From 7d5a67e9f0860be9e07a8950259fa4db5dd859b4 Mon Sep 17 00:00:00 2001 From: panfengfeng Date: Wed, 29 Jul 2020 20:53:47 +0800 Subject: [PATCH] googlenet-gpu --- model_zoo/official/cv/googlenet/eval.py | 17 ++- .../cv/googlenet/scripts/run_eval_gpu.sh | 43 ++++++++ .../cv/googlenet/scripts/run_train_gpu.sh | 45 ++++++++ .../official/cv/googlenet/src/googlenet.py | 102 ++++++++++++------ model_zoo/official/cv/googlenet/train.py | 51 ++++++--- 5 files changed, 204 insertions(+), 54 deletions(-) create mode 100644 model_zoo/official/cv/googlenet/scripts/run_eval_gpu.sh create mode 100644 model_zoo/official/cv/googlenet/scripts/run_train_gpu.sh diff --git a/model_zoo/official/cv/googlenet/eval.py b/model_zoo/official/cv/googlenet/eval.py index fc469879e7c..045aba56df4 100644 --- a/model_zoo/official/cv/googlenet/eval.py +++ b/model_zoo/official/cv/googlenet/eval.py @@ -16,6 +16,8 @@ ##############test googlenet example on cifar10################# python eval.py """ +import argparse + import mindspore.nn as nn from mindspore import context from mindspore.nn.optim.momentum import Momentum @@ -26,18 +28,27 @@ from src.config import cifar_cfg as cfg from src.dataset import create_dataset from src.googlenet import GoogleNet +parser = argparse.ArgumentParser(description='googlenet') +parser.add_argument('--checkpoint_path', type=str, default=None, help='Checkpoint file path') +args_opt = parser.parse_args() if __name__ == '__main__': + device_target = cfg.device_target context.set_context(mode=context.GRAPH_MODE, device_target=cfg.device_target) - context.set_context(device_id=cfg.device_id) + if device_target == "Ascend": + context.set_context(device_id=cfg.device_id) - net = GoogleNet(num_classes=cfg.num_classes) + net = GoogleNet(num_classes=cfg.num_classes, platform=device_target) opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), 0.01, cfg.momentum, weight_decay=cfg.weight_decay) loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean', is_grad=False) model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'}) - param_dict = load_checkpoint(cfg.checkpoint_path) + if device_target == "Ascend": + param_dict = load_checkpoint(cfg.checkpoint_path) + else: # GPU + param_dict = load_checkpoint(args_opt.checkpoint_path) + load_param_into_net(net, param_dict) net.set_train(False) dataset = create_dataset(cfg.data_path, 1, False) diff --git a/model_zoo/official/cv/googlenet/scripts/run_eval_gpu.sh b/model_zoo/official/cv/googlenet/scripts/run_eval_gpu.sh new file mode 100644 index 00000000000..b2e2a38737c --- /dev/null +++ b/model_zoo/official/cv/googlenet/scripts/run_eval_gpu.sh @@ -0,0 +1,43 @@ +#!/bin/bash +# Copyright 2020 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ + +ulimit -u unlimited + +if [ $# != 1 ] +then + echo "GPU: sh run_eval_gpu.sh [CHECKPOINT_PATH]" +exit 1 +fi + +# check checkpoint file +if [ ! -f $1 ] +then + echo "error: CHECKPOINT_PATH=$1 is not a file" +exit 1 +fi + +BASEPATH=$(cd "`dirname $0`" || exit; pwd) +export PYTHONPATH=${BASEPATH}:$PYTHONPATH +export DEVICE_ID=0 + +if [ -d "../eval" ]; +then + rm -rf ../eval +fi +mkdir ../eval +cd ../eval || exit + +python3 ${BASEPATH}/../eval.py --checkpoint_path=$1 > ./eval.log 2>&1 & diff --git a/model_zoo/official/cv/googlenet/scripts/run_train_gpu.sh b/model_zoo/official/cv/googlenet/scripts/run_train_gpu.sh new file mode 100644 index 00000000000..3f2419f7eb8 --- /dev/null +++ b/model_zoo/official/cv/googlenet/scripts/run_train_gpu.sh @@ -0,0 +1,45 @@ +#!