写的关于TF安装的教程

和TF不同的是，Theano不分gpu和cpu版，用哪个看配置文件设置，这一点是翻博客了解到的：
配置好Theano环境之后，只要 C:Users你的用户名 的路径下添加 .theanorc.txt 文件。

.theanorc.txt 文件内容:

[global]

openmp=False

device = cuda

floatX = float32

base_compiler = C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin

allow_input_downcast=True 

[lib]

cnmem = 0.75

[blas]

ldflags=

[gcc]

cxxflags=-IC:\Users\lyh\Anaconda2\MinGW

[nvcc]

fastmath = True

flags = -LC:\Users\lyh\Anaconda2\libs

compiler_bindir = C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin

flags =  -arch=sm_30

注意在新版本中，声明用gpu从device=gpu改为device=cuda

然后测试是否成功：

from theano import function,config,shared,tensor
import numpy
import time

vlen = 10 * 30 * 768  # 10 x #cores x # threads per core
iters = 1000

rng = numpy.random.RandomState(22)
x = shared(numpy.asarray(rng.rand(vlen),config.floatX))
f = function([],tensor.exp(x))
print(f.maker.fgraph.toposort())
t0 = time.time()
for i in range(iters):
    r = f()
t1 = time.time()
print("Looping %d times took %f seconds" % (iters,t1 - t0))
print("Result is %s" % (r,))
if numpy.any([isinstance(x.op,tensor.Elemwise) and
              ('Gpu' not in type(x.op).__name__)
              for x in f.maker.fgraph.toposort()]):
    print('Used the cpu')
else:
    print('Used the gpu')

输出：

[GpuElemwise{exp,no_inplace}(<GpuArrayType<None>(float32,vector)>),HostFromGpu(gpuarray)(GpuElemwise{exp,no_inplace}.0)]
Looping 1000 times took 0.377000 seconds
Result is [ 1.23178029  1.61879349  1.52278066 ...,2.20771813  2.29967761
  1.62323296]
Used the gpu

到这里就算配好了

然后在作业里面，显示Quadro卡启用

但是还是有个warning

WARNING (theano.tensor.blas): Using NumPy C-API based implementation for BLAS functions.

这个真不知道怎么处理

然后后面运行到：

with pm.Model() as logistic_model:
    # Since it is unlikely that the dependency between the age and salary is linear,we will include age squared
    # into features so that we can model dependency that favors certain ages.
    # Train Bayesian logistic regression model on the following features: sex,age,age^2,educ,hours
    # Use pm.sample to run MCMC to train this model.
    # To specify the particular sampler method (Metropolis-Hastings) to pm.sample,# use `pm.Metropolis`.
    # Train your model for 400 samples.
    # Save the output of pm.sample to a variable: this is the trace of the sampling procedure and will be used
    # to estimate the statistics of the posterior distribution.
    
    #### YOUR CODE HERE ####
    
    pm.glm.GLM.from_formula('income_more_50K ~  sex+age + age_square + educ + hours',data,family=pm.glm.families.Binomial())
    with logistic_model:
        trace = pm.sample(400,step=[pm.Metropolis()]) #nchains=1 works for gpu model
        
    ### END OF YOUR CODE ###

这里出现的报错：

GpuArrayException: cuMemcpyDtoHAsync(dst,src->ptr + srcoff,sz,ctx->mem_s): CUDA_ERROR_INVALID_VALUE: invalid argument

这个问题最后github大神解决了：
So njobs will spawn multiple chains to run in parallel. If the model uses the GPU there will be a conflict. We recently added nchains where you can still run multiple chains. So I think running pm.sample(niter,nchains=4,njobs=1) should give you what you want.
我把：

trace = pm.sample(400,step=[pm.Metropolis()]) #nchains=1 works for gpu model

加上nchains就好了，应该是并行方面的问题

trace = pm.sample(400,step=[pm.Metropolis()],nchains=1,njobs=1) #nchains=1 works for gpu model

另外

plot_traces(trace,burnin=200)

出现pm.df_summary报错，把pm.df_summary 换成 pm.summary就好了，也是github搜出来的。