diff --git a/model_zoo/research/cv/SE-Net/README.md b/model_zoo/research/cv/SE-Net/README.md new file mode 100644 index 0000000000..eb49583522 --- /dev/null +++ b/model_zoo/research/cv/SE-Net/README.md @@ -0,0 +1,245 @@ +# Contents + +- [SE-Net Description](#se-net-description) +- [Model Architecture](#model-architecture) +- [Dataset](#dataset) +- [Features](#features) + - [Mixed Precision](#mixed-precision) +- [Environment Requirements](#environment-requirements) +- [Quick Start](#quick-start) +- [Script Description](#script-description) + - [Script and Sample Code](#script-and-sample-code) + - [Script Parameters](#script-parameters) + - [Training Process](#training-process) + - [Evaluation Process](#evaluation-process) +- [Model Description](#model-description) + - [Performance](#performance) + - [Evaluation Performance](#evaluation-performance) + - [Inference Performance](#inference-performance) +- [Description of Random Situation](#description-of-random-situation) +- [ModelZoo Homepage](#modelzoo-homepage) + +# [SE-Net Description](#contents) + +## Description + +something should be written here. + +## Paper + +[paper](https://arxiv.org/abs/1709.01507):Jie Hu, Li Shen, Samuel Albanie, Gang Sun, Enhua Wu. "Squeeze-and-Excitation Networks" + +# [Model Architecture](#contents) + +The overall network architecture of Net is show below: +[Link](https://arxiv.org/pdf/1512.03385.pdf) + +# [Dataset](#contents) + +Dataset used: [ImageNet2012](http://www.image-net.org/) + +- Dataset size 224*224 colorful images in 1000 classes + - Train：1,281,167 images + - Test： 50,000 images +- Data format：jpeg + - Note：Data will be processed in dataset.py + +# [Features](#contents) + +## Mixed Precision + +The [mixed precision](https://www.mindspore.cn/tutorial/training/en/master/advanced_use/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. + +# [Environment Requirements](#contents) + +- Hardware（Ascend） + - Prepare hardware environment with Ascend processor. +- Framework + - [MindSpore](https://www.mindspore.cn/install/en) +- For more information, please check the resources below： + - [MindSpore Tutorials](https://www.mindspore.cn/tutorial/training/en/master/index.html) + - [MindSpore Python API](https://www.mindspore.cn/doc/api_python/en/master/index.html) + +# [Quick Start](#contents) + +After installing MindSpore via the official website, you can start training and evaluation as follows: + +- Running on Ascend + +```bash +# distributed training +Usage: +sh run_distribute_train.sh se-resnet50 imagenet2012 [RANK_TABLE_FILE] [DATASET_PATH] + +# standalone training +Usage: +export DEVICE_ID=0 +python train.py --net=se-resnet50 --dataset=imagenet2012 --dataset_path=[DATASET_PATH] + +# run evaluation example +Usage: +export DEVICE_ID=0 +python eval.py --net=se-resnet50 --dataset=imagenet2012 --checkpoint_path=[CHECKPOINT_PATH] --dataset_path=[DATASET_PATH] +``` + +# [Script Description](#contents) + +## [Script and Sample Code](#contents) + +```shell +. +└──SE-Net + ├── README.md + ├── scripts + ├── run_distribute_train.sh # launch ascend distributed training(8 pcs) + ├── run_eval.sh # launch ascend evaluation + ├── run_standalone_train.sh # launch ascend standalone training(1 pcs) + ├── src + ├── config.py # parameter configuration + ├── CrossEntropySmooth.py # loss definition for ImageNet2012 dataset + ├── dataset.py # data preprocessing + ├── lr_generator.py # generate learning rate for each step + ├── resnet.py # resnet50 backbone + └── se.py # se-block definition + ├── export.py # export model for inference + ├── eval.py # eval net + └── train.py # train net +``` + +## [Script Parameters](#contents) + +Parameters for both training and evaluation can be set in config.py. + +- Config for SE-ResNet50, ImageNet2012 dataset + +```bash +"class_num": 1001, # dataset class number +"batch_size": 256, # batch size of input tensor +"loss_scale": 1024, # loss scale +"momentum": 0.9, # momentum optimizer +"weight_decay": 1e-4, # weight decay +"epoch_size": 90, # only valid for taining, which is always 1 for inference +"pretrain_epoch_size": 0, # epoch size that model has been trained before loading pretrained checkpoint, actual training epoch size is equal to epoch_size minus pretrain_epoch_size +"save_checkpoint": True, # whether