From 2f3665e988502d2574849af126f5688cf4f1abca Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Mon, 6 Nov 2017 13:25:57 +0800 Subject: [PATCH] update reset script for benchmark --- benchmark/paddle/image/resnet.py | 213 +++++++++++++++++++++++++++ benchmark/paddle/image/run_mkldnn.sh | 35 +++-- 2 files changed, 233 insertions(+), 15 deletions(-) create mode 100644 benchmark/paddle/image/resnet.py diff --git a/benchmark/paddle/image/resnet.py b/benchmark/paddle/image/resnet.py new file mode 100644 index 00000000000..6ae1857642e --- /dev/null +++ b/benchmark/paddle/image/resnet.py @@ -0,0 +1,213 @@ +#!/usr/bin/env python +from paddle.trainer_config_helpers import * + +height = 224 +width = 224 +num_class = 1000 +batch_size = get_config_arg('batch_size', int, 64) +layer_num = get_config_arg("layer_num", int, 50) +is_test = get_config_arg("is_test", bool, False) + +args = {'height': height, 'width': width, 'color': True, 'num_class': num_class} +define_py_data_sources2( + "train.list", None, module="provider", obj="process", args=args) + +settings( + batch_size=batch_size, + learning_rate=0.01 / batch_size, + learning_method=MomentumOptimizer(0.9), + regularization=L2Regularization(0.0005 * batch_size)) + + +#######################Network Configuration ############# +def conv_bn_layer(name, + input, + filter_size, + num_filters, + stride, + padding, + channels=None, + active_type=ReluActivation()): + """ + A wrapper for conv layer with batch normalization layers. + Note: + conv layer has no activation. + """ + + tmp = img_conv_layer( + name=name + "_conv", + input=input, + filter_size=filter_size, + num_channels=channels, + num_filters=num_filters, + stride=stride, + padding=padding, + act=LinearActivation(), + bias_attr=False) + return batch_norm_layer( + name=name + "_bn", input=tmp, act=active_type, use_global_stats=is_test) + + +def bottleneck_block(name, input, num_filters1, num_filters2): + """ + A wrapper for bottlenect building block in ResNet. + Last conv_bn_layer has no activation. + Addto layer has activation of relu. + """ + last_name = conv_bn_layer( + name=name + '_branch2a', + input=input, + filter_size=1, + num_filters=num_filters1, + stride=1, + padding=0) + last_name = conv_bn_layer( + name=name + '_branch2b', + input=last_name, + filter_size=3, + num_filters=num_filters1, + stride=1, + padding=1) + last_name = conv_bn_layer( + name=name + '_branch2c', + input=last_name, + filter_size=1, + num_filters=num_filters2, + stride=1, + padding=0, + active_type=LinearActivation()) + + return addto_layer( + name=name + "_addto", input=[input, last_name], act=ReluActivation()) + + +def mid_projection(name, input, num_filters1, num_filters2, stride=2): + """ + A wrapper for middile projection in ResNet. + projection shortcuts are used for increasing dimensions, + and other shortcuts are identity + branch1: projection shortcuts are used for increasing + dimensions, has no activation. + branch2x: bottleneck building block, shortcuts are identity. + """ + # stride = 2 + branch1 = conv_bn_layer( + name=name + '_branch1', + input=input, + filter_size=1, + num_filters=num_filters2, + stride=stride, + padding=0, + active_type=LinearActivation()) + + last_name = conv_bn_layer( + name=name + '_branch2a', + input=input, + filter_size=1, + num_filters=num_filters1, + stride=stride, + padding=0) + last_name = conv_bn_layer( + name=name + '_branch2b', + input=last_name, + filter_size=3, + num_filters=num_filters1, + stride=1, + padding=1) + + last_name = conv_bn_layer( + name=name + '_branch2c', + input=last_name, + filter_size=1, + num_filters=num_filters2, + stride=1, + padding=0, + active_type=LinearActivation()) + + return addto_layer( + name=name + "_addto", input=[branch1, last_name], act=ReluActivation()) + + +img = data_layer(name='image', size=height * width * 3) + + +def