p04896132
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!1964 add backbone for resnet101 

Merge pull request !1964 from meixiaowei/master
This commit is contained in:
mindspore-ci-bot 2020-06-10 19:43:35 +08:00 committed by Gitee
parent 18e2a0f12e 0fe4f09fb5
commit 8ab436910f
14 changed files with 264 additions and 767 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

142
example/resnet101_imagenet2012/README.md
View File

@ -1,142 +0,0 @@
# ResNet101 Example
## Description
This is an example of training ResNet101 with ImageNet dataset in MindSpore.
## Requirements
- Install [MindSpore](https://www.mindspore.cn/install/en).
- Download the dataset ImageNet2012.
> Unzip the ImageNet2012 dataset to any path you want, the folder should include train and eval dataset as follows:
```
.
└─dataset
├─ilsvrc
│
└─validation_preprocess
```
## Example structure
```shell
.
├── crossentropy.py # CrossEntropy loss function
├── config.py # parameter configuration
├── dataset.py # data preprocessing
├── eval.py # eval net
├── lr_generator.py # generate learning rate
├── run_distribute_train.sh # launch distributed training(8p)
├── run_infer.sh # launch evaluating
├── run_standalone_train.sh # launch standalone training(1p)
└── train.py # train net
```
## Parameter configuration
Parameters for both training and evaluating can be set in config.py.
```
"class_num": 1001, # dataset class number
"batch_size": 32, # batch size of input tensor
"loss_scale": 1024, # loss scale
"momentum": 0.9, # momentum optimizer
"weight_decay": 1e-4, # weight decay
"epoch_size": 120, # epoch sizes for training
"pretrain_epoch_size": 0, # epoch size of pretrain checkpoint
"buffer_size": 1000, # number of queue size in data preprocessing
"image_height": 224, # image height
"image_width": 224, # image width
"save_checkpoint": True, # whether save checkpoint or not
"save_checkpoint_epochs": 1, # the epoch interval between two checkpoints. By default, the last checkpoint will be saved after the last epoch
"keep_checkpoint_max": 10, # only keep the last keep_checkpoint_max checkpoint
"save_checkpoint_path": "./", # path to save checkpoint relative to the executed path
"warmup_epochs": 0, # number of warmup epoch
"lr_decay_mode": "cosine" # decay mode for generating learning rate
"label_smooth": 1, # label_smooth
"label_smooth_factor": 0.1, # label_smooth_factor
"lr": 0.1 # base learning rate
```
## Running the example
### Train
#### Usage
```
# distributed training
sh run_distribute_train.sh [MINDSPORE_HCCL_CONFIG_PATH] [DATASET_PATH] [PRETRAINED_PATH](optional)
# standalone training
sh run_standalone_train.sh [DATASET_PATH] [PRETRAINED_PATH](optional)
```
#### Launch
```bash
# distributed training example(8p)
sh run_distribute_train.sh rank_table_8p.json dataset/ilsvrc
If you want to load pretrained ckpt file,
sh run_distribute_train.sh rank_table_8p.json dataset/ilsvrc ./ckpt/pretrained.ckpt
# standalone training example（1p）
sh run_standalone_train.sh dataset/ilsvrc
If you want to load pretrained ckpt file,
sh run_standalone_train.sh dataset/ilsvrc ./ckpt/pretrained.ckpt
```
> About rank_table.json, you can refer to the [distributed training tutorial](https://www.mindspore.cn/tutorial/en/master/advanced_use/distributed_training.html).
#### Result
Training result will be stored in the example path, whose folder name begins with "train" or "train_parallel". You can find checkpoint file together with result like the followings in log.
```
# distribute training result(8p)
epoch: 1 step: 5004, loss is 4.805483
epoch: 2 step: 5004, loss is 3.2121816
epoch: 3 step: 5004, loss is 3.429647
epoch: 4 step: 5004, loss is 3.3667371
epoch: 5 step: 5004, loss is 3.1718972
...
epoch: 67 step: 5004, loss is 2.2768745
epoch: 68 step: 5004, loss is 1.7223864
epoch: 69 step: 5004, loss is 2.0665488
epoch: 70 step: 5004, loss is 1.8717369
...
```
### Infer
#### Usage
```
# infer
sh run_infer.sh [VALIDATION_DATASET_PATH] [CHECKPOINT_PATH]
```
#### Launch
```bash
# infer with checkpoint
sh run_infer.sh dataset/validation_preprocess/ train_parallel0/resnet-120_5004.ckpt
```
> checkpoint can be produced in training process.
#### Result
Inference result will be stored in the example path, whose folder name is "infer". Under this, you can find result like the followings in log.
```
result: {'top_5_accuracy': 0.9429417413572343, 'top_1_accuracy': 0.7853513124199744} ckpt=train_parallel0/resnet-120_5004.ckpt
```