save checkpoint or not +"save_checkpoint_epochs": 5, # 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": "linear", # decay mode for generating learning rate +"use_label_smooth": True, # label smooth +"label_smooth_factor": 0.1, # label smooth factor +"lr_init": 0, # initial learning rate +"lr_max": 0.8, # maximum learning rate +"lr_end": 0.0, # minimum learning rate +``` + +## [Training Process](#contents) + +### Usage + +#### Running on Ascend + +```bash +# distributed training +Usage: +bash run_distribute_train.sh se-resnet50 imagenet2012 /imagenet/train /rank_table.json + +# standalone training +Usage: +export DEVICE_ID=0 +bash run_standalone_train.sh se-resnet50 imagenet2012 /data/imagenet/train/ +``` + +For distributed training, a hccl configuration file with JSON format needs to be created in advance. + +Please follow the instructions in the link [hccn_tools](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools). + +Training result will be stored in the example path, whose folder name begins with "train" or "train_parallel". Under this, you can find checkpoint file together with result like the following in log. + +### Result + +- Training SE-ResNet50 with ImageNet2012 dataset + +```bash +# distribute training result(8 pcs) +epoch: 1 step: 625, loss is 5.0938025 +epoch time: 303139.271 ms, per step time: 485.023 ms +epoch: 2 step: 625, loss is 4.152817 +epoch time: 205321.853 ms, per step time: 328.515 ms +epoch: 3 step: 625, loss is 3.7530446 +epoch time: 205214.637 ms, per step time: 328.343 ms +... +epoch: 89 step: 625, loss is 1.9109731 +epoch time: 205217.996 ms, per step time: 328.349 ms +epoch: 90 step: 625, loss is 1.5931969 +epoch time: 206295.838 ms, per step time: 330.073 ms +``` + +## [Evaluation Process](#contents) + +### Usage + +#### Running on Ascend + +```bash +export DEVICE_ID=0 +bash run_eval.sh /imagenet/val/ /path/to/resnet-90_625.ckpt +``` + +### Result + +- Evaluating SE-ResNet50 with ImageNet2012 dataset + +```bash +result: {'top_5_accuracy': 0.9385269007731959, 'top_1_accuracy': 0.7774645618556701} +``` + +# [Model Description](#contents) + +## [Performance](#contents) + +### Evaluation Performance + +#### SE-ResNet50 on ImageNet2012 + +| Parameters | Ascend 910 +| -------------------------- | ------------------------------------------------------------------------ | +| Model Version | SE-ResNet50 | +| Resource | Ascend 910，CPU 2.60GHz 192cores，Memory 755G | +| uploaded Date | 03/19/2021 (month/day/year) | +| MindSpore Version | 0.7.0-alpha | +| Dataset | ImageNet2012 | +| Training Parameters | epoch=90, steps per epoch=5004, batch_size = 256 | +| Optimizer | Momentum | +| Loss Function | Softmax Cross Entropy | +| outputs | probability | +| Loss | 1.5931969 | +| Speed | # ms/step（8pcs） | +| Total time | # mins | +| Parameters (M) | 285M | +| Checkpoint for Fine tuning | # M (.ckpt file) | +| Scripts | [Link](XXXXXXXhttps://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/resnet) | + +### Inference Performance + +#### SE-ResNet50 on ImageNet2012 + +| Parameters | Ascend | +| ------------------- | --------------------------- | +| Model Version | SE-ResNet50 | +| Resource | Ascend 910 | +| Uploaded Date | 03/19/2021 (month/day/year) | +| MindSpore Version | 0.7.0-alpha | +| Dataset | ImageNet2012 | +| batch_size | 256 | +| Accuracy | 77.74% | +| Model for inference | # (.air file) | + +# [Description of Random Situation](#contents) + +In dataset.py, we set the seed inside "create_dataset" function. We also use random seed in train.py. + +# [ModelZoo Homepage](#contents) + + Please check the official [homepage](https://gitee.com/mindspore/mindspore/tree/master/model_zoo). diff --git a/model_zoo/research/cv/SE-Net/eval.py b/model_zoo/research/cv/SE-Net/eval.py new file mode 100644 index 0000000000..8b19a05fe1 --- /dev/null +++ b/model_zoo/research/cv/SE-Net/eval.py @@ -0,0 +1,67 @@ +# Copyright 2021 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 net""" +import os +import argparse +from mindspore import context +from mindspore.common import set_seed +from mindspore.train.model import Model +from mindspore.train.serialization import load_checkpoint, load_param_into_net +from src.CrossEntropySmooth import CrossEntropySmooth +from src.resnet import se_resnet50 