deep_res_net(res2_num=3, res3_num=4, res4_num=6, res5_num=3): + """ + A wrapper for 50,101,152 layers of ResNet. + res2_num: number of blocks stacked in conv2_x + res3_num: number of blocks stacked in conv3_x + res4_num: number of blocks stacked in conv4_x + res5_num: number of blocks stacked in conv5_x + """ + # For ImageNet + # conv1: 112x112 + tmp = conv_bn_layer( + "conv1", + input=img, + filter_size=7, + channels=3, + num_filters=64, + stride=2, + padding=3) + tmp = img_pool_layer(name="pool1", input=tmp, pool_size=3, stride=2) + + # conv2_x: 56x56 + tmp = mid_projection( + name="res2_1", input=tmp, num_filters1=64, num_filters2=256, stride=1) + for i in xrange(2, res2_num + 1, 1): + tmp = bottleneck_block( + name="res2_" + str(i), input=tmp, num_filters1=64, num_filters2=256) + + # conv3_x: 28x28 + tmp = mid_projection( + name="res3_1", input=tmp, num_filters1=128, num_filters2=512) + for i in xrange(2, res3_num + 1, 1): + tmp = bottleneck_block( + name="res3_" + str(i), + input=tmp, + num_filters1=128, + num_filters2=512) + + # conv4_x: 14x14 + tmp = mid_projection( + name="res4_1", input=tmp, num_filters1=256, num_filters2=1024) + for i in xrange(2, res4_num + 1, 1): + tmp = bottleneck_block( + name="res4_" + str(i), + input=tmp, + num_filters1=256, + num_filters2=1024) + + # conv5_x: 7x7 + tmp = mid_projection( + name="res5_1", input=tmp, num_filters1=512, num_filters2=2048) + for i in xrange(2, res5_num + 1, 1): + tmp = bottleneck_block( + name="res5_" + str(i), + input=tmp, + num_filters1=512, + num_filters2=2048) + + tmp = img_pool_layer( + name='avgpool', + input=tmp, + pool_size=7, + stride=1, + pool_type=AvgPooling()) + + return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation()) + + +if layer_num == 50: + resnet = deep_res_net(3, 4, 6, 3) +elif layer_num == 101: + resnet = deep_res_net(3, 4, 23, 3) +elif layer_num == 152: + resnet = deep_res_net(3, 8, 36, 3) +else: + print("Wrong layer number.") + +lbl = data_layer(name="label", size=num_class) +loss = cross_entropy(name='loss', input=resnet, label=lbl) +inputs(img, lbl) +outputs(loss) diff --git a/benchmark/paddle/image/run_mkldnn.sh b/benchmark/paddle/image/run_mkldnn.sh index e31fec1cd85..4a196015073 100755 --- a/benchmark/paddle/image/run_mkldnn.sh +++ b/benchmark/paddle/image/run_mkldnn.sh @@ -3,24 +3,26 @@ set -e function train() { unset OMP_NUM_THREADS MKL_NUM_THREADS export OMP_DYNAMIC="FALSE" + # TODO(TJ): auto 1.0 or 0,0 for HT on or off export KMP_AFFINITY="granularity=fine,compact,0,0" topology=$1 - bs=$2 - use_mkldnn=$3 - if [ $3 == "True" ]; then + layer_num=$2 + bs=$3 + use_mkldnn=$4 + if [ $4 == "True" ]; then thread=1 - log="logs/${topology}-mkldnn-${bs}.log" - elif [ $3 == "False" ]; then + log="logs/${topology}-${layer_num}-mkldnn-${bs}.log" + elif [ $4 == "False" ]; then thread=`nproc` # each trainer_count use only 1 core to avoid conflict export OMP_NUM_THREADS=1 export MKL_NUM_THREADS=1 - log="logs/${topology}-${thread}mklml-${bs}.log" + log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log" else echo "Wrong input $3, use True or False." exit 0 fi - args="batch_size=${bs}" + args="batch_size=${bs},layer_num=${layer_num}" config="${topology}.py" paddle train --job=time \ --config=$config \ @@ -40,12 +42,15 @@ if [ ! -d "logs" ]; then mkdir logs fi -#========== mkldnn ==========# -train vgg 64 True -train vgg 128 True -train vgg 256 True +for use_mkldnn in True False; do + for batchsize in 64 128 256; do + # vgg-19 and vgg-16 + train vgg 19 $batchsize $use_mkldnn + train vgg 16 $batchsize $use_mkldnn -#========== mklml ===========# -train vgg 64 False -train vgg 128 False -train vgg 256 False + # resnet-50, 101 and 152 + train resnet 50 $batchsize $use_mkldnn + train resnet 101 $batchsize $use_mkldnn + train resnet 152 $batchsize $use_mkldnn + done +done -- GitLab