40
example/resnet101_imagenet2012/config.py
View File

@ -1,40 +0,0 @@
# 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.
# ============================================================================
"""
network config setting, will be used in train.py and eval.py
"""
from easydict import EasyDict as ed
config = ed({
"class_num": 1001,
"batch_size": 32,
"loss_scale": 1024,
"momentum": 0.9,
"weight_decay": 1e-4,
"epoch_size": 120,
"pretrain_epoch_size": 0,
"buffer_size": 1000,
"image_height": 224,
"image_width": 224,
"save_checkpoint": True,
"save_checkpoint_epochs": 5,
"keep_checkpoint_max": 10,
"save_checkpoint_path": "./",
"warmup_epochs": 0,
"lr_decay_mode": "cosine",
"label_smooth": 1,
"label_smooth_factor": 0.1,
"lr": 0.1
})

36
example/resnet101_imagenet2012/crossentropy.py
View File

@ -1,36 +0,0 @@
# 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.
# ============================================================================
"""define loss function for network"""
from mindspore.nn.loss.loss import _Loss
from mindspore.ops import operations as P
from mindspore.ops import functional as F
from mindspore import Tensor
from mindspore.common import dtype as mstype
import mindspore.nn as nn
class CrossEntropy(_Loss):
"""the redefined loss function with SoftmaxCrossEntropyWithLogits"""
def __init__(self, smooth_factor=0., num_classes=1001):
super(CrossEntropy, self).__init__()
self.onehot = P.OneHot()
self.on_value = Tensor(1.0 - smooth_factor, mstype.float32)
self.off_value = Tensor(1.0 * smooth_factor / (num_classes -1), mstype.float32)
self.ce = nn.SoftmaxCrossEntropyWithLogits()
self.mean = P.ReduceMean(False)
def construct(self, logit, label):
one_hot_label = self.onehot(label, F.shape(logit)[1], self.on_value, self.off_value)
loss = self.ce(logit, one_hot_label)
loss = self.mean(loss, 0)
return loss

89
example/resnet101_imagenet2012/dataset.py
View File

@ -1,89 +0,0 @@
# 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.
# ============================================================================
"""
create train or eval dataset.
"""
import os
import mindspore.common.dtype as mstype
import mindspore.dataset.engine as de
import mindspore.dataset.transforms.vision.c_transforms as C
import mindspore.dataset.transforms.c_transforms as C2
from config import config
def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32):
"""
create a train or evaluate dataset
Args:
dataset_path(string): the path of dataset.
do_train(bool): whether dataset is used for train or eval.
repeat_num(int): the repeat times of dataset. Default: 1
batch_size(int): the batch size of dataset. Default: 32
Returns:
dataset
"""
device_num = int(os.getenv("RANK_SIZE"))
rank_id = int(os.getenv("RANK_ID"))
if device_num == 1:
ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True)
else:
ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True,
num_shards=device_num, shard_id=rank_id)
resize_height = 224
rescale = 1.0 / 255.0
shift = 0.0
# define map operations
decode_op = C.Decode()
random_resize_crop_op = C.RandomResizedCrop(resize_height, (0.08, 1.0), (0.75, 1.33), max_attempts=100)
horizontal_flip_op = C.RandomHorizontalFlip(rank_id / (rank_id + 1))
resize_op_256 = C.Resize((256, 256))
center_crop = C.CenterCrop(224)
rescale_op = C.Rescale(rescale, shift)
normalize_op = C.Normalize((0.475, 0.451, 0.392), (0.275, 0.267, 0.278))
changeswap_op = C.HWC2CHW()
trans = []
if do_train:
trans = [decode_op,
random_resize_crop_op,
horizontal_flip_op,
rescale_op,
normalize_op,
changeswap_op]
else:
trans = [decode_op,
resize_op_256,
center_crop,
rescale_op,
normalize_op,
changeswap_op]
type_cast_op = C2.TypeCast(mstype.int32)
ds = ds.map(input_columns="image", operations=trans, num_parallel_workers=8)
ds = ds.map(input_columns="label", operations=type_cast_op, num_parallel_workers=8)
# apply shuffle operations
ds = ds.shuffle(buffer_size=config.buffer_size)
# apply batch operations
ds = ds.batch(batch_size, drop_remainder=True)
# apply dataset repeat operation
ds = ds.repeat(repeat_num)
return ds