as resnet +from src.config import config2 as config +from src.dataset import create_dataset2 as create_dataset + +parser = argparse.ArgumentParser(description='Image classification') +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') +parser.add_argument('--device_target', type=str, default='Ascend', choices=("Ascend", "GPU", "CPU"), + help="Device target, support Ascend, GPU and CPU.") +args_opt = parser.parse_args() +set_seed(1) + +if __name__ == '__main__': + target = args_opt.device_target + + # init context + context.set_context(mode=context.GRAPH_MODE, device_target=target, save_graphs=False) + if target == "Ascend": + device_id = int(os.getenv('DEVICE_ID')) + context.set_context(device_id=device_id) + + # create dataset + dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=False, batch_size=config.batch_size, + target=target) + step_size = dataset.get_dataset_size() + + # define net + net = resnet(class_num=config.class_num) + + # load checkpoint + param_dict = load_checkpoint(args_opt.checkpoint_path) + load_param_into_net(net, param_dict) + net.set_train(False) + + # define loss, model + if not config.use_label_smooth: + config.label_smooth_factor = 0.0 + loss = CrossEntropySmooth(sparse=True, reduction='mean', + smooth_factor=config.label_smooth_factor, num_classes=config.class_num) + # define model + model = Model(net, loss_fn=loss, metrics={'top_1_accuracy', 'top_5_accuracy'}) + + # eval model + res = model.eval(dataset) + print("result:", res, "ckpt=", args_opt.checkpoint_path) diff --git a/model_zoo/research/cv/SE-Net/export.py b/model_zoo/research/cv/SE-Net/export.py new file mode 100644 index 0000000000..4b1785e2f6 --- /dev/null +++ b/model_zoo/research/cv/SE-Net/export.py @@ -0,0 +1,50 @@ +# Copyright 2021 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. +# ============================================================================ +""" +##############export checkpoint file into air and onnx models################# +python export.py +""" +import argparse +import numpy as np + +from mindspore import Tensor, load_checkpoint, load_param_into_net, export, context +from src.config import config2 as config +from src.resnet import se_resnet50 as resnet + +parser = argparse.ArgumentParser(description='resnet export') +parser.add_argument('--network_dataset', type=str, default="se-resnet50", choices=["se-resnet50"], + help='network and dataset name.') +parser.add_argument("--device_id", type=int, default=0, help="Device id") +parser.add_argument("--batch_size", type=int, default=1, help="batch size") +parser.add_argument("--ckpt_file", type=str, required=True, help="Checkpoint file path.") +parser.add_argument("--file_name", type=str, default="resnet", help="output file name.") +parser.add_argument('--width', type=int, default=224, help='input width') +parser.add_argument('--height', type=int, default=224, help='input height') +parser.add_argument("--file_format", type=str, choices=["AIR", "ONNX", "MINDIR"], default="AIR", help="file format") +parser.add_argument("--device_target", type=str, default="Ascend", + choices=["Ascend", "GPU", "CPU"], help="device target(default: Ascend)") +args = parser.parse_args() + +context.set_context(mode=context.GRAPH_MODE, device_target="Ascend") +if args.device_target == "Ascend": + context.set_context(device_id=args.device_id) + +if __name__ == '__main__': + net = resnet(config.class_num) + param_dict = load_checkpoint(args.ckpt_file) + load_param_into_net(net, param_dict) + + input_arr = Tensor(np.zeros([args.batch_size, 3, args.height, args.width], np.float32)) + export(net, input_arr, file_name=args.file_name, file_format=args.file_format) diff --git a/model_zoo/research/cv/SE-Net/scripts/run_distribute_train.sh b/model_zoo/research/cv/SE-Net/scripts/run_distribute_train.sh new file mode 100644 index 0000000000..701167c295 --- /dev/null +++ b/model_zoo/research/cv/SE-Net/scripts/run_distribute_train.sh @@ -0,0 +1,48 @@ +#!