[
https://issues.apache.org/jira/browse/SINGA-248?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15516698#comment-15516698
]
wangwei commented on SINGA-248:
-------------------------------
Thanks for reporting this issue.
net.save() uses numpy and cPickle to dump the model parameters (about 500MB),
which use extra spaces.
Here is a simple experiment I did,
{code}
>>> import numpy as np
>>> import cPickle as pickle
>>> bigary = np.random.rand(20588 + 4096 + 1000, 4096)
>>> np.save('bigary', bigary)
>>> pickary = {}
>>> pickary['a'] = np.random.rand(20588, 4096)
>>> pickary['b'] = np.random.rand(4096, 4096)
>>> pickary['c'] = np.random.rand(1000, 4096)
>>> with open('pickary', 'wb') as fd:
... pickle.dump(pickary, fd)
...
>>> quit()
(pysinga)wangwei@slave2:~/vgg16$ ls -lh
total 4.5G
803M Sep 23 22:51 bigary.npy
1.5G Sep 23 22:39 model.bin
2.2K Sep 23 22:38 net.py
2.2G Sep 23 22:53 pickary
{code}
bigary.npy, and pickary have similar size in memory (i.e. the number of float
values), but the sizes of the disk files differ a lot.
One approach to reduce the checkpoint file size is to serialize the singa
tensors into protobuf objects and use classes in io/ to dump them.
We would test it and update the code.
> bug in checkpoint size in vgg-16 model
> ---------------------------------------
>
> Key: SINGA-248
> URL: https://issues.apache.org/jira/browse/SINGA-248
> Project: Singa
> Issue Type: Bug
> Environment: ubuntu 14.04
> Reporter: hacker99
>
> i created vgg-16 net, then saved it with python interface
> (python/dragon/net.py) net.save('model.bin'),then find model.bin is about
> 1.5GB.but same model in caffe just 528MB. can anyone may explain why?very
> appreciate.
> vgg-16 code :
> from dragon import layer
> from dragon import initializer
> from dragon import metric
> from dragon import loss
> from dragon import net as ffnet
> def ConvReLU(net, name, nb_filers, sample_shape=None):
> net.add(layer.Conv2D(name + '_1', nb_filers, 3, 1, pad=1,
> input_sample_shape=sample_shape))
> net.add(layer.Activation(name + '_3'))
> def create_net(use_cpu=False):
> if use_cpu:
> layer.engine = 'dragoncpp'
> net = ffnet.FeedForwardNet(loss.SoftmaxCrossEntropy(), metric.Accuracy())
> ConvReLU(net, 'conv1_1', 64, (3, 224, 224))
> ConvReLU(net, 'conv1_2', 64)
> net.add(layer.MaxPooling2D('pool1', 2, 2, border_mode='valid'))
> ConvReLU(net, 'conv2_1', 128)
> ConvReLU(net, 'conv2_2', 128)
> net.add(layer.MaxPooling2D('pool2', 2, 2, border_mode='valid'))
> ConvReLU(net, 'conv3_1', 256)
> ConvReLU(net, 'conv3_2', 256)
> ConvReLU(net, 'conv3_3', 256)
> net.add(layer.MaxPooling2D('pool3', 2, 2, border_mode='valid'))
> ConvReLU(net, 'conv4_1', 512)
> ConvReLU(net, 'conv4_2', 512)
> ConvReLU(net, 'conv4_3', 512)
> net.add(layer.MaxPooling2D('pool4', 2, 2, border_mode='valid'))
> ConvReLU(net, 'conv5_1', 512)
> ConvReLU(net, 'conv5_2', 512)
> ConvReLU(net, 'conv5_3', 512)
> net.add(layer.MaxPooling2D('pool5', 2, 2, border_mode='valid'))
> net.add(layer.Flatten('flat'))
> net.add(layer.Dense('ip1', 4096))
> net.add(layer.Dropout('drop_ip1', 0.5))
> net.add(layer.Activation('relu_ip1'))
> net.add(layer.Dense('ip2', 4096))
> net.add(layer.Activation('relu_ip2'))
> net.add(layer.Dropout('drop_ip2', 0.5))
> #net.add(layer.BatchNormalization('batchnorm_ip1'))
> net.add(layer.Dense('ip3', 1000))
> for (p, name) in zip(net.param_values(), net.param_names()):
> print name, p.shape
> if 'mean' in name or 'beta' in name:
> p.set_value(0.0)
> elif 'var' in name:
> p.set_value(1.0)
> elif 'gamma' in name:
> initializer.uniform(p, 0, 1)
> elif len(p.shape) > 1:
> if 'conv' in name:
> initializer.gaussian(p, 0, 3 * 3 * p.shape[0])
> else:
> p.gaussian(0, 0.02)
> else:
> p.set_value(0)
> print name, p.l1()
> return net
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