75
example/resnet101_imagenet2012/eval.py
View File

@ -1,75 +0,0 @@
# 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.
# ============================================================================
"""
eval.
"""
import os
import argparse
import random
import numpy as np
from dataset import create_dataset
from config import config
from mindspore import context
from mindspore.model_zoo.resnet import resnet101
from mindspore.parallel._auto_parallel_context import auto_parallel_context
from mindspore.train.model import Model, ParallelMode
from mindspore.train.serialization import load_checkpoint, load_param_into_net
import mindspore.dataset.engine as de
from mindspore.communication.management import init
from crossentropy import CrossEntropy
random.seed(1)
np.random.seed(1)
de.config.set_seed(1)
parser = argparse.ArgumentParser(description='Image classification')
parser.add_argument('--run_distribute', type=bool, default=False, help='Run distribute')
parser.add_argument('--device_num', type=int, default=1, help='Device num.')
parser.add_argument('--do_train', type=bool, default=False, help='Do train or not.')
parser.add_argument('--do_eval', type=bool, default=True, help='Do eval or not.')
parser.add_argument('--checkpoint_path', type=str, default=None, help='Checkpoint file path')
parser.add_argument('--dataset_path', type=str, default=None, help='Dataset path')
args_opt = parser.parse_args()
device_id = int(os.getenv('DEVICE_ID'))
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id)
if __name__ == '__main__':
if not args_opt.do_eval and args_opt.run_distribute:
context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True, parameter_broadcast=True)
auto_parallel_context().set_all_reduce_fusion_split_indices([180, 313])
init()
epoch_size = config.epoch_size
net = resnet101(class_num=config.class_num)
if not config.label_smooth:
config.label_smooth_factor = 0.0
loss = CrossEntropy(smooth_factor=config.label_smooth_factor, num_classes=config.class_num)
if args_opt.do_eval:
dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=False, batch_size=config.batch_size)
step_size = dataset.get_dataset_size()
if args_opt.checkpoint_path:
param_dict = load_checkpoint(args_opt.checkpoint_path)
load_param_into_net(net, param_dict)
net.set_train(False)
model = Model(net, loss_fn=loss, metrics={'top_1_accuracy', 'top_5_accuracy'})
res = model.eval(dataset)
print("result:", res, "ckpt=", args_opt.checkpoint_path)

56
example/resnet101_imagenet2012/lr_generator.py
View File

@ -1,56 +0,0 @@
# 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.
# ============================================================================
"""learning rate generator"""
import math
import numpy as np
def linear_warmup_lr(current_step, warmup_steps, base_lr, init_lr):
lr_inc = (float(base_lr) - float(init_lr)) / float(warmup_steps)
lr = float(init_lr) + lr_inc * current_step
return lr
def warmup_cosine_annealing_lr(lr, steps_per_epoch, warmup_epochs, max_epoch=120, global_step=0):
"""
generate learning rate array with cosine
Args:
lr(float): base learning rate
steps_per_epoch(int): steps size of one epoch
warmup_epochs(int): number of warmup epochs
max_epoch(int): total epochs of training
global_step(int): the current start index of lr array
Returns:
np.array, learning rate array
"""
base_lr = lr
warmup_init_lr = 0
total_steps = int(max_epoch * steps_per_epoch)
warmup_steps = int(warmup_epochs * steps_per_epoch)
decay_steps = total_steps - warmup_steps
lr_each_step = []
for i in range(total_steps):
if i < warmup_steps:
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
else:
linear_decay = (total_steps - i) / decay_steps
cosine_decay = 0.5 * (1 + math.cos(math.pi * 2 * 0.47 * i / decay_steps))
decayed = linear_decay * cosine_decay + 0.00001
lr = base_lr * decayed
lr_each_step.append(lr)
lr_each_step = np.array(lr_each_step).astype(np.float32)
learning_rate = lr_each_step[global_step:]
return learning_rate

