mindspore/model_zoo/official/cv/resnet_thor

ResNet-50-THOR Example

• Description
• Model Architecture
• Dataset
• Features
• Environment Requirements
• Quick Start
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
  • Evaluation Process
• Model Description
  • Evaluation Performance
  • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Description

This is an example of training ResNet-50 V1.5 with ImageNet2012 dataset by second-order optimizer THOR. THOR is a novel approximate seond-order optimization method in MindSpore. With fewer iterations, THOR can finish ResNet-50 V1.5 training in 72 minutes to top-1 accuracy of 75.9% using 8 Ascend 910, which is much faster than SGD with Momentum.

Model Architecture

The overall network architecture of ResNet-50 is show below:link

Dataset

Dataset used: ImageNet2012

• 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

• Download the dataset ImageNet2012

Unzip the ImageNet2012 dataset to any path you want and the folder structure should include train and eval dataset as follows:

    ├── ilsvrc                  # train dataset
    └── ilsvrc_eval             # infer dataset

Features

The classical first-order optimization algorithm, such as SGD, has a small amount of computation, but the convergence speed is slow and requires lots of iterations. The second-order optimization algorithm uses the second-order derivative of the target function to accelerate convergence, can converge faster to the optimal value of the model and requires less iterations. But the application of the second-order optimization algorithm in deep neural network training is not common because of the high computation cost. The main computational cost of the second-order optimization algorithm lies in the inverse operation of the second-order information matrix (Hessian matrix, Fisher information matrix, etc.), and the time complexity is about O (n^3). On the basis of the existing natural gradient algorithm, we developed the available second-order optimizer THOR in MindSpore by adopting approximation and shearing of Fisher information matrix to reduce the computational complexity of the inverse matrix. With eight Ascend 910 chips, THOR can complete ResNet50-v1.5-ImageNet training in 72 minutes.

Environment Requirements

• Hardware（Ascend/GPU）

  • Prepare hardware environment with Ascend or GPU processor.

• Framework

  • MindSpore

• For more information, please check the resources below：

  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

After installing MindSpore via the official website, you can start training and evaluation as follows:

• Running on Ascend

# run distributed training example
sh run_distribute_train.sh [RANK_TABLE_FILE] [DATASET_PATH] [DEVICE_NUM]

# run evaluation example
sh run_eval.sh [DATASET_PATH] [CHECKPOINT_PATH]

For distributed training, a hccl configuration file with JSON format needs to be created in advance. About the configuration file, you can refer to the HCCL_TOOL.

• Running on GPU

# run distributed training example
sh run_distribute_train_gpu.sh [DATASET_PATH] [DEVICE_NUM]

# run evaluation example
sh run_eval_gpu.sh [DATASET_PATH] [CHECKPOINT_PATH]

Script Description

Script Code Structure

└── resnet_thor
    ├── README.md                                 # descriptions about resnet_thor
    ├── scripts
    │    ├── run_distribute_train.sh               # launch distributed training for Ascend
    │    └── run_eval.sh                           # launch inference for Ascend
    │    ├── run_distribute_train_gpu.sh           # launch distributed training for GPU
    │    └── run_eval_gpu.sh                       # launch inference for GPU
    ├──src
    │    ├── crossentropy.py                       # CrossEntropy loss function
    │    ├── config.py                             # parameter configuration
    │    ├── dataset_helper.py                     # dataset help for minddata dataset
    │    ├── grad_reducer_thor.py                  # grad reducer for thor
    │    ├── model_thor.py                         # model for train
    │    ├── resnet_thor.py                        # resnet50_thor backone
    │    ├── thor.py                               # thor optimizer
    │    ├── thor_layer.py                         # thor layer
    │    └── dataset.py                            # data preprocessing
    ├── eval.py                                   # infer script
    ├── train.py                                  # train script
    ├── export.py                                 # export checkpoint file into air file
    └── mindspore_hub_conf.py                     # config file for mindspore hub repository

Script Parameters

Parameters for both training and inference can be set in config.py.

• Parameters for Ascend 910

"class_num": 1001,                # dataset class number
"batch_size": 32,                 # batch size of input tensor(only supports 32)
"loss_scale": 128,                # loss scale
"momentum": 0.9,                  # momentum of THOR optimizer
"weight_decay": 5e-4,             # weight decay
"epoch_size": 45,                 # only valid for taining, which is always 1 for inference
"save_checkpoint": True,          # whether save checkpoint or not
"save_checkpoint_epochs": 1,      # the epoch interval between two checkpoints. By default, the checkpoint will be saved every epoch
"keep_checkpoint_max": 15,        # only keep the last keep_checkpoint_max checkpoint
"save_checkpoint_path": "./",     # path to save checkpoint relative to the executed path
"use_label_smooth": True,         # label smooth
"label_smooth_factor": 0.1,       # label smooth factor
"lr_init": 0.045,                 # learning rate init value
"lr_decay": 6,                    # learning rate decay rate value
"lr_end_epoch": 70,               # learning rate end epoch value
"damping_init": 0.03,             # damping init value for Fisher information matrix
"damping_decay": 0.87,            # damping decay rate
"frequency": 834,                 # the step interval to update second-order information matrix(should be divisor of the steps of per epoch)

