Using MobileNetV2 to Implement Fine-Tuning

Linux Windows Ascend GPU CPU Model Development Intermediate Expert

View Source On Gitee  

Overview

In a computer vision task, training a network from scratch is time-consuming and requires a large amount of computing power. Pre-trained models often select open large datasets such as OpenImage, ImageNet, VOC, and COCO. The number of images in these datasets reaches hundreds of thousands or even millions. Most tasks have a large amount of data. If a pre-trained model is not used during network model training, the training from scratch consumes a large amount of time and computing power. As a result, the model is prone to local minimum and overfitting. Therefore, most tasks perform fine-tuning on pre-trained models.

MindSpore is a diversified machine learning framework. It can run on devices such as mobile phones and PCs, or on server clusters on the cloud. Currently, MobileNetV2 supports fine-tuning on a single CPU or on one or more Ascend AI Processors or GPUs on Windows, EulerOS, and Ubuntu systems. This tutorial describes how to perform fine-tuning training and validation in the MindSpore frameworks of different systems and processors.

Currently, only the CPU is supported on Windows, and the CPU, GPU, and Ascend AI Processor are supported on Ubuntu and EulerOS.

You can obtain the complete executable sample code at https://gitee.com/mindspore/mindspore/tree/r1.1/model_zoo/official/cv/mobilenetv2.

Task Description and Preparations

Environment Configuration

If running a task in a local environment, install the MindSpore framework and configure the CPU, GPU, or Ascend AI Processor. If running a task in the HUAWEI CLOUD environment, skip this section because the installation and configuration are not required.

On the Windows operating system, backslashes \ are used to separate directories of different levels in a path address. On the Linux operating system, slashes / are used. The following uses / by default. If you use Windows operating system, replace / in the path address with \.

  1. Install the MindSpore framework. Install a MindSpore framework based on the processor architecture and the EulerOS, Ubuntu, or Windows system.

  2. Configure the CPU environment.
    Set the following code before calling the CPU to start training or testing:

    if config.platform == "CPU":
        context.set_context(mode=context.GRAPH_MODE, device_target=config.platform, \
            save_graphs=False)
    
  3. Configure the GPU environment.
    Set the following code before calling the GPU to start training or testing:

    elif config.platform == "GPU":
        context.set_context(mode=context.GRAPH_MODE, device_target=config.platform, save_graphs=False)
        if config.run_distribute:
            init("nccl")
            context.set_auto_parallel_context(device_num=get_group_size(),
                                              parallel_mode=ParallelMode.DATA_PARALLEL,
                                              gradients_mean=True)
    
  4. Configure the Ascend environment.
    The following uses the JSON configuration file hccl_config.json in an environment with eight Ascend 910 AI processors as an example. Adjust "server_count" and device based on the following example to switch between the single-device and multi-device environments:

    {
        "version": "1.0",
        "server_count": "1",
        "server_list": [
            {
                "server_id": "10.*.*.*",
                "device": [
                    {"device_id": "0","device_ip": "192.1.27.6","rank_id": "0"},
                    {"device_id": "1","device_ip": "192.2.27.6","rank_id": "1"},
                    {"device_id": "2","device_ip": "192.3.27.6","rank_id": "2"},
                    {"device_id": "3","device_ip": "192.4.27.6","rank_id": "3"},
                    {"device_id": "4","device_ip": "192.1.27.7","rank_id": "4"},
                    {"device_id": "5","device_ip": "192.2.27.7","rank_id": "5"},
                    {"device_id": "6","device_ip": "192.3.27.7","rank_id": "6"},
                    {"device_id": "7","device_ip": "192.4.27.7","rank_id": "7"}],
                "host_nic_ip": "reserve"
            }
        ],
        "status": "completed"
    }
    

    Set the following code before calling the Ascend AI Processor to start training or testing:

    elif config.platform == "Ascend":
        context.set_context(mode=context.GRAPH_MODE, device_target=config.platform, device_id=config.device_id,
                            save_graphs=False)
        if config.run_distribute:
            context.set_auto_parallel_context(device_num=config.rank_size,
                                              parallel_mode=ParallelMode.DATA_PARALLEL,
                                              gradients_mean=True,
                                              all_reduce_fusion_config=[140])
            init()
    ...
    