/bin/bash +# Copyright 2020 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ + +if [ $# -lt 2 ] +then + echo "Usage:\n \ + sh run_train.sh [DEVICE_NUM] [VISIABLE_DEVICES(0,1,2,3,4,5,6,7)]\n \ + " +exit 1 +fi + +if [ $1 -lt 1 ] && [ $1 -gt 8 ] +then + echo "error: DEVICE_NUM=$1 is not in (1-8)" +exit 1 +fi + +export DEVICE_NUM=$1 +export RANK_SIZE=$1 + +BASEPATH=$(cd "`dirname $0`" || exit; pwd) +export PYTHONPATH=${BASEPATH}:$PYTHONPATH +if [ -d "../train" ]; +then + rm -rf ../train +fi +mkdir ../train +cd ../train || exit + +export CUDA_VISIBLE_DEVICES="$2" +mpirun -n $1 --allow-run-as-root \ +python3 ${BASEPATH}/../train.py > train.log 2>&1 & diff --git a/model_zoo/official/cv/googlenet/src/googlenet.py b/model_zoo/official/cv/googlenet/src/googlenet.py index 701b3aeb5a9..20bd96454b9 100644 --- a/model_zoo/official/cv/googlenet/src/googlenet.py +++ b/model_zoo/official/cv/googlenet/src/googlenet.py @@ -56,24 +56,35 @@ class Inception(nn.Cell): Inception Block """ - def __init__(self, in_channels, n1x1, n3x3red, n3x3, n5x5red, n5x5, pool_planes): + def __init__(self, in_channels, n1x1, n3x3red, n3x3, n5x5red, n5x5, pool_planes, platform="Ascend"): super(Inception, self).__init__() + self.platform = platform self.b1 = Conv2dBlock(in_channels, n1x1, kernel_size=1) self.b2 = nn.SequentialCell([Conv2dBlock(in_channels, n3x3red, kernel_size=1), Conv2dBlock(n3x3red, n3x3, kernel_size=3, padding=0)]) self.b3 = nn.SequentialCell([Conv2dBlock(in_channels, n5x5red, kernel_size=1), Conv2dBlock(n5x5red, n5x5, kernel_size=3, padding=0)]) - self.maxpool = P.MaxPoolWithArgmax(ksize=3, strides=1, padding="same") + if self.platform == "Ascend": + self.maxpool = P.MaxPoolWithArgmax(ksize=3, strides=1, padding="same") + else: # GPU + self.maxpool = P.MaxPool(ksize=3, strides=1, padding="same") self.b4 = Conv2dBlock(in_channels, pool_planes, kernel_size=1) self.concat = P.Concat(axis=1) def construct(self, x): + ''' + construct inception model + ''' branch1 = self.b1(x) branch2 = self.b2(x) branch3 = self.b3(x) - cell, argmax = self.maxpool(x) - branch4 = self.b4(cell) - _ = argmax + if self.platform == "Ascend": + cell, argmax = self.maxpool(x) + branch4 = self.b4(cell) + _ = argmax + else: # GPU + cell = self.maxpool(x) + branch4 = self.b4(cell) return self.concat((branch1, branch2, branch3, branch4)) @@ -82,61 +93,82 @@ class GoogleNet(nn.Cell): Googlenet architecture """ - def __init__(self, num_classes): + def __init__(self, num_classes, platform="Ascend"): super(GoogleNet, self).__init__() self.conv1 = Conv2dBlock(3, 64, kernel_size=7, stride=2, padding=0) - self.maxpool1 = P.MaxPoolWithArgmax(ksize=3, strides=2, padding="same") - + self.platform = platform self.conv2 = Conv2dBlock(64, 64, kernel_size=1) self.conv3 = Conv2dBlock(64, 192, kernel_size=3, padding=0) - self.maxpool2 = P.MaxPoolWithArgmax(ksize=3, strides=2, padding="same") - - self.block3a = Inception(192, 64, 96, 128, 16, 32, 32) - self.block3b = Inception(256, 128, 128, 192, 32, 96, 64) - self.maxpool3 = P.MaxPoolWithArgmax(ksize=3, strides=2, padding="same") - - self.block4a = Inception(480, 192, 96, 208, 16, 48, 64) - self.block4b = Inception(512, 160, 112, 224, 24, 64, 64) - self.block4c = Inception(512, 128, 128, 256, 24, 64, 64) - self.block4d = Inception(512, 112, 144, 288, 32, 64, 64) - self.block4e = Inception(528, 256, 160, 320, 32, 128, 128) - self.maxpool4 = P.MaxPoolWithArgmax(ksize=2, strides=2, padding="same") - - self.block5a = Inception(832, 256, 160, 320, 32, 128, 128) - self.block5b = Inception(832, 384, 192, 384, 48, 128, 128) - + self.block3a = Inception(192, 64, 96, 128, 16, 32, 32, platform=self.platform) + self.block3b = Inception(256, 128, 128, 192, 32, 96, 64, platform=self.platform) + self.block4a = Inception(480, 192, 96, 208, 16, 48, 64, platform=self.platform) + self.block4b = Inception(512, 160, 112, 224, 24, 64, 64, platform=self.platform) + self.block4c = Inception(512, 128, 128, 256, 24, 64, 64, platform=self.platform) + self.block4d = Inception(512, 112, 144, 288, 32, 64, 64, platform=self.platform) + self.block4e = Inception(528, 256, 160, 320, 32, 128, 128, platform=self.platform) + if self.platform == "Ascend": + self.maxpool1 = P.MaxPoolWithArgmax(ksize=3, strides=2, padding="same") + self.maxpool2 = P.MaxPoolWithArgmax(ksize=3, strides=2, padding="same") + self.maxpool3 = P.MaxPoolWithArgmax(ksize=3, strides=2, padding="same") + self.maxpool4 = P.MaxPoolWithArgmax(ksize=2, strides=2, padding="same") + else: # GPU + self.maxpool1 = P.MaxPool(ksize=3, strides=2, padding="same") + self.maxpool2 = P.MaxPool(ksize=3, strides=2, padding="same") + self.maxpool3 = P.MaxPool(ksize=3, strides=2, padding="same") + self.maxpool4 = P.MaxPool(ksize=2, strides=2, padding="same") + self.block5a = Inception(832, 256, 160, 320, 32, 128, 128, platform=self.platform) + self.block5b = Inception(832, 384, 192, 384, 48, 128, 128, platform=self.platform) self.mean = P.ReduceMean(keep_dims=True) self.dropout = nn.Dropout(keep_prob=0.8) self.flatten = nn.Flatten() self.classifier = nn.Dense(1024, num_classes, weight_init=weight_variable(), bias_init=weight_variable()) - def construct(self, x): + ''' + construct googlenet model + ''' x = self.conv1(x) - x, argmax = self.maxpool1(x) + if self.platform == "Ascend": + x, argmax = self.maxpool1(x) + else: # GPU + x = self.maxpool1(x) x = self.conv2(x) x = self.conv3(x) - x, argmax = self.maxpool2(x) + if self.platform == "Ascend": + x, argmax = self.maxpool2(x) + else: # GPU + x = self.maxpool2(x) x = self.block3a(x) x = self.block3b(x) - x, argmax = self.maxpool3(x) + if self.platform == "Ascend": + x, argmax = self.maxpool3(x) + else: # GPU + x = self.maxpool3(x) x = self.block4a(x) x = self.block4b(x) x = self.block4c(x) x = self.block4d(x) x = self.block4e(x) - x, argmax = self.maxpool4(x) + if self.platform == "Ascend": + x, argmax = self.maxpool4(x) + x = self.block5a(x) + x = self.block5b(x) - x = self.block5a(x) - x = self.block5b(x) + x = self.mean(x, (2, 3)) + x = self.flatten(x) + x = self.classifier(x) + _ = argmax + else: # GPU + x = self.maxpool4(x) + x = self.block5a(x) + x = self.block5b(x) - x = self.mean(x, (2, 3)) - x = self.flatten(x) - x = self.classifier(x) + x = self.mean(x, (2, 3)) + x = self.flatten(x) + x = self.classifier(x) - _ = argmax return x diff --git a/model_zoo/official/cv/googlenet/train.py b/model_zoo/official/cv/googlenet/train.py index a7c72e2763e..b7668a017f9 100644 --- a/model_zoo/official/cv/googlenet/train.py +++ b/model_zoo/official/cv/googlenet/train.py @@ -25,7 +25,7 @@ import numpy as np import mindspore.nn as nn from mindspore import Tensor from mindspore import context -from mindspore.communication.management import init +from