def main():
ctx = mx.cpu()
train_data, test_data = utils.load_data_mnist(batch_size=64)
net = nn.HybridSequential()
with net.name_scope():
net.add(Convolution_int(acti_bit=8, channels=64, kernel_size=(3,3), strides=(1,1), padding=(1,1), in_channels=1, use_bias=False),
Activation_int(8, 'relu'),
Convolution_int(acti_bit=8, channels=128, kernel_size=(3,3), strides=(1,1), padding=(1,1), in_channels=64, use_bias=False),
Activation_int(8, 'relu'),
Dense(units=10, acti_bit=8, in_units=100352),)
net.load_params('./mnist_quantize_bias_dense_conv.params', allow_missing=True, ignore_extra=True, ctx=ctx)
loss = gluon.loss.SoftmaxCrossEntropyLoss()
batch_size = 64
trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate': 0.001})
#utils.train(train_data, test_data, net, loss, trainer, ctx=ctx, num_epochs=10)
#如果不需要转为sym-module，将下一行注释掉即可。
net.hybridize()
#test_acc = utils.evaluate_accuracy(test_data, net, ctx)
#net.save_params('./mnist_quantize_bias_dense_conv.params')
#print('test acc : ', test_acc)
#net.export('mnist_quantize')
#net.save_params('./mnist_quantize.params')
# added
sym, params = nnvm.frontend.from_mxnet(net)
graph, lib, params = nnvm.compiler.build(sym, 'llvm', shape={'data': (1, 784)}, params=params)
remote = rpc.LocalSession()
remote_ctx = remote.gpu(0)
module = rumtime.create(graph, lib, ctx=remote_ctx)
module.set_input('data', train_data[0])
module.run()
output = moudule.get_output(0, tvm.nd.empty(1,), dtype='float32')
print(output)
def main():
ctx = mx.cpu()
train_data, test_data = utils.load_data_mnist(batch_size=64)
net = nn.HybridSequential()
with net.name_scope():
net.add(Dense(units=10, acti_bit=8, in_units=100352), )
net.initialize()
batch_size = 64
#utils.train(train_data, test_data, net, loss, trainer, ctx=ctx, num_epochs=10)
#如果不需要转为sym-module，将下一行注释掉即可。
net.hybridize()
#test_acc = utils.evaluate_accuracy(test_data, net, ctx)
#net.save_params('./mnist_quantize_bias_dense_conv.params')
#print('test acc : ', test_acc)
#net.export('mnist_quantize')
#net.save_params('./mnist_quantize.params')
print(net(mx.symbol.Variable('data')).tojson())
# added
sys, params = nnvm.frontend.from_mxnet(net)
def test_mnist(learning_rate=0.01,
L1_reg=0.00,
L2_reg=0.0001,
n_epochs=100,
batch_size=128,
n_hidden=500,
n_hiddenLayers=3,
normalization=True,
eps=1e-4,
verbose=False,
smaller_set=True,
loss='norm',
lr_decay=False,
binary=True):
"""
Wrapper function for training and testing MLP
:type learning_rate: float
:param learning_rate: learning rate used (factor for the stochastic
gradient.
:type L1_reg: float
:param L1_reg: L1-norm's weight when added to the cost (see
regularization).
:type L2_reg: float
:param L2_reg: L2-norm's weight when added to the cost (see
regularization).
:type n_epochs: int
:param n_epochs: maximal number of epochs to run the optimizer.
:type batch_size: int
:param batch_szie: number of examples in minibatch.
:type n_hidden: int or list of ints
:param n_hidden: number of hidden units. If a list, it specifies the
number of units in each hidden layers, and its length should equal to
n_hiddenLayers.
:type n_hiddenLayers: int
:param n_hiddenLayers: number of hidden layers.
:type verbose: boolean
:param verbose: to print out epoch summary or not to.
:type smaller_set: boolean
:param smaller_set: to use the smaller dataset or not to.
:type loss: string
:param loss: to use hinge loss or normal loss.
