攒了台PC，准备玩GPU上的深度学习了，谁有经验，多指教

首先，应该吧显示器接在i5的核显接口上吧？
在bios里设置，首选cpu的核显，对吧？

这个显卡只能玩玩吧，跑些中小规模的网络，并且training时间还很长能跑大的网络，可以在上面做研究发paper的起码要titan x了。百度可以搜到很多theano和caffe的安装教程。

要不要学新出的mxnet

GTX960 一般是2G的内存，我还专门买了个4G的，
再高档的就太贵了！

sudo apt-get install build-essential
sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libboost-all-dev libhdf5-serial-dev
sudo apt-get install libatlas-base-dev
sudo apt-get install python-dev python-pip
sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev protobuf-compiler
sudo apt-get install python-numpy python-scipy python-matplotlib python-sklearn python-skimage python-h5py python-protobuf python-leveldb python-networkx python-nose python-pandas python-gflags Cython ipython
sudo apt-get install git
sudo pip install --upgrade --no-deps git+git://github.com/Theano/Theano.git
cd keras
sudo python setup.py install
或者
sudo pip install keras -i http://pypi.douban.com/simple
keras包安装在/usr/local/lib下
http://s3.amazonaws.com/img-datasets/mnist.pkl.gz
数据在用户的.keras目录下

成功啦！
yin@deepthink:~/keras/examples$ THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python mnist_cnn.py
Using Theano backend.
Using gpu device 0: GeForce GTX 960 (CNMeM is disabled)
X_train shape: (60000, 1, 28, 28)
60000 train samples
10000 test samples
Train on 60000 samples, validate on 10000 samples
Epoch 1/12
60000/60000 [==============================] - 18s - loss: 0.2593 - acc: 0.9202 - val_loss: 0.0641 - val_acc: 0.9788
Epoch 2/12
60000/60000 [==============================] - 18s - loss: 0.0932 - acc: 0.9728 - val_loss: 0.0389 - val_acc: 0.9871
Epoch 3/12
60000/60000 [==============================] - 18s - loss: 0.0708 - acc: 0.9790 - val_loss: 0.0419 - val_acc: 0.9850
Epoch 4/12
60000/60000 [==============================] - 18s - loss: 0.0564 - acc: 0.9824 - val_loss: 0.0414 - val_acc: 0.9872
Epoch 5/12
60000/60000 [==============================] - 18s - loss: 0.0500 - acc: 0.9844 - val_loss: 0.0323 - val_acc: 0.9902

10秒内，把mnist跑到99%以上
比以前快了40倍

gpu部分
在https://developer.nvidia.com/cuda-downloads 下载CUDA Toolkit 7.5 deb包
sudo dpkg -i cuda-repo-<distro>_<version>_<architecture>.deb
sudo apt-get update
sudo apt-get install cuda
export PATH=/usr/local/cuda-7.5/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib64:$LD_LIBRARY_PATH
sudo ldconfig /usr/local/cuda-7.0/lib64
THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python mnist_cnn.py

写错了，不收10秒，是100秒
100秒内，把mnist跑到99%以上
比以前快了40倍

鼓掌

看keras的example里面写的
Get to 99.25% test accuracy after 12 epochs (there is still a lot of margin for parameter tuning). 16 seconds per epoch on a GRID K520 GPU.
而我的18 seconds per epoch
是不是说明我的显卡和GRID K520 GPU差不多？

用在word2vec之类的算法效果如何?

mnist_cnn.py只用了几百兆的显存