mxnet系列教程之1-第一个例子

第一个例子当然是mnist的例子

假设已经成功安装了mxnet

例子的代码如下：

cd mxnet/example/image-classification
python train_mnist.py

这样就会运行下去

train_mnist.py的代码为

"""
Train mnist, see more explanation at http://mxnet.io/tutorials/python/mnist.html
"""
import os
import argparse
import logging
logging.basicConfig(level=logging.DEBUG)
from common import find_mxnet, fit
from common.util import download_file
import mxnet as mx
import numpy as np
import gzip, structdef read_data(label, image):"""download and read data into numpy"""base_url = 'http://yann.lecun.com/exdb/mnist/'with gzip.open(download_file(base_url+label, os.path.join('data',label))) as flbl:magic, num = struct.unpack(">II", flbl.read(8))label = np.fromstring(flbl.read(), dtype=np.int8)with gzip.open(download_file(base_url+image, os.path.join('data',image)), 'rb') as fimg:magic, num, rows, cols = struct.unpack(">IIII", fimg.read(16))image = np.fromstring(fimg.read(), dtype=np.uint8).reshape(len(label), rows, cols)return (label, image)def to4d(img):"""reshape to 4D arrays"""return img.reshape(img.shape[0], 1, 28, 28).astype(np.float32)/255def get_mnist_iter(args, kv):"""create data iterator with NDArrayIter"""(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')train = mx.io.NDArrayIter(to4d(train_img), train_lbl, args.batch_size, shuffle=True)val = mx.io.NDArrayIter(to4d(val_img), val_lbl, args.batch_size)return (train, val)if __name__ == '__main__':# parse argsparser = argparse.ArgumentParser(description="train mnist",formatter_class=argparse.ArgumentDefaultsHelpFormatter)parser.add_argument('--num-classes', type=int, default=10,help='the number of classes')parser.add_argument('--num-examples', type=int, default=60000,help='the number of training examples')fit.add_fit_args(parser)parser.set_defaults(# networknetwork        = 'mlp',# traingpus           = '0,1',batch_size      = 64,disp_batches = 100,num_epochs     = 20,lr             = .05,lr_step_epochs = '10',model_predix = './my')args = parser.parse_args()# load networkfrom importlib import import_modulenet = import_module('symbols.'+args.network)sym = net.get_symbol(**vars(args))# trainfit.fit(args, sym, get_mnist_iter)

net写在了symbol文件夹中，相当于caffe的prototxt文件

model_predix相当于caffe的保存模型前缀

打印的信息为Saved checkpoint to"./my-0001.params"

"""
a simple multilayer perceptron
"""
import mxnet as mxdef get_symbol(num_classes=10, **kwargs):data = mx.symbol.Variable('data')data = mx.sym.Flatten(data=data)fc1  = mx.symbol.FullyConnected(data = data, name='fc1', num_hidden=128)act1 = mx.symbol.Activation(data = fc1, name='relu1', act_type="relu")fc2  = mx.symbol.FullyConnected(data = act1, name = 'fc2', num_hidden = 64)act2 = mx.symbol.Activation(data = fc2, name='relu2', act_type="relu")fc3  = mx.symbol.FullyConnected(data = act2, name='fc3', num_hidden=num_classes)mlp  = mx.symbol.SoftmaxOutput(data = fc3, name = 'softmax')return mlp

