提交 54bd17fe 编写于 作者: Y Yu Yang

Complete Flowers

上级 50e7e25d
...@@ -31,7 +31,13 @@ std::string OpHandleBase::DebugString() const { ...@@ -31,7 +31,13 @@ std::string OpHandleBase::DebugString() const {
return ss.str(); return ss.str();
} }
OpHandleBase::~OpHandleBase() {} OpHandleBase::~OpHandleBase() {
#ifdef PADDLE_WITH_CUDA
for (auto &ev : events_) {
cudaEventDestroy(ev.second);
}
#endif
}
void OpHandleBase::Run(bool use_event) { void OpHandleBase::Run(bool use_event) {
#ifdef PADDLE_WITH_CUDA #ifdef PADDLE_WITH_CUDA
......
...@@ -21,7 +21,7 @@ void SSAGraphBuilder::PolishGraphToSupportDataHazards(SSAGraph *graph) { ...@@ -21,7 +21,7 @@ void SSAGraphBuilder::PolishGraphToSupportDataHazards(SSAGraph *graph) {
for (auto &var_map : graph->vars_) { for (auto &var_map : graph->vars_) {
for (auto &name_pair : var_map) { for (auto &name_pair : var_map) {
if (name_pair.second.size() <= 1) { if (name_pair.second.size() <= 1) {
return; continue;
} }
auto it_new = name_pair.second.rbegin(); auto it_new = name_pair.second.rbegin();
auto it_old = name_pair.second.rbegin(); auto it_old = name_pair.second.rbegin();
......
...@@ -2,3 +2,4 @@ mnist.recordio ...@@ -2,3 +2,4 @@ mnist.recordio
mnist_0.recordio mnist_0.recordio
mnist_1.recordio mnist_1.recordio
mnist_2.recordio mnist_2.recordio
flowers.recordio
...@@ -16,6 +16,7 @@ import unittest ...@@ -16,6 +16,7 @@ import unittest
import paddle.fluid as fluid import paddle.fluid as fluid
import paddle.v2 as paddle import paddle.v2 as paddle
import paddle.v2.dataset.mnist as mnist import paddle.v2.dataset.mnist as mnist
import paddle.v2.dataset.flowers as flowers
import numpy import numpy
...@@ -64,6 +65,119 @@ def fc_with_batchnorm(): ...@@ -64,6 +65,119 @@ def fc_with_batchnorm():
return loss return loss
def squeeze_excitation(input, num_channels, reduction_ratio):
# pool = fluid.layers.pool2d(
# input=input, pool_size=0, pool_type='avg', global_pooling=True)
conv = input
shape = conv.shape
reshape = fluid.layers.reshape(
x=conv, shape=[-1, shape[1], shape[2] * shape[3]])
pool = fluid.layers.reduce_mean(input=reshape, dim=2)
squeeze = fluid.layers.fc(input=pool,
size=num_channels / reduction_ratio,
act='relu')
excitation = fluid.layers.fc(input=squeeze,
size=num_channels,
act='sigmoid')
scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0)
return scale
def conv_bn_layer(input, num_filters, filter_size, stride=1, groups=1,
act=None):
conv = fluid.layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
stride=stride,
padding=(filter_size - 1) / 2,
groups=groups,
act=None,
bias_attr=False)
return fluid.layers.batch_norm(input=conv, act=act, momentum=0.1)
def shortcut(input, ch_out, stride):
ch_in = input.shape[1]
if ch_in != ch_out:
if stride == 1:
filter_size = 1
else:
filter_size = 3
return conv_bn_layer(input, ch_out, filter_size, stride)
else:
return input
def bottleneck_block(input, num_filters, stride, cardinality, reduction_ratio):