• Parameters for GPU

"class_num": 1001,                # dataset class number
"batch_size": 32,                 # batch size of input tensor
"loss_scale": 128,                # loss scale
"momentum": 0.9,                  # momentum of THOR optimizer
"weight_decay": 5e-4,             # weight decay
"epoch_size": 40,                 # only valid for taining, which is always 1 for inference
"save_checkpoint": True,          # whether save checkpoint or not
"save_checkpoint_epochs": 1,      # the epoch interval between two checkpoints. By default, the checkpoint will be saved every epoch
"keep_checkpoint_max": 15,        # only keep the last keep_checkpoint_max checkpoint
"save_checkpoint_path": "./",     # path to save checkpoint relative to the executed path
"use_label_smooth": True,         # label smooth
"label_smooth_factor": 0.1,       # label smooth factor
"lr_init": 0.05672,               # learning rate init value
"lr_decay": 4.9687,               # learning rate decay rate value
"lr_end_epoch": 50,               # learning rate end epoch value
"damping_init": 0.02345,          # damping init value for Fisher information matrix
"damping_decay": 0.5467,          # damping decay rate
"frequency": 834,                 # the step interval to update second-order information matrix(should be divisor of the steps of per epoch)

Due to the limitation of operators, the value of batch size only supports 32 in Ascend currently. And the update frequency of second-order information matrix must be set the divisor of the steps of per epoch(for example, 834 is the divisor of 5004). As a word, our algorithm is not very flexible in setting those parameters due to the limitations of the framework and operators. But we will solve these problems in the future versions.

Training Process

Ascend 910

  sh run_distribute_train.sh [RANK_TABLE_FILE] [DATASET_PATH] [DEVICE_NUM]

We need three parameters for this scripts.

• RANK_TABLE_FILE：the path of rank_table.json
• DATASET_PATH：the path of train dataset.
• DEVICE_NUM: the device number for distributed train.

Training result will be stored in the current path, whose folder name begins with "train_parallel". Under this, you can find checkpoint file together with result like the following in log.

...
epoch: 1 step: 5004, loss is 4.4182425
epoch: 2 step: 5004, loss is 3.740064
epoch: 3 step: 5004, loss is 4.0546017
epoch: 4 step: 5004, loss is 3.7598825
epoch: 5 step: 5004, loss is 3.3744206
......
epoch: 40 step: 5004, loss is 1.6907625
epoch: 41 step: 5004, loss is 1.8217756
epoch: 42 step: 5004, loss is 1.6453942
...

GPU

sh run_distribute_train_gpu.sh [DATASET_PATH] [DEVICE_NUM]

Training result will be stored in the current path, whose folder name begins with "train_parallel". Under this, you can find checkpoint file together with result like the following in log.

...
epoch: 1 step: 5004, loss is 4.2546034
epoch: 2 step: 5004, loss is 4.0819564
epoch: 3 step: 5004, loss is 3.7005644
epoch: 4 step: 5004, loss is 3.2668946
epoch: 5 step: 5004, loss is 3.023509
......
epoch: 36 step: 5004, loss is 1.645802
...

Evaluation Process

Before running the command below, please check the checkpoint path used for evaluation. Please set the checkpoint path to be the absolute full path, e.g., "username/resnet_thor/train_parallel0/resnet-42_5004.ckpt".

Ascend 910

  sh run_eval.sh [DATASET_PATH] [CHECKPOINT_PATH]

We need two parameters for this scripts.

• DATASET_PATH：the path of evaluation dataset.
• CHECKPOINT_PATH: the absolute path for checkpoint file.

checkpoint can be produced in training process.

Inference result will be stored in the example path, whose folder name is "eval". Under this, you can find result like the following in log.

  result: {'top_5_accuracy': 0.9295574583866837, 'top_1_accuracy': 0.761443661971831} ckpt=train_parallel0/resnet-42_5004.ckpt

GPU

  sh run_eval_gpu.sh [DATASET_PATH] [CHECKPOINT_PATH]

Inference result will be stored in the example path, whose folder name is "eval". Under this, you can find result like the following in log.

  result: {'top_5_accuracy': 0.9287972151088348, 'top_1_accuracy': 0.7597031049935979} ckpt=train_parallel/resnet-36_5004.ckpt

Model Description

Evaluation Performance

Parameters	Ascend 910	GPU
Model Version	ResNet50-v1.5	ResNet50-v1.5
Resource	Ascend 910，CPU 2.60GHz 192cores，Memory 755G	GPU(Tesla V100 SXM2)，CPU 2.1GHz 24cores，Memory 128G
uploaded Date	06/01/2020 (month/day/year)	09/23/2020(month/day/year)
MindSpore Version	0.3.0-alpha	1.0.0
Dataset	ImageNet2012	ImageNet2012
Training Parameters	epoch=45, steps per epoch=5004, batch_size = 32	epoch=40, steps per epoch=5004, batch_size = 32
Optimizer	THOR	THOR
Loss Function	Softmax Cross Entropy	Softmax Cross Entropy
outputs	probability	probability
Loss	1.6453942	1.645802
Speed	20.4ms/step（8pcs）	76ms/step（8pcs）
Total time(acc to 75.9%)	72 mins	229 mins
Parameters (M)	25.5	25.5
Checkpoint for Fine tuning	491M (.ckpt file)	380M (.ckpt file)
Scripts	Link	Link

Inference Performance

Parameters	Ascend 910	GPU
Model Version	ResNet50-v1.5	ResNet50-v1.5
Resource	Ascend 910	GPU
Uploaded Date	06/01/2020 (month/day/year)	09/23/2020(month/day/year)
MindSpore Version	0.3.0-alpha	1.0.0
Dataset	ImageNet2012	ImageNet2012
batch_size	32	32
outputs	probability	probability
Accuracy	76.14%	75.97%
Model for inference	98M (.air file)

Description of Random Situation

In dataset.py, we set the seed inside “create_dataset" function. We also use random seed in train.py.

ModelZoo HomePage

Please check the official homepage.