fit.py里面写了运行的代码

import mxnet as mx
import logging
import os
import timedef _get_lr_scheduler(args, kv):if 'lr_factor' not in args or args.lr_factor >= 1:return (args.lr, None)epoch_size = args.num_examples / args.batch_sizeif 'dist' in args.kv_store:epoch_size /= kv.num_workersbegin_epoch = args.load_epoch if args.load_epoch else 0step_epochs = [int(l) for l in args.lr_step_epochs.split(',')]lr = args.lrfor s in step_epochs:if begin_epoch >= s:lr *= args.lr_factorif lr != args.lr:logging.info('Adjust learning rate to %e for epoch %d' %(lr, begin_epoch))steps = [epoch_size * (x-begin_epoch) for x in step_epochs if x-begin_epoch > 0]return (lr, mx.lr_scheduler.MultiFactorScheduler(step=steps, factor=args.lr_factor))def _load_model(args, rank=0):if 'load_epoch' not in args or args.load_epoch is None:return (None, None, None)assert args.model_prefix is not Nonemodel_prefix = args.model_prefixif rank > 0 and os.path.exists("%s-%d-symbol.json" % (model_prefix, rank)):model_prefix += "-%d" % (rank)sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, args.load_epoch)logging.info('Loaded model %s_%04d.params', model_prefix, args.load_epoch)return (sym, arg_params, aux_params)def _save_model(args, rank=0):if args.model_prefix is None:return Nonedst_dir = os.path.dirname(args.model_prefix)if not os.path.isdir(dst_dir):os.mkdir(dst_dir)return mx.callback.do_checkpoint(args.model_prefix if rank == 0 else "%s-%d" % (args.model_prefix, rank))def add_fit_args(parser):"""parser : argparse.ArgumentParserreturn a parser added with args required by fit"""train = parser.add_argument_group('Training', 'model training')train.add_argument('--network', type=str,help='the neural network to use')train.add_argument('--num-layers', type=int,help='number of layers in the neural network, required by some networks such as resnet')train.add_argument('--gpus', type=str,help='list of gpus to run, e.g. 0 or 0,2,5. empty means using cpu')train.add_argument('--kv-store', type=str, default='device',help='key-value store type')train.add_argument('--num-epochs', type=int, default=100,help='max num of epochs')train.add_argument('--lr', type=float, default=0.1,help='initial learning rate')train.add_argument('--lr-factor', type=float, default=0.1,help='the ratio to reduce lr on each step')train.add_argument('--lr-step-epochs', type=str,help='the epochs to reduce the lr, e.g. 30,60')train.add_argument('--optimizer', type=str, default='sgd',help='the optimizer type')train.add_argument('--mom', type=float, default=0.9,help='momentum for sgd')train.add_argument('--wd', type=float, default=0.0001,help='weight decay for sgd')train.add_argument('--batch-size', type=int, default=128,help='the batch size')train.add_argument('--disp-batches', type=int, default=20,help='show progress for every n batches')train.add_argument('--model-prefix', type=str,help='model prefix')parser.add_argument('--monitor', dest='monitor', type=int, default=0,help='log network parameters every N iters if larger than 0')train.add_argument('--load-epoch', type=int,help='load the model on an epoch using the model-load-prefix')train.add_argument('--top-k', type=int, default=0,help='report the top-k accuracy. 0 means no report.')train.add_argument('--test-io', type=int, default=0,help='1 means test reading speed without training')return traindef fit(args, network, data_loader, **kwargs):"""train a modelargs : argparse returnsnetwork : the symbol definition of the nerual networkdata_loader : function that returns the train and val data iterators"""# kvstorekv = mx.kvstore.create(args.kv_store)# logginghead = '%(asctime)-15s Node[' + str(kv.rank) + '] %(message)s'logging.basicConfig(level=logging.DEBUG, format=head)logging.info('start with arguments %s', args)# data iterators(train, val) = data_loader(args, kv)if args.test_io:tic = time.time()for i, batch in enumerate(train):for j in batch.data:j.wait_to_read()if (i+1) % args.disp_batches == 0:logging.info('Batch [%d]\tSpeed: %.2f samples/sec' % (i, args.disp_batches*args.batch_size/(time.time()-tic)))tic = time.time()return# load modelif 'arg_params' in kwargs and 'aux_params' in kwargs:arg_params = kwargs['arg_params']aux_params = kwargs['aux_params']else:sym, arg_params, aux_params = _load_model(args, kv.rank)if sym is not None:assert sym.tojson() == network.tojson()# save modelcheckpoint = _save_model(args, kv.rank)# devices for trainingdevs = mx.cpu() if args.gpus is None or args.gpus is '' else [mx.gpu(int(i)) for i in args.gpus.split(',')]# learning ratelr, lr_scheduler = _get_lr_scheduler(args, kv)# create modelmodel = mx.mod.Module(context       = devs,symbol        = network)lr_scheduler  = lr_scheduleroptimizer_params = {'learning_rate': lr,'momentum' : args.mom,'wd' : args.wd,'lr_scheduler': lr_scheduler}monitor = mx.mon.Monitor(args.monitor, pattern=".*") if args.monitor > 0 else Noneinitializer   = mx.init.Xavier(rnd_type='gaussian', factor_type="in", magnitude=2)# initializer   = mx.init.Xavier(factor_type="in", magnitude=2.34),# evaluation metriceseval_metrics = ['accuracy']if args.top_k > 0:eval_metrics.append(mx.metric.create('top_k_accuracy', top_k=args.top_k))# callbacks that run after each batchbatch_end_callbacks = [mx.callback.Speedometer(args.batch_size, args.disp_batches)]if 'batch_end_callback' in kwargs:cbs = kwargs['batch_end_callback']batch_end_callbacks += cbs if isinstance(cbs, list) else [cbs]# runmodel.fit(train,begin_epoch        = args.load_epoch if args.load_epoch else 0,num_epoch          = args.num_epochs,eval_data          = val,eval_metric        = eval_metrics,kvstore            = kv,optimizer          = args.optimizer,optimizer_params   = optimizer_params,initializer        = initializer,arg_params         = arg_params,aux_params         = aux_params,batch_end_callback = batch_end_callbacks,epoch_end_callback = checkpoint,allow_missing      = True,monitor            = monitor)