# The number of first 1x1 convolutional channels for each bottleneck build block
# was halved to reduce the compution cost.
conv0 = conv_bn_layer(
input=input, num_filters=num_filters, filter_size=1, act='relu')
conv1 = conv_bn_layer(
input=conv0,
num_filters=num_filters * 2,
filter_size=3,
stride=stride,
groups=cardinality,
act='relu')
conv2 = conv_bn_layer(
input=conv1, num_filters=num_filters * 2, filter_size=1, act=None)
scale = squeeze_excitation(
input=conv2,
num_channels=num_filters * 2,
reduction_ratio=reduction_ratio)
short = shortcut(input, num_filters * 2, stride)
return fluid.layers.elementwise_add(x=short, y=scale, act='relu')
def SE_ResNeXt152():
reader = fluid.layers.open_recordio_file(
filename='./flowers.recordio',
shapes=[[-1, 3, 224, 224], [-1, 1]],
lod_levels=[0, 0],
dtypes=['float32', 'int64'])
img, label = fluid.layers.read_file(reader)
conv = conv_bn_layer(
input=img, num_filters=64, filter_size=3, stride=2, act='relu')
conv = conv_bn_layer(
input=conv, num_filters=64, filter_size=3, stride=1, act='relu')
conv = conv_bn_layer(
input=conv, num_filters=128, filter_size=3, stride=1, act='relu')
conv = fluid.layers.pool2d(
input=conv, pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
cardinality = 64
reduction_ratio = 16
depth = [3, 8, 36, 3]
num_filters = [128, 256, 512, 1024]
for block in range(len(depth)):
for i in range(depth[block]):
conv = bottleneck_block(
input=conv,
num_filters=num_filters[block],
stride=2 if i == 0 and block != 0 else 1,
cardinality=cardinality,
reduction_ratio=reduction_ratio)
shape = conv.shape
reshape = fluid.layers.reshape(
x=conv, shape=[-1, shape[1], shape[2] * shape[3]])
pool = fluid.layers.reduce_mean(input=reshape, dim=2)
dropout = fluid.layers.dropout(x=pool, dropout_prob=0.2)
# Classifier layer:
prediction = fluid.layers.fc(input=dropout, size=1000, act='softmax')
loss = fluid.layers.cross_entropy(input=prediction, label=label)
loss = fluid.layers.mean(loss)
return loss
class ParallelExecutor(unittest.TestCase): class ParallelExecutor(unittest.TestCase):
@classmethod @classmethod
def setUpClass(cls): def setUpClass(cls):
...@@ -81,24 +195,40 @@ class ParallelExecutor(unittest.TestCase): ...@@ -81,24 +195,40 @@ class ParallelExecutor(unittest.TestCase):
fluid.recordio_writer.convert_reader_to_recordio_file( fluid.recordio_writer.convert_reader_to_recordio_file(
'./mnist.recordio', reader, feeder) './mnist.recordio', reader, feeder)
with fluid.program_guard(fluid.Program(), fluid.Program()):
reader = paddle.batch(flowers.train(), batch_size=4)
feeder = fluid.DataFeeder(
feed_list=[
fluid.layers.data(
name='image', shape=[3, 224, 224]),
fluid.layers.data(
name='label', shape=[1], dtype='int64'),
],
place=fluid.CPUPlace())
fluid.recordio_writer.convert_reader_to_recordio_file(
"./flowers.recordio", reader, feeder)
def test_simple_fc(self): def test_simple_fc(self):
self.check_network_convergence(simple_fc_net) self.check_network_convergence(simple_fc_net)
def test_batchnorm_fc(self): def test_batchnorm_fc(self):
self.check_network_convergence(fc_with_batchnorm) self.check_network_convergence(fc_with_batchnorm)
def check_network_convergence(self, method): def check_network_convergence(self, method, memory_opt=True, iter=10):
main = fluid.Program() main = fluid.Program()
startup = fluid.Program() startup = fluid.Program()
with fluid.program_guard(main, startup): with fluid.program_guard(main, startup):
loss = method() loss = method()
adam = fluid.optimizer.Adam() adam = fluid.optimizer.Adam()
adam.minimize(loss) adam.minimize(loss)
if memory_opt:
fluid.memory_optimize(main)
exe = fluid.ParallelExecutor(loss_name=loss.name, use_cuda=True) exe = fluid.ParallelExecutor(loss_name=loss.name, use_cuda=True)
first_loss, = exe.run([loss.name]) first_loss, = exe.run([loss.name])
first_loss = numpy.array(first_loss) first_loss = numpy.array(first_loss)
for i in xrange(10): for i in xrange(iter):
exe.run([]) exe.run([])
last_loss, = exe.run([loss.name]) last_loss, = exe.run([loss.name])
...@@ -106,3 +236,6 @@ class ParallelExecutor(unittest.TestCase): ...@@ -106,3 +236,6 @@ class ParallelExecutor(unittest.TestCase):
print first_loss, last_loss print first_loss, last_loss
self.assertGreater(first_loss[0], last_loss[0]) self.assertGreater(first_loss[0], last_loss[0])
def test_resnet(self):
self.check_network_convergence(SE_ResNeXt152, iter=20)
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册
反馈
建议
客服 返回
顶部