mindspore/model_zoo/official/cv/xception

• Contents
• Xception 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
  • • Training Performance
    • • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• Contents
• Xception 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
    • Training Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Xception Description

Xception by Google is extreme version of Inception. With a modified depthwise separable convolution, it is even better than Inception-v3. This paper was published in 2017.

Paper Franois Chollet. Xception: Deep Learning with Depthwise Separable Convolutions. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) IEEE, 2017.

Model architecture

The overall network architecture of Xception 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. 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

.
└─Xception
  ├─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
    ├─Xception.py                     # network definition
    ├─loss.py                         # Customized CrossEntropy loss function
    └─lr_generator.py                 # learning rate generator
  ├─train.py                          # train net
  ├─export.py                         # export net
  └─eval.py                           # eval net

Script Parameters

Major parameters in train.py and config.py are:
'num_classes': 1000                # dataset class numbers
'batch_size': 128                  # input batchsize
'loss_scale': 1024                 # loss scale
'momentum': 0.9                    # momentum
'weight_decay': 1e-4               # weight decay
'epoch_size': 250                  # total epoch numbers
'save_checkpoint': True            # save checkpoint
'save_checkpoint_epochs': 1        # save checkpoint epochs
'keep_checkpoint_max': 5           # max numbers to keep checkpoints
'save_checkpoint_path': "./"       # save checkpoint path
'warmup_epochs': 1                 # warmup epoch numbers
'lr_decay_mode': "liner"           # lr decay mode
'use_label_smooth': True           # use label smooth
'finish_epoch': 0                  # finished epochs numbers
'label_smooth_factor': 0.1         # label smoothing factor
'lr_init': 0.00004                 # initiate learning rate
'lr_max': 0.4                      # max bound of learning rate
'lr_end': 0.00004                  # min bound of learning rate

Training process

Usage

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

• Ascend:

# distribute training example(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.
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 --device_target Ascend --dataset_path /dataset/train

  shell:
      Ascend:
      # distribute training example(8p)
      sh scripts/run_distribute_train.sh RANK_TABLE_FILE DATA_PATH
      # standalone training
      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 . /ckpt_0 by default, and training log will be redirected to log.txt like followings.

epoch: 1 step: 1251, loss is 4.8427444
epoch time: 701242.350 ms, per step time: 560.545 ms
epoch: 2 step: 1251, loss is 4.0637593
epoch time: 598591.422 ms, per step time: 478.490ms

Eval process

Usage

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

sh scripts/run_eval.sh DEVICE_ID DATA_DIR PATH_CHECKPOINT

Launch

# eval example
  python:
      Ascend: python eval.py --device_target Ascend --checkpoint_path PATH_CHECKPOINT --dataset_path DATA_DIR

  shell:
      Ascend: sh scripts/run_eval.sh DEVICE_ID DATA_DIR 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.

result: {'Loss': 1.7797744848789312, 'Top_1_Acc': 0.7985777243589743, 'Top_5_Acc': 0.9485777243589744}

Model description

Performance

Training Performance

Parameters	Ascend
Model Version	Xception
Resource	HUAWEI CLOUD Modelarts
uploaded Date	12/10/2020
MindSpore Version	1.1.0
Dataset	1200k images
Batch_size	128
Training Parameters	src/config.py
Optimizer	Momentum
Loss Function	CrossEntropySmooth
Loss	1.78
Accuracy (8p)	Top1[79.8%] Top5[94.8%]
Per step time (8p)	479 ms/step
Total time (8p)	42h
Params (M)	180M
Scripts	Xception script

Inference Performance

Parameters	Ascend
Model Version	Xception
Resource	HUAWEI CLOUD Modelarts
Uploaded Date	11/15/2020
MindSpore Version	1.0.0
Dataset	50k images
Batch_size	128
Accuracy	Top1[79.9%] Top5[94.9%]
Total time	3mins

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.