86
example/resnet101_imagenet2012/run_distribute_train.sh
View File

@ -1,86 +0,0 @@
#!/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 [ $# != 2 ] && [ $# != 3 ]
then
echo "Usage: sh run_distribute_train.sh [MINDSPORE_HCCL_CONFIG_PATH] [DATASET_PATH] [PRETRAINED_PATH](optional)"
exit 1
fi
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
PATH1=$(get_real_path $1)
PATH2=$(get_real_path $2)
echo $PATH1
echo $PATH2
if [ $# == 3 ]
then
PATH3=$(get_real_path $3)
echo $PATH3
fi
if [ ! -f $PATH1 ]
then
echo "error: MINDSPORE_HCCL_CONFIG_PATH=$PATH1 is not a file"
exit 1
fi
if [ ! -d $PATH2 ]
then
echo "error: DATASET_PATH=$PATH2 is not a directory"
exit 1
fi
if [ $# == 3 ] && [ ! -f $PATH3 ]
then
echo "error: PRETRAINED_PATH=$PATH3 is not a file"
exit 1
fi
ulimit -u unlimited
export DEVICE_NUM=8
export RANK_SIZE=8
export MINDSPORE_HCCL_CONFIG_PATH=$PATH1
export RANK_TABLE_FILE=$PATH1
for((i=0; i<${DEVICE_NUM}; i++))
do
export DEVICE_ID=$i
export RANK_ID=$i
rm -rf ./train_parallel$i
mkdir ./train_parallel$i
cp *.py ./train_parallel$i
cp *.sh ./train_parallel$i
cd ./train_parallel$i || exit
echo "start training for rank $RANK_ID, device $DEVICE_ID"
env > env.log
if [ $# == 2 ]
then
python train.py --do_train=True --run_distribute=True --device_num=$DEVICE_NUM --dataset_path=$PATH2 &> log &
fi
if [ $# == 3 ]
then
python train.py --do_train=True --run_distribute=True --device_num=$DEVICE_NUM --dataset_path=$PATH2 --pre_trained=$PATH3 &> log &
fi
cd ..
done

64
example/resnet101_imagenet2012/run_infer.sh
View File

@ -1,64 +0,0 @@
#!/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 [ $# != 2 ]
then
echo "Usage: sh run_infer.sh [DATASET_PATH] [CHECKPOINT_PATH]"
exit 1
fi
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
PATH1=$(get_real_path $1)
PATH2=$(get_real_path $2)
echo $PATH1
echo $PATH2
if [ ! -d $PATH1 ]
then
echo "error: DATASET_PATH=$PATH1 is not a directory"
exit 1
fi
if [ ! -f $PATH2 ]
then
echo "error: CHECKPOINT_PATH=$PATH2 is not a file"
exit 1
fi
ulimit -u unlimited
export DEVICE_NUM=1
export DEVICE_ID=0
export RANK_SIZE=$DEVICE_NUM
export RANK_ID=0
if [ -d "infer" ];
then
rm -rf ./infer
fi
mkdir ./infer
cp *.py ./infer
cp *.sh ./infer
cd ./infer || exit
env > env.log
echo "start infering for device $DEVICE_ID"
python eval.py --do_eval=True --dataset_path=$PATH1 --checkpoint_path=$PATH2 &> log &
cd ..

75
example/resnet101_imagenet2012/run_standalone_train.sh
View File

@ -1,75 +0,0 @@
#!/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 [ $# != 1 ] && [ $# != 2 ]
then
echo "Usage: sh run_standalone_train.sh [DATASET_PATH] [PRETRAINED_PATH](optional)"
exit 1
fi
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
PATH1=$(get_real_path $1)
echo $PATH1
if [ $# == 2 ]
then
PATH2=$(get_real_path $2)
echo $PATH2
fi
if [ ! -d $PATH1 ]
then
echo "error: DATASET_PATH=$PATH1 is not a directory"
exit 1
fi
if [ $# == 2 ] && [ ! -f $PATH2 ]
then
echo "error: PRETRAINED_PATH=$PATH2 is not a file"
exit 1
fi
ulimit -u unlimited
export DEVICE_NUM=1
export DEVICE_ID=0
export RANK_ID=0
export RANK_SIZE=1
if [ -d "train" ];
then
rm -rf ./train
fi
mkdir ./train
cp *.py ./train
cp *.sh ./train
cd ./train || exit
echo "start training for device $DEVICE_ID"
env > env.log
if [ $# == 1 ]
then
python train.py --do_train=True --dataset_path=$PATH1 &> log &
fi
if [ $# == 2 ]
then
python train.py --do_train=True --dataset_path=$PATH1 --pre_trained=$PATH2 &> log &
fi
cd ..