:type lr_decay: boolean
:param lr_decay: to use learning_rate decay
:type binary: boolean
:param binary: to binarize the output
:type normalization: boolean
:param normalization: normalization output or not
:type eps: float
:param eps: normalization variable
"""
# load the dataset; download the dataset if it is not present
datasets = load_data_mnist(theano_shared=True)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
train_data_y_mat = datasets[3]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] // batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] // batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] // batch_size
######################
# BUILD ACTUAL MODEL #
######################
print('... building the model')
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
y_mat = T.matrix('y_mat')
epoch = T.lscalar('epoch')
rng = numpy.random.RandomState(1234)
# construct a neural network, either MLP or CNN.
classifier = myMLP(rng=rng,
input=x,
n_in=28 * 28,
n_hidden=n_hidden,
n_hiddenLayers=n_hiddenLayers,
n_out=10,
binary=binary,
normalization=normalization,
eps=eps)
# the cost we minimize during training is the negative log likelihood of
# the model plus the regularization terms (L1 and L2); cost is expressed
# here symbolically
# Loss can chosen as hinge loss or nll loss
if loss == 'norm':
cost = (classifier.negative_log_likelihood(y) +
L1_reg * classifier.L1 + L2_reg * classifier.L2_sqr)
else:
cost = classifier.logRegressionLayer.hinge(y_mat)
# compiling a Theano function that computes the mistakes that are made
# by the model on a minibatch
test_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size],
})
validate_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size],
})
# compute the gradient of cost with respect to theta (sotred in params)
# the resulting gradients will be stored in a list gparams
# According to paper gradient is calculated using binarized weights as
# same weights are used during forward propagation
if binary:
gparams = [T.grad(cost, param) for param in classifier.params_bin]
else:
gparams = [T.grad(cost, param) for param in classifier.params]
# specify how to update the parameters of the model as a list of
# (variable, update expression) pairs
# change learning rate depending upon input
# we are following exponential decay with
# learning rate = learning_rate_start * (learning_rate_final / learning_rate_start) ** (epoch/ n_epochs)
if lr_decay:
lr_start = learning_rate
lr_final = 1e-6
updates = [
(param_i,
T.cast(
T.clip(
param_i - (lr_start *
(lr_final / lr_start)**(epoch / 25)) * grad_i,
-1, 1), theano.config.floatX))
for param_i, grad_i in zip(classifier.params, gparams)
]
else:
updates = [(param_i,
T.cast(T.clip(param_i - learning_rate * grad_i, -1, 1),
theano.config.floatX))
for param_i, grad_i in zip(classifier.params, gparams)]
# compiling a Theano function `train_model` that returns the cost, but
# in the same time updates the parameter of the model based on the rules
# defined in `updates`
# Dependin upon lr_decay and loss function calculation we need to pass different
# parameters to training model
if loss == 'norm':
if lr_decay:
train_model = theano.function(
inputs=[index, epoch],
outputs=cost,
updates=updates,
givens={
x:
train_set_x[index * batch_size:(index + 1) * batch_size],
y:
train_set_y[index * batch_size:(index + 1) * batch_size],
})
else:
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x:
train_set_x[index * batch_size:(index + 1) * batch_size],
y:
train_set_y[index * batch_size:(index + 1) * batch_size],
})
else:
if lr_decay:
train_model = theano.function(
inputs=[index, epoch],
outputs=cost,
updates=updates,
givens={
x:
train_set_x[index * batch_size:(index + 1) * batch_size],
y_mat:
train_set_y[index * batch_size:(index + 1) * batch_size],
})
else:
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x:
train_set_x[index * batch_size:(index + 1) * batch_size],
y_mat:
