测试了一下keras和mxnet的速度

这两个都很好用啊，适合我这样的入门小白

win10 64 cuda8.0 cudnn5.1 gtx1060

cnn mnist

import numpy
import os
import urllib
import gzip
import struct
def read_data(label_name, image_name):
    s=os.getenv('DATA')
    with gzip.open(os.getenv('DATA')+'\\MNIST\\'+label_name) as flbl:
        magic, num = struct.unpack(">II", flbl.read(8))
        label = numpy.fromstring(flbl.read(), dtype=numpy.int8)
    with gzip.open(os.getenv('DATA')+'\\MNIST\\'+image_name, 'rb') as fimg:
        magic, num, rows, cols = struct.unpack(">IIII", fimg.read(16))
        image = numpy.fromstring(fimg.read(), dtype=numpy.uint8).reshape(len(label), rows, cols)
    return (label, image)
(train_lbl, train_img) = read_data('train-labels-idx1-ubyte.gz', 'train-images-idx3-ubyte.gz')
(val_lbl, val_img) = read_data('t10k-labels-idx1-ubyte.gz','t10k-images-idx3-ubyte.gz')
def to4d(img):
    return img.reshape(img.shape[0], 1, 28, 28).astype(numpy.float32)/255
def repack_data(d):
    t = numpy.zeros((d.size, 10))
    for i in range(d.size):
        t[i][d[i]] = 1
    return t
train_img=to4d(train_img)
val_img=to4d(val_img)
batch_size = 100
num_epoch =5
#backend='mxnet'
backend='keras'
if backend=='keras':
    from keras.models import *
    from keras.layers import *
    from keras.optimizers import *
    model = Sequential()
    model.add(Convolution2D(64, 5, 5, input_shape=(1,28,28), init='uniform', activation='relu'))
    model.add(MaxPooling2D())
    model.add(Convolution2D(128, 5, 5, init='uniform', activation='relu'))
    model.add(MaxPooling2D())
    model.add(Flatten())
    model.add(Dense(1024, init='uniform', activation='relu'))
    model.add(Dense(1024, init='uniform', activation='relu'))
    model.add(Dense(10, init='uniform', activation='softmax'))
    model.summary()
    model.compile(loss='categorical_crossentropy', optimizer=adadelta(), metrics=['accuracy'])
    model.fit(train_img,repack_data(train_lbl),batch_size=batch_size,nb_epoch=num_epoch,validation_data=(val_img,repack_data(val_lbl)))
else:
    import mxnet
    train_iter = mxnet.io.NDArrayIter(train_img, train_lbl, batch_size, shuffle=True)
    val_iter = mxnet.io.NDArrayIter(val_img, val_lbl, batch_size)
    data = mxnet.symbol.Variable('data')
    conv1 = mxnet.sym.Convolution(data=data, kernel=(5, 5), num_filter=64)
    relu1 = mxnet.sym.Activation(data=conv1, act_type="relu")
    pool1 = mxnet.sym.Pooling(data=relu1, pool_type="max", kernel=(2, 2), stride=(2, 2))
    conv2 = mxnet.sym.Convolution(data=pool1, kernel=(5, 5), num_filter=128)
    relu2 = mxnet.sym.Activation(data=conv2, act_type="relu")
    pool2 = mxnet.sym.Pooling(data=relu2, pool_type="max", kernel=(2, 2), stride=(2, 2))
    flatten = mxnet.sym.Flatten(data=pool2)
    fc1 = mxnet.symbol.FullyConnected(data=flatten, num_hidden=1024)
    relu3 = mxnet.sym.Activation(data=fc1, act_type="relu")
    fc2 = mxnet.symbol.FullyConnected(data=relu3, num_hidden=1024)
    relu4 = mxnet.sym.Activation(data=fc2, act_type="relu")
    fc3 = mxnet.sym.FullyConnected(data=relu4, num_hidden=10)
    net = mxnet.sym.SoftmaxOutput(data=fc3, name='softmax')
    mxnet.viz.plot_network(symbol=net, shape= {"data" : (batch_size, 1, 28, 28)}).render('mxnet')
    model = mxnet.model.FeedForward(
        ctx=mxnet.gpu(0),  # use GPU 0 for training, others are same as before
        symbol=net,
        num_epoch=num_epoch,
        learning_rate=0.1,
        optimizer='AdaDelta',
        initializer=mxnet.initializer.Uniform())
    import logging
    logging.getLogger().setLevel(logging.DEBUG)
    model.fit(
        X=train_iter,
        eval_data=val_iter,
        batch_end_callback=mxnet.callback.Speedometer(batch_size, 200)
    )