102
example/resnet101_imagenet2012/train.py
View File

@ -1,102 +0,0 @@
# 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.
# ============================================================================
"""train_imagenet."""
import os
import argparse
import random
import numpy as np
from dataset import create_dataset
from lr_generator import warmup_cosine_annealing_lr
from config import config
from mindspore import context
from mindspore import Tensor
from mindspore.model_zoo.resnet import resnet101
from mindspore.parallel._auto_parallel_context import auto_parallel_context
from mindspore.nn.optim.momentum import Momentum
from mindspore.train.model import Model, ParallelMode
from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor
from mindspore.train.loss_scale_manager import FixedLossScaleManager
from mindspore.train.serialization import load_checkpoint, load_param_into_net
import mindspore.dataset.engine as de
from mindspore.communication.management import init
import mindspore.nn as nn
import mindspore.common.initializer as weight_init
from crossentropy import CrossEntropy
random.seed(1)
np.random.seed(1)
de.config.set_seed(1)
parser = argparse.ArgumentParser(description='Image classification')
parser.add_argument('--run_distribute', type=bool, default=False, help='Run distribute')
parser.add_argument('--device_num', type=int, default=1, help='Device num.')
parser.add_argument('--do_train', type=bool, default=True, help='Do train or not.')
parser.add_argument('--do_eval', type=bool, default=False, help='Do eval or not.')
parser.add_argument('--dataset_path', type=str, default=None, help='Dataset path')
parser.add_argument('--pre_trained', type=str, default=None, help='Pretrained checkpoint path')
args_opt = parser.parse_args()
device_id = int(os.getenv('DEVICE_ID'))
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id,
enable_auto_mixed_precision=True)
if __name__ == '__main__':
if not args_opt.do_eval and args_opt.run_distribute:
context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True, parameter_broadcast=True)
auto_parallel_context().set_all_reduce_fusion_split_indices([180, 313])
init()
epoch_size = config.epoch_size
net = resnet101(class_num=config.class_num)
# weight init
for _, cell in net.cells_and_names():
if isinstance(cell, nn.Conv2d):
cell.weight.default_input = weight_init.initializer(weight_init.XavierUniform(),
cell.weight.default_input.shape(),
cell.weight.default_input.dtype()).to_tensor()
if isinstance(cell, nn.Dense):
cell.weight.default_input = weight_init.initializer(weight_init.TruncatedNormal(),
cell.weight.default_input.shape(),
cell.weight.default_input.dtype()).to_tensor()
if not config.label_smooth:
config.label_smooth_factor = 0.0
loss = CrossEntropy(smooth_factor=config.label_smooth_factor, num_classes=config.class_num)
if args_opt.do_train:
dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=True,
repeat_num=epoch_size, batch_size=config.batch_size)
step_size = dataset.get_dataset_size()
loss_scale = FixedLossScaleManager(config.loss_scale, drop_overflow_update=False)
if args_opt.pre_trained:
param_dict = load_checkpoint(args_opt.pre_trained)
load_param_into_net(net, param_dict)
# learning rate strategy with cosine
lr = Tensor(warmup_cosine_annealing_lr(config.lr, step_size, config.warmup_epochs, 120,
config.pretrain_epoch_size*step_size))
opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr, config.momentum,
config.weight_decay, config.loss_scale)
model = Model(net, loss_fn=loss, optimizer=opt, amp_level='O2', keep_batchnorm_fp32=False,
loss_scale_manager=loss_scale, metrics={'acc'})
time_cb = TimeMonitor(data_size=step_size)
loss_cb = LossMonitor()
cb = [time_cb, loss_cb]
if config.save_checkpoint:
config_ck = CheckpointConfig(save_checkpoint_steps=config.save_checkpoint_epochs*step_size,
keep_checkpoint_max=config.keep_checkpoint_max)
ckpt_cb = ModelCheckpoint(prefix="resnet", directory=config.save_checkpoint_path, config=config_ck)
cb += [ckpt_cb]
model.train(epoch_size, dataset, callbacks=cb)