/bin/bash +# Copyright 2021 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. +# ============================================================================ + +#Usage: sh run_distribute_train.sh [se-resnet50] [cifar10|imagenet2012] [RANK_TABLE_FILE] [DATASET_PATH] [PRETRAINED_CKPT_PATH](optional) + + +ulimit -u unlimited +export DEVICE_NUM=8 +export RANK_SIZE=8 +#export RANK_TABLE_FILE=/home/wks/hccl_8p_01234567_127.0.0.1.json +export NET=$1 +export DATASET=$2 +export DATASET_PATH=$3 +export RANK_TABLE_FILE=$4 + +export SERVER_ID=0 +rank_start=$((DEVICE_NUM * SERVER_ID)) + +for((i=0; i<${DEVICE_NUM}; i++)) +do + export DEVICE_ID=${i} + export RANK_ID=$((rank_start + i)) + rm -rf ./train_parallel$i + mkdir ./train_parallel$i + cp ../*.py ./train_parallel$i + cp *.sh ./train_parallel$i + cp -r ../src ./train_parallel$i + cd ./train_parallel$i || exit + echo "start training for rank $RANK_ID, device $DEVICE_ID" + env > env.log + python train.py --net=$NET --dataset=$DATASET --run_distribute=True --device_num=$DEVICE_NUM --dataset_path=$DATASET_PATH &> log & + #python train.py --net=se-resnet50 --dataset=imagenet2012 --dataset_path=/data/imagenet/train # single one + + cd .. +done diff --git a/model_zoo/research/cv/SE-Net/scripts/run_eval.sh b/model_zoo/research/cv/SE-Net/scripts/run_eval.sh new file mode 100644 index 0000000000..24d2c97afd --- /dev/null +++ b/model_zoo/research/cv/SE-Net/scripts/run_eval.sh @@ -0,0 +1,46 @@ +#!/bin/bash +# Copyright 2021 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 [ $# != 2 ] +then + echo "Usage: sh run_eval.sh [DATASET_PATH] [CHECKPOINT_PATH]" +exit 1 +fi + +ulimit -u unlimited +export DEVICE_NUM=1 +export DEVICE_ID=0 +export RANK_SIZE=$DEVICE_NUM +export RANK_ID=0 +export DATA_PATH=$1 +export CKPT_PATH=$2 + +if [ -d "eval" ]; +then + rm -rf ./eval +fi +mkdir ./eval +cp ../*.py ./eval +cp *.sh ./eval +cp -r ../src ./eval +cd ./eval || exit +env > env.log +echo "start evaluation for device $DEVICE_ID" + +#python eval.py --dataset_path /data/imagenet/val/ --checkpoint_path /home/wks/train8s/train_paralleltrain_parallel0/resnet-90_625.ckpt | tee eval.log # export ID=0 +python eval.py --dataset_path=$DATA_PATH --checkpoint_path=$CKPT_PATH &> log & + +cd .. diff --git a/model_zoo/research/cv/SE-Net/scripts/run_standalone_train.sh b/model_zoo/research/cv/SE-Net/scripts/run_standalone_train.sh new file mode 100644 index 0000000000..bab61cda1c --- /dev/null +++ b/model_zoo/research/cv/SE-Net/scripts/run_standalone_train.sh @@ -0,0 +1,47 @@ +#!/bin/bash +# Copyright 2021 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 [ $# != 3 ] +then + echo "Usage: sh run_standalone_train.sh [NET] [DATASET_NAME] [DATASET_PATH]" +exit 1 +fi + +ulimit -u unlimited +export DEVICE_NUM=1 +export DEVICE_ID=0 +export RANK_ID=0 +export RANK_SIZE=1 +export NET=$1 +export DATASET=$2 +export DATASET_PATH=$3 + + +if [ -d "train" ]; +then + rm -rf ./train +fi + +mkdir ./train +cp ../*.py ./train +cp *.sh ./train +cp -r ../src ./train +cd ./train || exit +echo "start training for device $DEVICE_ID" +env > env.log +# python train.py --net=se-resnet50 --dataset=imagenet2012 --dataset_path=/data/imagenet/train/ +python train.py --net=$NET --dataset=$DATASET --dataset_path=$DATASET_PATH &> log & +cd .. diff --git a/model_zoo/research/cv/SE-Net/src/CrossEntropySmooth.py b/model_zoo/research/cv/SE-Net/src/CrossEntropySmooth.py new file mode 100644 index 0000000000..6d63b66694 --- /dev/null +++ b/model_zoo/research/cv/SE-Net/src/CrossEntropySmooth.py @@ -0,0 +1,38 @@ +# Copyright 2021 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""" +import mindspore.nn as nn +from mindspore import Tensor +from mindspore.common import dtype as mstype +from mindspore.nn.loss.loss import _Loss +from mindspore.ops import functional as F +from mindspore.ops import operations as P + + +class CrossEntropySmooth(_Loss): + """CrossEntropy""" + def __init__(self, sparse=True, reduction='mean', smooth_factor=0., num_classes=1000): + super(CrossEntropySmooth, self).