Downloading Code

Run the following command to clone MindSpore open-source project repository in Gitee and go to ./model_zoo/official/cv/mobilenetv2/.

git clone https://gitee.com/mindspore/mindspore.git -b r1.1
cd ./mindspore/model_zoo/official/cv/mobilenetv2

The code structure is as follows:

├─MobileNetV2
    ├─README.md     # descriptions about MobileNetV2
    ├─scripts
       run_train.sh   # Shell script for train with Ascend or GPU
       run_eval.sh    # Shell script for evaluation with Ascend or GPU
    ├─src
       config.py      # parameter configuration
       dataset.py     # creating dataset
       launch.py      # start Python script
       lr_generator.py     # learning rate config
       mobilenetV2.py      # MobileNetV2 architecture
       mobilenetV2_fusion.py      # MobileNetV2 fusion architecture
       models.py        # net utils to load ckpt_file, define_net...
       utils.py        # net utils to switch precision, set_context and so on
    ├─train.py      # training script
    └─eval.py       #  evaluation script

During fine-tuning training and testing, python files train.py and eval.py can be used on Windows, Ubuntu, and EulerOS, and shell script files run_train.sh and run_eval.sh can be used on Ubuntu and EulerOS.

If the script file run_train.sh is used, it runs launch.py and inputs parameters to launch.py which starts one or more processes to run train.py based on the number of allocated CPUs, GPUs, or Ascend AI Processors. Each process is allocated with a processor.

Preparing a Pre-Trained Model

Download a CPU/GPU pre-trained model or Ascend pre-trained model to the following directories based on the processor type:
./pretrain_checkpoint/

  • CPU/GPU

    mkdir pretrain_checkpoint
    wget -P ./pretrain_checkpoint https://download.mindspore.cn/model_zoo/official/lite/mobilenetv2_openimage_lite/mobilenetv2_cpu_gpu.ckpt
    
  • Ascend AI Processor

    mkdir pretrain_checkpoint
    wget -P ./pretrain_checkpoint https://download.mindspore.cn/model_zoo/official/lite/mobilenetv2_openimage_lite/mobilenetv2_ascend.ckpt
    

Preparing Data

Prepare the dataset managed in ImageFolder format. Add the <dataset_path> parameter when running run_train.sh, and add the --dataset_path <dataset_path> parameter when running train.py.

The dataset structure is as follows:

└─ImageFolder
    ├─train
       class1Folder
       class2Folder
       ......
    └─eval
        class1Folder
        class2Folder
        ......

Code for Loading a Pre-Trained Model

During fine-tuning, you need to load a pre-trained model. The distribution of the feature extraction layer (convolutional layer) in different datasets and tasks tends to be consistent. However, the combination of feature vectors (fully connected layer) is different, and the number of classes (output_size of the fully connected layer) is usually different. During fine-tuning, parameters of the feature extraction layer are loaded and trained, while those of the fully connected layer are not. During fine-tuning and initial training, both feature extraction layer parameters and fully connected layer parameters are loaded and trained.

Before training and testing, build a backbone network and a head network of MobileNetV2 on the first line of the code, and build a MobileNetV2 network containing the two subnets. Lines 3 to 10 of the code show how to define backbone_net and head_net and how to add the two subnets to mobilenet_v2. Lines 12 to 23 of the code show that in fine-tuning training mode, the pre-trained model needs to be loaded to the backbone_net subnet, and parameters in backbone_net are frozen and do not participate in training. Lines 21 to 23 of the code show how to freeze network parameters.