train_set_y[index * batch_size:(index + 1) * batch_size],
})
###############
# TRAIN MODEL #
###############
print('... training')
result = train_nn(train_model, validate_model, test_model, n_train_batches,
n_valid_batches, n_test_batches, n_epochs, verbose,
lr_decay)
# plot_graph(result[2])
return result
elif FLAGS.exp == 'celeba_waemmd':
confs = config.conf_celeba_wae_mmd
model = WAE.WAE_MMD(confs)
optimizer = torch.optim.Adam(model.myparameters,
lr=confs['lr'],
betas=(confs['B1'], confs['B2']))
scheduler1 = MultiStepLR(optimizer,
milestones=confs['milestones1'],
gamma=0.5)
scheduler2 = MultiStepLR(optimizer,
milestones=confs['milestones2'],
gamma=0.2)
if confs['dataset'] == 'MNIST':
trainloader, testloader = load_data_mnist(test=True)
elif confs['dataset'] == 'celeba':
trainloader, testloader = load_data_celeba()
if confs['CUDA']:
model.cuda()
#if confs['pretrain']:
# optimizer_p = torch.optim.Adam(model.encoder.parameters(),lr = confs['lr'], betas=(confs['B1_disc'],confs['B2_disc']))
# pretrain(model,trainloader,optimizer_p,confs)
train_losses = []
train_loss_data = {'total': [], 'recon': [], 'match': []}
test_loss_data = {'total': [], 'recon': [], 'match': []}
test_losses = []
return nd.mean(nd.sum(L, 1))
if __name__ == "__main__":
# setting the hyper parameters
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size', default=16, type=int)
parser.add_argument('--epochs', default=1, type=int)
parser.add_argument('--train', default=False, type=bool)
args = parser.parse_args()
print(args)
ctx = utils.try_gpu()
# ctx = mx.cpu()
train_data, test_data = utils.load_data_mnist(batch_size=args.batch_size,resize=28)
net = CapsNet(batch_size=args.batch_size,ctx=ctx)
margin_loss = CapsuleMarginLoss()
print('====================================net====================================')
print(net)
if args.train:
print('====================================train====================================')
trainer = Trainer(net.collect_params(),'adam', {'learning_rate': 0.01})
utils.train(train_data, test_data, net, margin_loss, trainer, ctx, num_epochs=args.epochs)
sys.path.append('./dependencies')
import utils
ctx = utils.try_gpu()
import VAE as vaemodule
vae = vaemodule.VAE()
filename1 = './params/vae.params.get'
#vae.load_params(filename1, ctx = ctx)
vae.collect_params()
print(vae)
vae.initialize(ctx=ctx)
batch_size = 128
total_epoch = 100
epoch_size = 500
train_data, test_data = utils.load_data_mnist(batch_size)
trainer = gluon.Trainer(vae.collect_params(), 'sgd', {'learning_rate': 0.01})
train_last_loss = 2.
train_curr_loss = 0.1
for epoch in range(total_epoch):
if abs(train_last_loss - train_curr_loss) / train_last_loss < 1e-3:
break
train_loss = 0.
tic = time.time()
num = 0
for data, label in train_data:
num += 1
# Control the number of training data in each epoch
if num > epoch_size:
# 手写字体MNIST 多层感知器Multilayer Percepton (MLP)识别
# 多层神经网络
import mxnet as mx
from mxnet import gluon, autograd, ndarray
import numpy as np
import sys
sys.path.append('..')
import utils #包含了自己定义的一些通用函数 如下载 载入数据集等
##########################################################
#### 准备输入数据 ###
#我们通过gluon的data.vision模块自动下载这个数据
batch_size = 256#每次训练 输入的图片数量
train_data, test_data = utils.load_data_mnist(batch_size)
'''
def transform(data, label):
return data.astype('float32')/255, label.astype('float32')
#下载数据
#mnist_train = gluon.data.vision.FashionMNIST(train=True, transform=transform)
#mnist_test = gluon.data.vision.FashionMNIST(train=False, transform=transform)
mnist_train = gluon.data.vision.MNIST(train=True, transform=transform)
mnist_test = gluon.data.vision.MNIST(train=False, transform=transform)
# 使用gluon.data.DataLoader载入训练数据和测试数据。
# 这个DataLoader是一个iterator对象类，非常适合处理规模较大的数据集。
train_data = gluon.data.DataLoader(mnist_train, batch_size=32, shuffle=True)
test_data = gluon.data.DataLoader(train_data, batch_size=32, shuffle=False)
'''
###########################################