MXNet学习～试用卷积～跑CIFAR-10

第一次用卷积，看的别人的模型跑的CIFAR-10，不过吐槽一下。。。我觉着我的965m加速之后比我的cpu算起来没快多少。。正确率64%的样子，没达到模型里说的75%，不知道问题出在哪里

import numpy as np

import os

import mxnet as mx

import logging

import cPickle

def unpickle(file):

    with open(file,'rb') as fo:

        dict = cPickle.load(fo)

    return np.array(dict['data']).reshape(10000,3072),np.array(dict['labels']).reshape(10000)

def to4d(img):

    return img.reshape(img.shape[0],3,32,32).astype(np.float32)/255

def fit(batch_num,model,val_iter,batch_size):

    (train_img, train_lbl) = unpickle('cifar-10/data_batch_'+str(batch_num))

    train_iter = mx.io.NDArrayIter(to4d(train_img), train_lbl, batch_size, shuffle=True)

    model.fit(

        X=train_iter,

        eval_data=val_iter,

        batch_end_callback=mx.callback.Speedometer(batch_size,200)

    )

(val_img, val_lbl) = unpickle('cifar-10/test_batch')

batch_size = 100

val_iter = mx.io.NDArrayIter(to4d(val_img),val_lbl,batch_size)

data = mx.sym.Variable('data')

cv1 = mx.sym.Convolution(data=data,name='cv1',num_filter=32,kernel=(3,3))

act1 = mx.sym.Activation(data=cv1,name='relu1',act_type='relu')

poing1 = mx.sym.Pooling(data=act1,name='poing1',kernel=(2,2),pool_type='max')

do1 = mx.sym.Dropout(data=poing1,name='do1',p=0.25)

cv2 = mx.sym.Convolution(data=do1,name='cv2',num_filter=32,kernel=(3,3))

act2 = mx.sym.Activation(data=cv2,name='relu2',act_type='relu')

poing2 = mx.sym.Pooling(data=act2,name='poing2',kernel=(2,2),pool_type='avg')

do2 = mx.sym.Dropout(data=poing2,name='do2',p=0.25)

cv3 = mx.sym.Convolution(data=do2,name='cv3',num_filter=64,kernel=(3,3))

act3 = mx.sym.Activation(data=cv3,name='relu3',act_type='relu')

poing3 = mx.sym.Pooling(data=act3,name='poing3',kernel=(2,2),pool_type='avg')

do3 = mx.sym.Dropout(data=poing3,name='do3',p=0.25)

data = mx.sym.Flatten(data=do3)

fc1 = mx.sym.FullyConnected(data=data,name='fc1',num_hidden=64)

act4 = mx.sym.Activation(data=fc1,name='relu4',act_type='relu')

do4 = mx.sym.Dropout(data=act4,name='do4',p=0.25)

fc2 = mx.sym.FullyConnected(data=do4,name='fc2',num_hidden=10)

mlp = mx.sym.SoftmaxOutput(data=fc2,name='softmax')

logging.getLogger().setLevel(logging.DEBUG)

model = mx.model.FeedForward(

    ctx=mx.gpu(0),

    symbol=mlp,

    num_epoch=10,

    learning_rate=0.1

)

for batch_num in range(1,6):

    fit(batch_num, model, val_iter, batch_size)