mindspore/model_zoo/official/cv/inceptionv4

InceptionV4 for Ascend/GPU

• 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/GPU）

  • Prepare hardware environment with Ascend processor.
  • or prepare GPU processor.

• Framework

  • MindSpore

• For more information, please check the resources below：

  • MindSpore Tutorials
  • MindSpore Python API

• Running on ModelArts

  # Train 8p with Ascend
  # (1) Perform a or b.
  #       a. Set "enable_modelarts=True" on default_config.yaml file.
  #          Set "distribute=True" on default_config.yaml file.
  #          Set "need_modelarts_dataset_unzip=True" on default_config.yaml file.
  #          Set "modelarts_dataset_unzip_name='ImageNet_Original'" on default_config.yaml file.
  #          Set "lr_init=0.00004" on default_config.yaml file.
  #          Set "dataset_path='/cache/data'" on default_config.yaml file.
  #          Set "epoch_size=250" on default_config.yaml file.
  #          (optional)Set "checkpoint_url='s3://dir_to_your_pretrained/'" on default_config.yaml file.
  #          Set other parameters on default_config.yaml file you need.
  #       b. Add "enable_modelarts=True" on the website UI interface.
  #          Add "need_modelarts_dataset_unzip=True" on the website UI interface.
  #          Add "modelarts_dataset_unzip_name='ImageNet_Original'" on the website UI interface.
  #          Add "distribute=True" on the website UI interface.
  #          Add "lr_init=0.00004" on the website UI interface.
  #          Add "dataset_path=/cache/data" on the website UI interface.
  #          Add "epoch_size=250" on the website UI interface.
  #          (optional)Add "checkpoint_url='s3://dir_to_your_pretrained/'" on the website UI interface.
  #          Add other parameters on the website UI interface.
  # (2) Prepare model code
  # (3) Upload or copy your pretrained model to S3 bucket if you want to finetune.
  # (4) Perform a or b. (suggested option a)
  #       a. First, zip MindRecord dataset to one zip file.
  #          Second, upload your zip dataset to S3 bucket.(you could also upload the origin mindrecord dataset, but it can be so slow.)
  #       b. Upload the original coco dataset to S3 bucket.
  #           (Data set conversion occurs during training process and costs a lot of time. it happens every time you train.)
  # (5) Set the code directory to "/path/inceptionv4" on the website UI interface.
  # (6) Set the startup file to "train.py" on the website UI interface.
  # (7) Set the "Dataset path" and "Output file path" and "Job log path" to your path on the website UI interface.
  # (8) Create your job.
  #
  # Train 1p with Ascend
  # (1) Perform a or b.
  #       a. Set "enable_modelarts=True" on default_config.yaml file.
  #          Set "need_modelarts_dataset_unzip=True" on default_config.yaml file.
  #          Set "modelarts_dataset_unzip_name='ImageNet_Original'" on default_config.yaml file.
  #          Set "dataset_path='/cache/data'" on default_config.yaml file.
  #          Set "epoch_size=250" on default_config.yaml file.
  #          (optional)Set "checkpoint_url='s3://dir_to_your_pretrained/'" on default_config.yaml file.
  #          Set other parameters on default_config.yaml file you need.
  #       b. Add "enable_modelarts=True" on the website UI interface.
  #          Add "need_modelarts_dataset_unzip=True" on the website UI interface.
  #          Add "modelarts_dataset_unzip_name='ImageNet_Original'" on the website UI interface.
  #          Add "dataset_path='/cache/data'" on the website UI interface.
  #          Add "epoch_size=250" on the website UI interface.
  #          (optional)Add "checkpoint_url='s3://dir_to_your_pretrained/'" on the website UI interface.
  #          Add other parameters on the website UI interface.
  # (2) Prepare model code
  # (3) Upload or copy your pretrained model to S3 bucket if you want to finetune.
  # (4) Perform a or b. (suggested option a)
  #       a. zip MindRecord dataset to one zip file.
  #          Second, upload your zip dataset to S3 bucket.(you could also upload the origin mindrecord dataset, but it can be so slow.)
  #       b. Upload the original coco dataset to S3 bucket.
  #           (Data set conversion occurs during training process and costs a lot of time. it happens every time you train.)
  # (5) Set the code directory to "/path/inceptionv4" on the website UI interface.
  # (6) Set the startup file to "train.py" on the website UI interface.
  # (7) Set the "Dataset path" and "Output file path" and "Job log path" to your path on the website UI interface.
  # (8) Create your job.
  #
  # Eval 1p with Ascend
  # (1) Perform a or b.
  #       a. Set "enable_modelarts=True" on default_config.yaml file.
  #          Set "need_modelarts_dataset_unzip=True" on default_config.yaml file.
  #          Set "modelarts_dataset_unzip_name='ImageNet_Original'" on default_config.yaml file.
  #          Set "checkpoint_url='s3://dir_to_your_trained_model/'" on base_config.yaml file.
  #          Set "checkpoint_path='./inceptionv4/inceptionv4-train-250_1251.ckpt'" on default_config.yaml file.
  #          Set "dataset_path='/cache/data'" on default_config.yaml file.
  #          Set other parameters on default_config.yaml file you need.
  #       b. Add "enable_modelarts=True" on the website UI interface.
  #          Add "need_modelarts_dataset_unzip=True" on the website UI interface.
  #          Add "modelarts_dataset_unzip_name='ImageNet_Original'" on the website UI interface.
  #          Add "checkpoint_url='s3://dir_to_your_trained_model/'" on the website UI interface.
  #          Add "checkpoint_path='./inceptionv4/inceptionv4-train-250_1251.ckpt'" on the website UI interface.
  #          Add "dataset_path='/cache/data'" on the website UI interface.
  # (2) Prepare model code
  #          Add other parameters on the website UI interface.
  # (3) Upload or copy your trained model to S3 bucket.
  # (4) Perform a or b. (suggested option a)
  #       a. First, zip MindRecord dataset to one zip file.
  #          Second, upload your zip dataset to S3 bucket.(you could also upload the origin mindrecord dataset, but it can be so slow.)
  #       b. Upload the original coco dataset to S3 bucket.
  #           (Data set conversion occurs during training process and costs a lot of time. it happens every time you train.)
  # (5) Set the code directory to "/path/inceptionv4" on the website UI interface.
  # (6) Set the startup file to "eval.py" on the website UI interface.
  # (7) Set the "Dataset path" and "Output file path" and "Job log path" to your path on the website UI interface.
  # (8) Create your job.
  

