mindspore/model_zoo/official/cv/inceptionv4

• InceptionV4 for Ascend
• InceptionV4 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
      • Training performance results
• Description of Random Situation
• ModelZoo Homepage

InceptionV4 for Ascend

• InceptionV4 Description
• Model Architecture
• Dataset
• Features
  • Mixed Precision
• Environment Requirements
• Script Description
  • Script and Sample Code
  • Training Process
  • Evaluation Process
    • Evaluation
• Model Description
  • Performance
    • Training Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

InceptionV4 Description

Inception-v4 is a convolutional neural network architecture that builds on previous iterations of the Inception family by simplifying the architecture and using more inception modules than Inception-v3. This idea was proposed in the paper Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning, published in 2016.

Paper Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi. Computer Vision and Pattern Recognition[J]. 2016.

Model architecture

The overall network architecture of InceptionV4 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-v4
  ├─README.md
  ├─scripts
    ├─run_standalone_train_ascend.sh    # launch standalone training with ascend platform(1p)
    ├─run_distribute_train_ascend.sh    # launch distributed training with ascend platform(8p)
    └─run_eval_ascend.sh                # launch evaluating with ascend platform
  ├─src
    ├─config.py                       # parameter configuration
    ├─dataset.py                      # data preprocessing
    ├─inceptionv4.py                  # network definition
    └─callback.py                     # eval callback function
  ├─eval.py                           # eval net
  ├─export.py                         # export checkpoint, surpport .onnx, .air, .mindir convert
  └─train.py                          # train net

Script Parameters

Major parameters in train.py and config.py are:
'is_save_on_master'          # save checkpoint only on master device
'batch_size'                 # input batchsize
'epoch_size'                 # total epoch numbers
'num_classes'                # dataset class numbers
'work_nums'                  # number of workers to read data
'loss_scale'                 # loss scale
'smooth_factor'              # label smoothing factor
'weight_decay'               # weight decay
'momentum'                   # momentum
'amp_level'                  # precision training, Supports [O0, O2, O3]
'decay'                      # decay used in optimize function
'epsilon'                    # epsilon used in iptimize function
'keep_checkpoint_max'        # max numbers to keep checkpoints
'save_checkpoint_epochs'     # save checkpoints per n epoch
'lr_init'                    # init leaning rate
'lr_end'                     # end of learning rate
'lr_max'                     # max bound of learning rate
'warmup_epochs'              # warmup epoch numbers
'start_epoch'                # number of start epoch range[1, epoch_size]

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_ascend.sh RANK_TABLE_FILE DATA_PATH DATA_DIR
# standalone training
sh scripts/run_standalone_train_ascend.sh DEVICE_ID DATA_DIR

Notes:
RANK_TABLE_FILE can refer to Link , and the device_ip can be got as Link. For large models like InceptionV4, 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
  shell:
      Ascend:
      # distribute training example(8p)
      sh scripts/run_distribute_train_ascend.sh RANK_TABLE_FILE DATA_PATH DATA_DIR
      # standalone training
      sh scripts/run_standalone_train_ascend.sh DEVICE_ID DATA_DIR

Result

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

epoch: 1 step: 1251, loss is 5.861846
Epoch time: 701416.649, per step time: 560.685
epoch: 2 step: 1251, loss is 4.295785
Epoch time: 472524.154, per step time: 377.717
epoch: 3 step: 1251, loss is 3.691987
Epoch time: 472505.767, per step time: 377.702

Eval process

Usage

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

• Ascend:

  sh scripts/run_eval_ascend.sh DEVICE_ID DATA_DIR CHECKPOINT_PATH

Launch

# eval example
  shell:
      Ascend:
            sh scripts/run_eval_ascend.sh DEVICE_ID DATA_DIR CHECKPOINT_PATH

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': 0.9849, 'Top1-Acc':0.7985, 'Top5-Acc':0.9460}

Model description

Performance

Training Performance

Parameters	Ascend
Model Version	InceptionV4
Resource	Ascend 910, cpu:2.60GHz 192cores, memory:755G
uploaded Date	11/04/2020
MindSpore Version	1.0.0
Dataset	1200k images
Batch_size	128
Training Parameters	src/config.py
Optimizer	RMSProp
Loss Function	SoftmaxCrossEntropyWithLogits
Outputs	probability
Loss	0.98486
Accuracy (8p)	ACC1[79.85%] ACC5[94.60%]
Total time (8p)	33h
Params (M)	153M
Checkpoint for Fine tuning	2135M
Scripts	inceptionv4 script

Inference Performance

Parameters	Ascend
Model Version	InceptionV4
Resource	Ascend 910, cpu:2.60GHz 192cores, memory:755G
Uploaded Date	11/04/2020
MindSpore Version	1.0.0
Dataset	50k images
Batch_size	128
Outputs	probability
Accuracy	ACC1[79.85%] ACC5[94.60%]
Total time	2mins
Model for inference	2135M (.ckpt file)

Training performance results

Ascend	train performance
1p	345 img/s

Ascend	train performance
8p	2708img/s

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.