mindspore.communication.management import init, get_rank from mindspore.nn.optim.momentum import Momentum from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor from mindspore.train.model import Model, ParallelMode @@ -38,7 +38,6 @@ from src.googlenet import GoogleNet random.seed(1) np.random.seed(1) - def lr_steps(global_step, lr_max=None, total_epochs=None, steps_per_epoch=None): """Set learning rate.""" lr_each_step = [] @@ -65,23 +64,36 @@ if __name__ == '__main__': parser.add_argument('--device_id', type=int, default=None, help='device id of GPU or Ascend. (Default: None)') args_opt = parser.parse_args() - context.set_context(mode=context.GRAPH_MODE, device_target=cfg.device_target) - if args_opt.device_id is not None: - context.set_context(device_id=args_opt.device_id) - else: - context.set_context(device_id=cfg.device_id) + device_target = cfg.device_target + context.set_context(mode=context.GRAPH_MODE, device_target=cfg.device_target) device_num = int(os.environ.get("DEVICE_NUM", 1)) - if device_num > 1: - context.reset_auto_parallel_context() - context.set_auto_parallel_context(device_num=device_num, parallel_mode=ParallelMode.DATA_PARALLEL, - mirror_mean=True) - init() + + if device_target == "Ascend": + if args_opt.device_id is not None: + context.set_context(device_id=args_opt.device_id) + else: + context.set_context(device_id=cfg.device_id) + + if device_num > 1: + context.reset_auto_parallel_context() + context.set_auto_parallel_context(device_num=device_num, parallel_mode=ParallelMode.DATA_PARALLEL, + mirror_mean=True) + init() + elif device_target == "GPU": + init("nccl") + + if device_num > 1: + context.reset_auto_parallel_context() + context.set_auto_parallel_context(device_num=device_num, parallel_mode=ParallelMode.DATA_PARALLEL, + mirror_mean=True) + else: + raise ValueError("Unsupport platform.") dataset = create_dataset(cfg.data_path, 1) batch_num = dataset.get_dataset_size() - net = GoogleNet(num_classes=cfg.num_classes) + net = GoogleNet(num_classes=cfg.num_classes, platform=device_target) # Continue training if set pre_trained to be True if cfg.pre_trained: param_dict = load_checkpoint(cfg.checkpoint_path) @@ -90,12 +102,19 @@ if __name__ == '__main__': opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), Tensor(lr), cfg.momentum, weight_decay=cfg.weight_decay) loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean', is_grad=False) - model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'}, - amp_level="O2", keep_batchnorm_fp32=False, loss_scale_manager=None) + + if device_target == "Ascend": + model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'}, + amp_level="O2", keep_batchnorm_fp32=False, loss_scale_manager=None) + ckpt_save_dir = "./" + else: # GPU + model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'}, + amp_level="O2", keep_batchnorm_fp32=True, loss_scale_manager=None) + ckpt_save_dir = "./ckpt_" + str(get_rank()) + "/" config_ck = CheckpointConfig(save_checkpoint_steps=batch_num * 5, keep_checkpoint_max=cfg.keep_checkpoint_max) time_cb = TimeMonitor(data_size=batch_num) - ckpoint_cb = ModelCheckpoint(prefix="train_googlenet_cifar10", directory="./", config=config_ck) + ckpoint_cb = ModelCheckpoint(prefix="train_googlenet_cifar10", directory=ckpt_save_dir, config=config_ck) loss_cb = LossMonitor() model.train(cfg.epoch_size, dataset, callbacks=[time_cb, ckpoint_cb, loss_cb]) print("train success")