1
model_zoo/resnet101/README.md
View File

@ -35,6 +35,7 @@ This is an example of training ResNet101 with ImageNet dataset in MindSpore.
├─crossentropy.py # CrossEntropy loss function
├─dataset.py # data preprocessin
├─lr_generator.py # generate learning rate
├─resnet101.py # resnet101 backbone
├─eval.py # eval net
└─train.py # train net
```

2
model_zoo/resnet101/eval.py
View File

@ -20,12 +20,12 @@ import argparse
import random
import numpy as np
from mindspore import context
from mindspore.model_zoo.resnet import resnet101
from mindspore.parallel._auto_parallel_context import auto_parallel_context
from mindspore.train.model import Model, ParallelMode
from mindspore.train.serialization import load_checkpoint, load_param_into_net
import mindspore.dataset.engine as de
from mindspore.communication.management import init
from src.resnet101 import resnet101
from src.dataset import create_dataset
from src.config import config
from src.crossentropy import CrossEntropy

261
model_zoo/resnet101/src/resnet101.py Executable file
View File

@ -0,0 +1,261 @@
# 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.
# ============================================================================
"""ResNet101."""
import numpy as np
import mindspore.nn as nn
from mindspore.ops import operations as P
from mindspore.common.tensor import Tensor
def _weight_variable(shape, factor=0.01):
init_value = np.random.randn(*shape).astype(np.float32) * factor
return Tensor(init_value)
def _conv3x3(in_channel, out_channel, stride=1):
weight_shape = (out_channel, in_channel, 3, 3)
weight = _weight_variable(weight_shape)
return nn.Conv2d(in_channel, out_channel,
kernel_size=3, stride=stride, padding=0, pad_mode='same', weight_init=weight)
def _conv1x1(in_channel, out_channel, stride=1):
weight_shape = (out_channel, in_channel, 1, 1)
weight = _weight_variable(weight_shape)
return nn.Conv2d(in_channel, out_channel,
kernel_size=1, stride=stride, padding=0, pad_mode='same', weight_init=weight)
def _conv7x7(in_channel, out_channel, stride=1):
weight_shape = (out_channel, in_channel, 7, 7)
weight = _weight_variable(weight_shape)
return nn.Conv2d(in_channel, out_channel,
kernel_size=7, stride=stride, padding=0, pad_mode='same', weight_init=weight)
def _bn(channel):
return nn.BatchNorm2d(channel, eps=1e-4, momentum=0.9,
gamma_init=1, beta_init=0, moving_mean_init=0, moving_var_init=1)
def _bn_last(channel):
return nn.BatchNorm2d(channel, eps=1e-4, momentum=0.9,
gamma_init=0, beta_init=0, moving_mean_init=0, moving_var_init=1)
def _fc(in_channel, out_channel):
weight_shape = (out_channel, in_channel)
weight = _weight_variable(weight_shape)
return nn.Dense(in_channel, out_channel, has_bias=True, weight_init=weight, bias_init=0)
class ResidualBlock(nn.Cell):
"""
ResNet V1 residual block definition.
Args:
in_channel (int): Input channel.
out_channel (int): Output channel.
stride (int): Stride size for the first convolutional layer. Default: 1.
Returns:
Tensor, output tensor.
Examples:
>>> ResidualBlock(3, 256, stride=2)
"""
expansion = 4
def __init__(self,
in_channel,
out_channel,
stride=1):
super(ResidualBlock, self).__init__()
channel = out_channel // self.expansion
self.conv1 = _conv1x1(in_channel, channel, stride=1)
self.bn1 = _bn(channel)
self.conv2 = _conv3x3(channel, channel, stride=stride)
self.bn2 = _bn(channel)
self.conv3 = _conv1x1(channel, out_channel, stride=1)
self.bn3 = _bn_last(out_channel)
self.relu = nn.ReLU()
self.down_sample = False
if stride != 1 or in_channel != out_channel:
self.down_sample = True
self.down_sample_layer = None
if self.down_sample:
self.down_sample_layer = nn.SequentialCell([_conv1x1(in_channel, out_channel, stride),
_bn(out_channel)])
self.add = P.TensorAdd()
def construct(self, x):