__init__() + self.onehot = P.OneHot() + self.sparse = sparse + 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(reduction=reduction) + + def construct(self, logit, label): + if self.sparse: + label = self.onehot(label, F.shape(logit)[1], self.on_value, self.off_value) + loss = self.ce(logit, label) + return loss diff --git a/model_zoo/research/cv/SE-Net/src/__init__.py b/model_zoo/research/cv/SE-Net/src/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/model_zoo/research/cv/SE-Net/src/config.py b/model_zoo/research/cv/SE-Net/src/config.py new file mode 100644 index 0000000000..28434a643b --- /dev/null +++ b/model_zoo/research/cv/SE-Net/src/config.py @@ -0,0 +1,39 @@ +# Copyright 2021 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 for se-resnet50, imagenet2012 +config2 = ed({ + "class_num": 1001, + "batch_size": 256, + "loss_scale": 1024, + "momentum": 0.9, + "weight_decay": 1e-4, + "epoch_size": 90, + "pretrain_epoch_size": 0, + "save_checkpoint": True, + "save_checkpoint_epochs": 5, + "keep_checkpoint_max": 10, + "save_checkpoint_path": "./", + "warmup_epochs": 0, + "lr_decay_mode": "linear", + "use_label_smooth": True, + "label_smooth_factor": 0.1, + "lr_init": 0, + "lr_max": 0.8, + "lr_end": 0.0 +}) diff --git a/model_zoo/research/cv/SE-Net/src/dataset.py b/model_zoo/research/cv/SE-Net/src/dataset.py new file mode 100644 index 0000000000..0b8671e277 --- /dev/null +++ b/model_zoo/research/cv/SE-Net/src/dataset.py @@ -0,0 +1,104 @@ +# Copyright 2021 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 as ds +import mindspore.dataset.vision.c_transforms as C +import mindspore.dataset.transforms.c_transforms as C2 +from mindspore.communication.management import init, get_rank, get_group_size + +def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False): + """ + create a train or eval imagenet2012 dataset for resnet50 + + 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 + target(str): the device target. Default: Ascend + distribute(bool): data for distribute or not. Default: False + + Returns: + dataset + """ + if target == "Ascend": + device_num, rank_id = _get_rank_info() + else: + if distribute: + init() + rank_id = get_rank() + device_num = get_group_size() + else: + device_num = 1 + + if device_num == 1: + data_set = ds.ImageFolderDataset(dataset_path, num_parallel_workers=8, shuffle=True) + else: + data_set = ds.ImageFolderDataset(dataset_path, num_parallel_workers=8, shuffle=True, + num_shards=device_num, shard_id=rank_id) + + image_size = 224 + mean = [0.485 * 255, 0.456 * 255, 0.406 * 255] + std = [0.229 * 255, 0.224 * 255, 0.225 * 255] + + # define map operations + if do_train: + trans = [ + C.RandomCropDecodeResize(image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)), + C.RandomHorizontalFlip(prob=0.5), + C.Normalize(mean=mean, std=std), + C.HWC2CHW() + ] + else: + trans = [ + C.Decode(), + C.Resize(256), + C.CenterCrop(image_size), + C.Normalize(mean=mean, std=std), + C.HWC2CHW() + ] + + type_cast_op = C2.TypeCast(mstype.int32) + + data_set = data_set.map(operations=trans, input_columns="image", num_parallel_workers=8) + data_set = data_set.map(operations=type_cast_op, input_columns="label", num_parallel_workers=8) + + # apply batch operations + data_set = data_set.batch(batch_size, drop_remainder=True) + + # apply dataset repeat operation + data_set = data_set.repeat(repeat_num) + + return data_set + + +def _get_rank_info(): + """ + get rank size and rank id + """ + rank_size = int(os.environ.get("RANK_SIZE", 1)) + + if rank_size > 1: + rank_size = get_group_size() + rank_id = get_rank() + else: + rank_size = 1 + rank_id = 0 + + return rank_size, rank_id diff --git a/model_zoo/research/cv/SE-Net/src/lr_generator.py b/model_zoo/research/cv/SE-Net/src/lr_generator.py new file mode 100644 index 0000000000..b096f2945e --- /dev/null +++ b/model_zoo/research/cv/SE-Net/src/lr_generator.py @@ -0,0 +1,207 @@ +# Copyright 2021 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 _generate_steps_lr(lr_init, lr_max, total_steps, warmup_steps): + """ + Applies three steps decay to generate learning rate array. + + Args: + lr_init(float): init learning rate. + lr_max(float): max learning rate. + total_steps(int): all steps in training. + warmup_steps(int): all steps in warmup epochs. + + Returns: + np.array, learning rate array. + """ + decay_epoch_index = [0.3 * total_steps, 0.6 * total_steps, 0.8 * total_steps] + lr_each_step = [] + for i in range(total_steps): + if i < warmup_steps: + lr = lr_init + (lr_max - lr_init) * i / warmup_steps + else: + if i < decay_epoch_index[0]: + lr = lr_max + elif i < decay_epoch_index[1]: + lr = lr_max * 0.1 + elif i < decay_epoch_index[2]: + lr = lr_max * 0.01 + else: + lr = lr_max * 0.001 + lr_each_step.append(lr) + return lr_each_step + + +def _generate_poly_lr(lr_init, lr_end, lr_max, total_steps, warmup_steps): + """ + Applies polynomial decay to generate learning rate array. + + Args: + lr_init(float): init learning rate. + lr_end(float): end learning rate + lr_max(float): max learning rate. + total_steps(int): all steps in training. + warmup_steps(int): all steps in warmup epochs. + + Returns: + np.array, learning rate array. + """ + lr_each_step = [] + if warmup_steps != 0: + inc_each_step = (float(lr_max) - float(lr_init)) / float(warmup_steps) + else: + inc_each_step = 0 + for i in range(total_steps): + if i < warmup_steps: + lr = float(lr_init) + inc_each_step * float(i) + else: + base = (1.0 - (float(i) - float(warmup_steps)) / (float(total_steps) - float(warmup_steps))) + lr = float(lr_max) * base * base + if lr < 0.0: + lr = 0.0 + lr_each_step.append(lr) + return lr_each_step + + +def _generate_cosine_lr(lr_init, lr_end, lr_max, total_steps, warmup_steps): + """ + Applies cosine decay to generate learning rate array. + + Args: + lr_init(float): init learning rate. + lr_end(float): end learning rate + lr_max(float): max learning rate. + total_steps(int): all steps in training. + warmup_steps(int): all steps in warmup epochs. + + Returns: + np.array, learning rate array. + """ + decay_steps = total_steps - warmup_steps + lr_each_step = [] + for i in range(total_steps): + if i < warmup_steps: + lr_inc = (float(lr_max) - float(lr_init)) / float(warmup_steps) + lr = float(lr_init) + lr_inc * (i + 1) + 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 = lr_max * decayed + lr_each_step.append(lr) + return lr_each_step + + +def _generate_liner_lr(lr_init, lr_end, lr_max, total_steps, warmup_steps): + """ + Applies liner decay to generate learning rate array. + + Args: + lr_init(float): init learning rate. + lr_end(float): end learning rate + lr_max(float): max learning rate. + total_steps(int): all steps in training. + warmup_steps(int): all steps in warmup epochs. + + Returns: + np.array, learning rate array. + """ + lr_each_step = [] + for i in range(total_steps): + if i < warmup_steps: + lr = lr_init + (lr_max - lr_init) * i / warmup_steps + else: + lr = lr_max - (lr_max - lr_end) * (i - warmup_steps) / (total_steps - warmup_steps) + lr_each_step.append(lr) + return lr_each_step + + + +def get_lr(lr_init, lr_end, lr_max, warmup_epochs, total_epochs, steps_per_epoch, lr_decay_mode): + """ + generate learning rate array + + Args: + lr_init(float): init learning rate + lr_end(float): end learning rate + lr_max(float): max learning rate + warmup_epochs(int): number of warmup epochs + total_epochs(int): total epoch of training + steps_per_epoch(int): steps of one epoch + lr_decay_mode(string): learning rate decay mode, including steps, poly, cosine or liner(default) + + Returns: + np.array, learning rate array + """ + lr_each_step = [] + total_steps = steps_per_epoch * total_epochs + warmup_steps = steps_per_epoch * warmup_epochs + + if lr_decay_mode == 'steps': + lr_each_step = _generate_steps_lr(lr_init, lr_max, total_steps, warmup_steps) + elif lr_decay_mode == 'poly': + lr_each_step = _generate_poly_lr(lr_init, lr_end, lr_max, total_steps, warmup_steps) + elif lr_decay_mode == 'cosine': + lr_each_step = _generate_cosine_lr(lr_init, lr_end, lr_max, total_steps, warmup_steps) + else: + lr_each_step = _generate_liner_lr(lr_init, lr_end, lr_max, total_steps, warmup_steps) + + lr_each_step = np.array(lr_each_step).astype(np.float32) + return lr_each_step + + +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 diff --git a/model_zoo/research/cv/SE-Net/src/resnet.py b/model_zoo/research/cv/SE-Net/src/resnet.py new file mode 100644 index 0000000000..e942d2ba4f --- /dev/null +++ b/model_zoo/research/cv/SE-Net/src/resnet.py @@ -0,0 +1,265 @@ +# Copyright 2021 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. +# ============================================================================ +"""Se_ResNet.""" +import numpy as np +import mindspore.nn as nn +import mindspore.common.dtype as mstype +from mindspore.ops import operations as P +from mindspore.common.tensor import Tensor +from scipy.stats import truncnorm +from .se import SELayer + +def _conv_variance_scaling_initializer(in_channel, out_channel, kernel_size): + fan_in = in_channel * kernel_size * kernel_size + scale = 1.0 + scale /= max(1., fan_in) + stddev = (scale ** 0.5) / .87962566103423978 + mu, sigma = 0, stddev + weight = truncnorm(-2, 2, loc=mu, scale=sigma).rvs(out_channel * in_channel * kernel_size * kernel_size) + weight = np.reshape(weight, (out_channel, in_channel, kernel_size, kernel_size)) + return Tensor(weight, dtype=mstype.float32) + +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 Se_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. + use_se (bool): enable SE-ResNet50 net. Default: False. + se_block(bool): use se block in SE-ResNet50 net. Default: False. + + Returns: + Tensor, output tensor. + + """ + expansion = 4 + + def __init__(self, + in_channel, + out_channel, + stride=1, + reduction=16): + super(Se_ResidualBlock, self).__init__() + self.stride = stride + 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)])#use_se=self.use_se + self.add = P.TensorAdd() + self.se = SELayer(out_channel, reduction) + + def construct(self, x): + """se_block""" + 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) + + out = self.se(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. + use_se (bool): enable SE-ResNet50 net. Default: False. + se_block(bool): use se block in SE-ResNet50 net in layer 3 and layer 4. Default: False. + Returns: + Tensor, output tensor. + + """ + + 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. + se_block(bool): use se block in SE-ResNet50 net. Default: False. + Returns: + SequentialCell, the output layer. + + """ + 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): + """construct network""" + 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 se_resnet50(class_num=10): + """ + Get ResNet50 neural network. + + Args: + class_num (int): Class number. + + Returns: + Cell, cell instance of ResNet50 neural network. + + """ + return ResNet(Se_ResidualBlock, + [3, 4, 6, 3], + [64, 256, 512, 1024], + [256, 512, 1024, 2048], + [1, 2, 2, 2], + class_num) diff --git a/model_zoo/research/cv/SE-Net/src/se.py b/model_zoo/research/cv/SE-Net/src/se.py new file mode 100644 index 0000000000..bfcea87a5f --- /dev/null +++ b/model_zoo/research/cv/SE-Net/src/se.py @@ -0,0 +1,53 @@ +# Copyright 2021 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. +# ============================================================================ +"""SE Module""" +from mindspore.ops import operations as P +from mindspore import Tensor + +import mindspore.nn as nn +from mindspore.ops import functional as F +import numpy as np + +def _weight_variable(shape, factor=0.01): + init_value = np.random.randn(*shape).astype(np.float32) * factor + return Tensor(init_value) + +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 SELayer(nn.Cell): + """SE Layer""" + def __init__(self, out_channel, reduction=16): + super(SELayer, self).__init__() + self.se_global_pool = P.ReduceMean(keep_dims=False) + self.se_dense_0 = _fc(out_channel, int(out_channel / 4)) + self.se_dense_1 = _fc(int(out_channel / 4), out_channel) + self.se_sigmoid = nn.Sigmoid() + self.relu = nn.ReLU() + self.se_mul = P.Mul() + + def construct(self, out): + out_se = out + out = self.se_global_pool(out, (2, 3)) + out = self.se_dense_0(out) + out = self.relu(out) + out = self.se_dense_1(out) + out = self.se_sigmoid(out) + out = F.reshape(out, F.shape(out) + (1, 1)) + out = self.se_mul(out, out_se) + return out diff --git a/model_zoo/research/cv/SE-Net/train.py b/model_zoo/research/cv/SE-Net/train.py new file mode 100644 index 0000000000..1b1613e9d3 --- /dev/null +++ b/model_zoo/research/cv/SE-Net/train.py @@ -0,0 +1,145 @@ +# Copyright 2021 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 net.""" +import os +import argparse +import ast +from mindspore import context +from mindspore import Tensor +from mindspore.nn.optim.momentum import Momentum +from mindspore.train.model import Model +from mindspore.context import 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 +from mindspore.communication.management import init +from mindspore.common import set_seed +from mindspore.parallel import set_algo_parameters +import mindspore.nn as nn +import mindspore.common.initializer as weight_init +from src.lr_generator import get_lr +from src.CrossEntropySmooth import CrossEntropySmooth + +parser = argparse.ArgumentParser(description='Image