Script description

Script and sample code

.
└─Inception-v4
  ├─README.md
  ├─ascend310_infer                     # application for 310 inference
  ├─scripts
    ├─run_distribute_train_gpu.sh       # launch distributed training with gpu platform(8p)
    ├─run_eval_gpu.sh                   # launch evaluating with gpu platform
    ├─run_eval_cpu.sh                   # launch evaluating with cpu platform
    ├─run_standalone_train_cpu.sh       # launch standalone training with cpu platform(1p)
    ├─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_infer_310.sh                  # shell script for 310 inference
    └─run_eval_ascend.sh                # launch evaluating with ascend platform
  ├─src
    ├─dataset.py                      # data preprocessing
    ├─inceptionv4.py                  # network definition
    ├─callback.py                     # eval callback function
    └─model_utils
      ├─config.py               # Processing configuration parameters
      ├─device_adapter.py       # Get cloud ID
      ├─local_adapter.py        # Get local ID
      └─moxing_adapter.py       # Parameter processing
  ├─default_config.yaml             # Training parameter profile(ascend)
  ├─default_config_cpu.yaml         # Training parameter profile(cpu)
  ├─default_config_gpu.yaml         # Training parameter profile(gpu)
  ├─eval.py                         # eval net
  ├─export.py                       # export checkpoint, surpport .onnx, .air, .mindir convert
  ├─postprogress.py                 # post process for 310 inference
  └─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

• GPU:

# distribute training example(8p)
sh scripts/run_distribute_train_gpu.sh DATA_PATH

• CPU:

# standalone training example with shell
sh scripts/run_standalone_train_cpu.sh DATA_PATH

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
      GPU:
      # distribute training example(8p)
      sh scripts/run_distribute_train_gpu.sh DATA_PATH
      CPU:
      # standalone training example with shell
      sh scripts/run_standalone_train_cpu.sh DATA_PATH

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 following.

