mindspore/model_zoo/official/cv/inceptionv3

• Contents
• InceptionV3 Description
• Model architecture
• Dataset
• Features
• • Mixed Precision(Ascend)
• Environment Requirements
• Script description
• • Script and sample code
  • Script Parameters
  • Training process
  • • Usage
    • Launch
    • Result
  • Eval process
  • • Usage
    • Launch
    • Result
• Model description
• • Performance
  • • Evaluation Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• InceptionV3 Description
• Model Architecture
• Dataset
• Features
  • Mixed Precision
• Environment Requirements
• Script Description
  • Script and Sample Code
  • Training Process
  • Evaluation Process
    • Evaluation
• Model Description
  • Performance
    • Evaluation Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

InceptionV3 Description

InceptionV3 by Google is the 3rd version in a series of Deep Learning Convolutional Architectures. Inception v3 mainly focuses on burning less computational power by modifying the previous Inception architectures. This idea was proposed in the paper Rethinking the Inception Architecture for Computer Vision, published in 2015.

Paper Min Sun, Ali Farhadi, Steve Seitz. Ranking Domain-Specific Highlights by Analyzing Edited Videos[J]. 2014.

Model architecture

The overall network architecture of InceptionV3 is show below:

Link

Dataset

Dataset used can refer to paper.

• Dataset size: 125G, 1250k colorful images in 1000 classes
  • Train: 120G, 1200k images
  • Test: 5G, 50k images
• Data format: RGB images.
  • Note: Data will be processed in src/dataset.py

Features

Mixed Precision(Ascend)

The mixed precision training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, 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

• Hardware（Ascend）
• Prepare hardware environment with Ascend processor. If you want to try Ascend, please send the application form to ascend@huawei.com. Once approved, you can get the resources.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Script description

Script and sample code

.
└─Inception-v3
  ├─README.md
  ├─scripts
    ├─run_standalone_train.sh                 # launch standalone training with ascend platform(1p)
    ├─run_distribute_train.sh                 # launch distributed training with ascend platform(8p)
    └─run_eval.sh                             # launch evaluating with ascend platform
  ├─src
    ├─config.py                       # parameter configuration
    ├─dataset.py                      # data preprocessing
    ├─inception_v3.py                 # network definition
    ├─loss.py                         # Customized CrossEntropy loss function
    ├─lr_generator.py                 # learning rate generator
  ├─eval.py                           # eval net
  ├─export.py                         # convert checkpoint
  └─train.py                          # train net

Script Parameters

Major parameters in train.py and config.py are:
'random_seed'                # fix random seed
'work_nums'                  # number of workers to read the data
'decay_method'               # learning rate scheduler mode
"loss_scale"                 # loss scale
'batch_size'                 # input batchsize
'epoch_size'                 # total epoch numbers
'num_classes'                # dataset class numbers
'smooth_factor'              # label smoothing factor
'aux_factor'                 # loss factor of aux logit
'lr_init'                    # initiate learning rate
'lr_max'                     # max bound of learning rate
'lr_end'                     # min bound of learning rate
'warmup_epochs'              # warmup epoch numbers
'weight_decay'               # weight decay
'momentum'                   # momentum
'opt_eps'                    # epsilon
'keep_checkpoint_max'        # max numbers to keep checkpoints
'ckpt_path'                  # save checkpoint path
'is_save_on_master'          # save checkpoint on rank0, distributed parameters
'dropout_keep_prob'          # the keep rate, between 0 and 1, e.g. keep_prob = 0.9, means dropping out 10% of input units
'has_bias'                   # specifies whether the layer uses a bias vector.
'amp_level'                  # option for argument `level` in `mindspore.amp.build_train_network`, level for mixed
                             # precision training. Supports [O0, O2, O3].

Training process

Usage

You can start training using python or shell scripts. The usage of shell scripts as follows:

• Ascend:

# distribute training(8p)
sh scripts/run_distribute_train.sh RANK_TABLE_FILE DATA_PATH
# standalone training
sh scripts/run_standalone_train.sh DEVICE_ID DATA_PATH

Notes: RANK_TABLE_FILE can refer to Link, and the device_ip can be got as Link. For large models like InceptionV3, it's better to export an external environment variable export HCCL_CONNECT_TIMEOUT=600 to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size.

This is processor cores binding operation regarding the device_num and total processor numbers. If you are not expect to do it, remove the operations taskset in scripts/run_distribute_train.sh

Launch

# training example
  python:
      Ascend: python train.py --dataset_path DATA_PATH --platform Ascend

  shell:
      Ascend:
      # distribute training example(8p)
      sh scripts/run_distribute_train.sh RANK_TABLE_FILE DATA_PATH
      # standalone training example
      sh scripts/run_standalone_train.sh DEVICE_ID DATA_PATH

Result

Training result will be stored in the example path. Checkpoints will be stored at . /checkpoint by default, and training log will be redirected to ./log.txt like followings.

epoch: 0 step: 1251, loss is 5.7787247
epoch time: 360760.985 ms, per step time: 288.378 ms
epoch: 1 step: 1251, loss is 4.392868
epoch time: 160917.911 ms, per step time: 128.631 ms

Eval process

Usage

You can start training using python or shell scripts. The usage of shell scripts as follows:

• Ascend:

    sh scripts/run_eval.sh DEVICE_ID DATA_PATH PATH_CHECKPOINT

Launch

# eval example
  python:
      Ascend: python eval.py --dataset_path DATA_PATH --checkpoint PATH_CHECKPOINT --platform Ascend

  shell:
      Ascend: sh scripts/run_eval.sh DEVICE_ID DATA_PATH PATH_CHECKPOINT

checkpoint can be produced in training process.

Result

Evaluation result will be stored in the example path, you can find result like the followings in eval.log.

metric: {'Loss': 1.778, 'Top1-Acc':0.788, 'Top5-Acc':0.942}

Model description

Performance

Evaluation Performance

Parameters	Ascend
Model Version	InceptionV3
Resource	Ascend 910, cpu:2.60GHz 192cores, memory:755G
uploaded Date	08/21/2020
MindSpore Version	0.6.0-beta
Dataset	1200k images
Batch_size	128
Training Parameters	src/config.py
Optimizer	RMSProp
Loss Function	SoftmaxCrossEntropy
Outputs	probability
Loss	1.98
Total time (8p)	11h
Params (M)	103M
Checkpoint for Fine tuning	313M
Model for inference	92M (.onnx file)
Speed	1pc:1050 img/s;8pc:8000 img/s
Scripts	inceptionv3 script

Inference Performance

Parameters	Ascend
Model Version	InceptionV3
Resource	Ascend 910, cpu:2.60GHz 192cores, memory:755G
Uploaded Date	08/22/2020
MindSpore Version	0.6.0-beta
Dataset	50k images
Batch_size	128
Outputs	probability
Accuracy	ACC1[78.8%] ACC5[94.2%]
Total time	2mins

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.