classification') +parser.add_argument('--net', type=str, default=None, help='Resnet Model, either resnet50 or resnet101') +parser.add_argument('--dataset', type=str, default="cifar10", help='Dataset, either cifar10 or imagenet2012') +parser.add_argument('--run_distribute', type=ast.literal_eval, default=False, help='Run distribute') +parser.add_argument('--device_num', type=int, default=1, help='Device num.') + +parser.add_argument('--dataset_path', type=str, default=None, help='Dataset path') +parser.add_argument('--device_target', type=str, default='Ascend', choices=("Ascend", "GPU", "CPU"), + help="Device target, support Ascend, GPU and CPU.") +parser.add_argument('--pre_trained', type=str, default=None, help='Pretrained checkpoint path') +parser.add_argument('--parameter_server', type=ast.literal_eval, default=False, help='Run parameter server train') +args_opt = parser.parse_args() + +set_seed(1) + +if args_opt.net == "se-resnet50": + from src.resnet import se_resnet50 as resnet + from src.config import config2 as config + from src.dataset import create_dataset2 as create_dataset + +if __name__ == '__main__': + target = args_opt.device_target + ckpt_save_dir = config.save_checkpoint_path + + # init context + context.set_context(mode=context.GRAPH_MODE, device_target=target, save_graphs=False) + if args_opt.parameter_server: + context.set_ps_context(enable_ps=True) + if args_opt.run_distribute: + if target == "Ascend": + device_id = int(os.getenv('DEVICE_ID')) + context.set_context(device_id=device_id, enable_auto_mixed_precision=True) + context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL, + gradients_mean=True) + set_algo_parameters(elementwise_op_strategy_follow=True) + if args_opt.net == "se-resnet50": + context.set_auto_parallel_context(all_reduce_fusion_config=[85, 160]) + else: + context.set_auto_parallel_context(all_reduce_fusion_config=[180, 313]) + init() + + + # create dataset + dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=True, repeat_num=1, + batch_size=config.batch_size, target=target, distribute=args_opt.run_distribute) + step_size = dataset.get_dataset_size() + + # define net + net = resnet(class_num=config.class_num) + if args_opt.parameter_server: + net.set_param_ps() + + # init weight + if args_opt.pre_trained: + param_dict = load_checkpoint(args_opt.pre_trained) + load_param_into_net(net, param_dict) + else: + for _, cell in net.cells_and_names(): + if isinstance(cell, nn.Conv2d): + cell.weight.set_data(weight_init.initializer(weight_init.XavierUniform(), + cell.weight.shape, + cell.weight.dtype)) + if isinstance(cell, nn.Dense): + cell.weight.set_data(weight_init.initializer(weight_init.TruncatedNormal(), + cell.weight.shape, + cell.weight.dtype)) + + # init lr + if args_opt.net == "se-resnet50": + lr = get_lr(lr_init=config.lr_init, lr_end=config.lr_end, lr_max=config.lr_max, + warmup_epochs=config.warmup_epochs, total_epochs=config.epoch_size, steps_per_epoch=step_size, + lr_decay_mode=config.lr_decay_mode) + lr = Tensor(lr) + + # define opt + decayed_params = [] + no_decayed_params = [] + for param in net.trainable_params(): + if 'beta' not in param.name and 'gamma' not in param.name and 'bias' not in param.name: + decayed_params.append(param) + else: + no_decayed_params.append(param) + + group_params = [{'params': decayed_params, 'weight_decay': config.weight_decay}, + {'params': no_decayed_params}, + {'order_params': net.trainable_params()}] + opt = Momentum(group_params, lr, config.momentum, loss_scale=config.loss_scale) + # define loss, model + if target == "Ascend": + if args_opt.dataset == "imagenet2012": + if not config.use_label_smooth: + config.label_smooth_factor = 0.0 + loss = CrossEntropySmooth(sparse=True, reduction="mean", + smooth_factor=config.label_smooth_factor, num_classes=config.class_num) + loss_scale = FixedLossScaleManager(config.loss_scale, drop_overflow_update=False) + model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'}, + amp_level="O2", keep_batchnorm_fp32=False) + + # define callbacks + 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=ckpt_save_dir, config=config_ck) + cb += [ckpt_cb] + + # train model + model.train(config.epoch_size - config.pretrain_epoch_size, dataset, callbacks=cb, + sink_size=dataset.get_dataset_size(), dataset_sink_mode=True)