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
if self.down_sample:
identity = self.down_sample_layer(identity)
out = self.add(out, identity)
out = self.relu(out)
return out
class ResNet(nn.Cell):
"""
ResNet architecture.
Args:
block (Cell): Block for network.
layer_nums (list): Numbers of block in different layers.
in_channels (list): Input channel in each layer.
out_channels (list): Output channel in each layer.
strides (list): Stride size in each layer.
num_classes (int): The number of classes that the training images are belonging to.
Returns:
Tensor, output tensor.
Examples:
>>> ResNet(ResidualBlock,
>>> [3, 4, 6, 3],
>>> [64, 256, 512, 1024],
>>> [256, 512, 1024, 2048],
>>> [1, 2, 2, 2],
>>> 10)
"""
def __init__(self,
block,
layer_nums,
in_channels,
out_channels,
strides,
num_classes):
super(ResNet, self).__init__()
if not len(layer_nums) == len(in_channels) == len(out_channels) == 4:
raise ValueError("the length of layer_num, in_channels, out_channels list must be 4!")
self.conv1 = _conv7x7(3, 64, stride=2)
self.bn1 = _bn(64)
self.relu = P.ReLU()
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, pad_mode="same")
self.layer1 = self._make_layer(block,
layer_nums[0],
in_channel=in_channels[0],
out_channel=out_channels[0],
stride=strides[0])
self.layer2 = self._make_layer(block,
layer_nums[1],
in_channel=in_channels[1],
out_channel=out_channels[1],
stride=strides[1])
self.layer3 = self._make_layer(block,
layer_nums[2],
in_channel=in_channels[2],
out_channel=out_channels[2],
stride=strides[2])
self.layer4 = self._make_layer(block,
layer_nums[3],
in_channel=in_channels[3],
out_channel=out_channels[3],
stride=strides[3])
self.mean = P.ReduceMean(keep_dims=True)
self.flatten = nn.Flatten()
self.end_point = _fc(out_channels[3], num_classes)
def _make_layer(self, block, layer_num, in_channel, out_channel, stride):
"""
Make stage network of ResNet.
Args:
block (Cell): Resnet block.
layer_num (int): Layer number.
in_channel (int): Input channel.
out_channel (int): Output channel.
stride (int): Stride size for the first convolutional layer.
Returns:
SequentialCell, the output layer.
Examples:
>>> _make_layer(ResidualBlock, 3, 128, 256, 2)
"""
layers = []
resnet_block = block(in_channel, out_channel, stride=stride)
layers.append(resnet_block)
for _ in range(1, layer_num):
resnet_block = block(out_channel, out_channel, stride=1)
layers.append(resnet_block)
return nn.SequentialCell(layers)
def construct(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
c1 = self.maxpool(x)
c2 = self.layer1(c1)
c3 = self.layer2(c2)
c4 = self.layer3(c3)
c5 = self.layer4(c4)
out = self.mean(c5, (2, 3))
out = self.flatten(out)
out = self.end_point(out)
return out
def resnet101(class_num=1001):
"""
Get ResNet101 neural network.
Args:
class_num (int): Class number.
Returns:
Cell, cell instance of ResNet101 neural network.
Examples:
>>> net = resnet101(1001)
"""
return ResNet(ResidualBlock,
[3, 4, 23, 3],
[64, 256, 512, 1024],
[256, 512, 1024, 2048],
[1, 2, 2, 2],
class_num)

2
model_zoo/resnet101/train.py
View File

@ -19,7 +19,6 @@ import random
import numpy as np
from mindspore import context
from mindspore import Tensor
from mindspore.model_zoo.resnet import resnet101
from mindspore.parallel._auto_parallel_context import auto_parallel_context
from mindspore.nn.optim.momentum import Momentum
from mindspore.train.model import Model, ParallelMode
@ -30,6 +29,7 @@ import mindspore.dataset.engine as de
from mindspore.communication.management import init
import mindspore.nn as nn
import mindspore.common.initializer as weight_init
from src.resnet101 import resnet101
from src.dataset import create_dataset
from src.lr_generator import warmup_cosine_annealing_lr
from src.config import config