____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
convolution2d_1 (Convolution2D)  (None, 64, 24, 24)    1664        convolution2d_input_1[0][0]      
____________________________________________________________________________________________________
maxpooling2d_1 (MaxPooling2D)    (None, 64, 12, 12)    0           convolution2d_1[0][0]            
____________________________________________________________________________________________________
convolution2d_2 (Convolution2D)  (None, 128, 8, 8)     204928      maxpooling2d_1[0][0]             
____________________________________________________________________________________________________
maxpooling2d_2 (MaxPooling2D)    (None, 128, 4, 4)     0           convolution2d_2[0][0]            
____________________________________________________________________________________________________
flatten_1 (Flatten)              (None, 2048)          0           maxpooling2d_2[0][0]             
____________________________________________________________________________________________________
dense_1 (Dense)                  (None, 1024)          2098176     flatten_1[0][0]                  
____________________________________________________________________________________________________
dense_2 (Dense)                  (None, 1024)          1049600     dense_1[0][0]                    
____________________________________________________________________________________________________
dense_3 (Dense)                  (None, 10)            10250       dense_2[0][0]                    
====================================================================================================
Total params: 3364618
____________________________________________________________________________________________________

keras+theano

Train on 60000 samples, validate on 10000 samples
Epoch 1/5
60000/60000 [==============================] - 7s - loss: 0.1975 - acc: 0.9379 - val_loss: 0.0450 - val_acc: 0.9856
Epoch 2/5
60000/60000 [==============================] - 7s - loss: 0.0449 - acc: 0.9857 - val_loss: 0.0351 - val_acc: 0.9891
Epoch 3/5
60000/60000 [==============================] - 7s - loss: 0.0303 - acc: 0.9907 - val_loss: 0.0248 - val_acc: 0.9921
Epoch 4/5
60000/60000 [==============================] - 7s - loss: 0.0207 - acc: 0.9932 - val_loss: 0.0257 - val_acc: 0.9920
Epoch 5/5
60000/60000 [==============================] - 7s - loss: 0.0151 - acc: 0.9954 - val_loss: 0.0232 - val_acc: 0.9929

mxnet

INFO:root:Start training with [gpu(0)]
INFO:root:Epoch[0] Batch [200] Speed: 2960.54 samples/sec Train-accuracy=0.845600
INFO:root:Epoch[0] Batch [400] Speed: 2878.78 samples/sec Train-accuracy=0.975150
INFO:root:Epoch[0] Batch [600] Speed: 2875.59 samples/sec Train-accuracy=0.980750
INFO:root:Epoch[0] Resetting Data Iterator
INFO:root:Epoch[0] Time cost=21.459
INFO:root:Epoch[0] Validation-accuracy=0.986700
INFO:root:Epoch[1] Batch [200] Speed: 2888.17 samples/sec Train-accuracy=0.985850
INFO:root:Epoch[1] Batch [400] Speed: 2867.33 samples/sec Train-accuracy=0.988150
INFO:root:Epoch[1] Batch [600] Speed: 2867.63 samples/sec Train-accuracy=0.990200
INFO:root:Epoch[1] Resetting Data Iterator
INFO:root:Epoch[1] Time cost=20.874
INFO:root:Epoch[1] Validation-accuracy=0.980700
INFO:root:Epoch[2] Batch [200] Speed: 2894.78 samples/sec Train-accuracy=0.992200
INFO:root:Epoch[2] Batch [400] Speed: 2876.13 samples/sec Train-accuracy=0.993150
INFO:root:Epoch[2] Batch [600] Speed: 2858.85 samples/sec Train-accuracy=0.994650
INFO:root:Epoch[2] Resetting Data Iterator
INFO:root:Epoch[2] Time cost=20.875
INFO:root:Epoch[2] Validation-accuracy=0.990300
INFO:root:Epoch[3] Batch [200] Speed: 2879.48 samples/sec Train-accuracy=0.994600
INFO:root:Epoch[3] Batch [400] Speed: 2859.86 samples/sec Train-accuracy=0.995800
INFO:root:Epoch[3] Batch [600] Speed: 2860.25 samples/sec Train-accuracy=0.995800
INFO:root:Epoch[3] Resetting Data Iterator
INFO:root:Epoch[3] Time cost=20.951
INFO:root:Epoch[3] Validation-accuracy=0.990300
INFO:root:Epoch[4] Batch [200] Speed: 2887.86 samples/sec Train-accuracy=0.995750
INFO:root:Epoch[4] Batch [400] Speed: 2865.84 samples/sec Train-accuracy=0.997100
INFO:root:Epoch[4] Batch [600] Speed: 2868.30 samples/sec Train-accuracy=0.997700
INFO:root:Epoch[4] Resetting Data Iterator
INFO:root:Epoch[4] Time cost=20.915
INFO:root:Epoch[4] Validation-accuracy=0.988300

keras的速度我挺满意的，基本上达到了同类卡应该有的效果，而且gpu经常100%

但是theano后端的编译速度好慢好慢好慢！

mxnet好慢啊，三倍时间啊！跑一个官方例子也比gtx980慢一倍，感觉是什么地方配置跪了

不过我发现mxnet训练的时候cpu一直是100，可能是这个原因。。。。

悲伤的故事