• Ascend

epoch: 1 step: 1251, loss is 5.4833196
Epoch time: 520274.060, per step time: 415.887
epoch: 2 step: 1251, loss is 4.093194
Epoch time: 288520.628, per step time: 230.632
epoch: 3 step: 1251, loss is 3.6242008
Epoch time: 288507.506, per step time: 230.622

• GPU

epoch: 1 step: 1251, loss is 6.49775
Epoch time: 1487493.604, per step time: 1189.044
epoch: 2 step: 1251, loss is 5.6884665
Epoch time: 1421838.433, per step time: 1136.561
epoch: 3 step: 1251, loss is 5.5168786
Epoch time: 1423009.501, per step time: 1137.498

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

• GPU

  sh scripts/run_eval_gpu.sh DATA_DIR CHECKPOINT_PATH

Launch

# eval example
  shell:
      Ascend:
            sh scripts/run_eval_ascend.sh DEVICE_ID DATA_DIR CHECKPOINT_PATH
      GPU:
            sh scripts/run_eval_gpu.sh 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 following in eval.log.

• Ascend

metric: {'Loss': 0.9849, 'Top1-Acc':0.7985, 'Top5-Acc':0.9460}

• GPU(8p)

metric: {'Loss': 0.8144, 'Top1-Acc': 0.8009, 'Top5-Acc': 0.9457}

Model Export

python export.py --config_path [CONFIG_FILE] --ckpt_file [CKPT_PATH] --device_target [DEVICE_TARGET] --file_format[EXPORT_FORMAT]

EXPORT_FORMAT should be in ["AIR", "MINDIR"]

Inference Process

Usage

Before performing inference, the model file must be exported by export script on the Ascend910 environment.

# Ascend310 inference
sh run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DEVICE_ID]

-NOTE:Ascend310 inference use Imagenet dataset . The label of the image is the number of folder which is started from 0 after sorting.

result

Inference result is saved in current path, you can find result like this in acc.log file.

accuracy:80.044

Model description

Performance

Training Performance

Parameters	Ascend	GPU
Model Version	InceptionV4	InceptionV4
Resource	Ascend 910; cpu 2.60GHz, 192cores; memory 755G; OS Euler2.8	NV SMX2 V100-32G
uploaded Date	11/04/2020	03/05/2021
MindSpore Version	1.0.0	1.0.0
Dataset	1200k images	1200K images
Batch_size	128	128
Training Parameters	src/model_utils/default_config.yaml (Ascend)	src/model_utils/default_config.yaml (GPU)
Optimizer	RMSProp	RMSProp
Loss Function	SoftmaxCrossEntropyWithLogits	SoftmaxCrossEntropyWithLogits
Outputs	probability	probability
Loss	0.98486	0.8144
Accuracy (8p)	ACC1[79.85%] ACC5[94.60%]	ACC1[80.09%] ACC5[94.57%]
Total time (8p)	20h	95h
Params (M)	153M	153M
Checkpoint for Fine tuning	2135M	489M
Scripts	inceptionv4 script	inceptionv4 script

Inference Performance

Parameters	Ascend	GPU
Model Version	InceptionV4	InceptionV4
Resource	Ascend 910; cpu 2.60GHz, 192cores; memory 755G; OS Euler2.8	NV SMX2 V100-32G
Uploaded Date	11/04/2020	03/05/2021
MindSpore Version	1.0.0	1.0.0
Dataset	50k images	50K images
Batch_size	128	128
Outputs	probability	probability
Accuracy	ACC1[79.85%] ACC5[94.60%]	ACC1[80.09%] ACC5[94.57%]
Total time	2mins	2mins
Model for inference	2135M (.ckpt file)	489M (.ckpt file)

Training performance results

Ascend	train performance
1p	556 img/s

Ascend	train performance
8p	4430 img/s

GPU	train performance
8p	906 img/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.