 1:  backbone_net, head_net, net = define_net(args_opt, config)
 2:  ...
 3:  def define_net(config, is_training):
 4:      backbone_net = MobileNetV2Backbone()
 5:      activation = config.activation if not is_training else "None"
 6:      head_net = MobileNetV2Head(input_channel=backbone_net.out_channels,
 7:                                 num_classes=config.num_classes,
 8:                                 activation=activation)
 9:      net = mobilenet_v2(backbone_net, head_net)
10:      return backbone_net, head_net, net
11:  ...
12:  if args_opt.pretrain_ckpt and args_opt.freeze_layer == "backbone":
13:     load_ckpt(backbone_net, args_opt.pretrain_ckpt, trainable=False)
14:  ...
15:  def load_ckpt(network, pretrain_ckpt_path, trainable=True):
16:      """
17:      train the param weight or not
18:      """
19:      param_dict = load_checkpoint(pretrain_ckpt_path)
20:      load_param_into_net(network, param_dict)
21:      if not trainable:
22:          for param in network.get_parameters():
23:              param.requires_grad = False

Parameter Description

Change the value of each parameter based on the local processor type, data path, and pre-trained model path.

Running Python Files

When using train.py for training on Windows and Linux, input dataset_path, platform, pretrain_ckpt, and freeze_layer. When using eval.py for validation, input dataset_path, platform, and pretrain_ckpt.

# Windows/Linux train with Python file
python train.py --platform [PLATFORM] --dataset_path <DATASET_PATH>  --pretrain_ckpt [PRETRAIN_CHECKPOINT_PATH] --freeze_layer[("none", "backbone")]

# Windows/Linux eval with Python file
python eval.py --platform [PLATFORM] --dataset_path <DATASET_PATH> --pretrain_ckpt <PRETRAIN_CHECKPOINT_PATH>
  • --dataset_path: path of the training or validation dataset. There is no default value. This parameter is mandatory for training or validation.

  • --platform: processor type. The default value is Ascend. You can set it to CPU or GPU.

  • --pretrain_ckpt: path of the pretrain_checkpoint file required for loading a weight of a pre-trained model parameter during incremental training or optimization.

  • --freeze_layer: frozen network layer. Enter none or backbone.

Running Shell Scripts

You can run the shell scripts ./scripts/run_train.sh and ./scripts/run_eval.sh on Linux. Input parameters on the interaction interface.

# Windows doesn't support Shell
# Linux train with Shell script
sh run_train.sh <PLATFORM> <DEVICE_NUM> <VISIABLE_DEVICES(0,1,2,3,4,5,6,7)> <RANK_TABLE_FILE> <DATASET_PATH> <CKPT_PATH> [FREEZE_LAYER]

# Linux eval with Shell script for fine tune
sh run_eval.sh <PLATFORM> <DATASET_PATH> <PRETRAIN_CKPT_PATH>
  • <PLATFORM>: processor type. The default value is Ascend. You can set it to GPU.

  • <DEVICE_NUM>: number of processes on each node (equivalent to a server or PC). You are advised to set this parameter to the number of Ascend AI Processors or GPUs on a server.

  • <VISIABLE_DEVICES(0,1,2,3,4,5,6,7)>: device ID of character string type. During training, a process is bound to a device with the corresponding ID based on <VISIABLE_DEVICES>. Multiple device IDs are separated by commas (,). It is recommended that the number of IDs be the same as the number of processes.

  • <RANK_TABLE_FILE>: a JSON file configured when platform is set to Ascend

  • <DATASET_PATH>: path of the training or validation dataset. There is no default value. This parameter is mandatory for training or validation.

  • <CKPT_PATH>: path of the checkpoint file required for loading a weight of a pre-trained model parameter during incremental training or optimization.

  • [FREEZE_LAYER]: frozen network layer during fine-tuned model validation. Enter none or backbone.

Loading Fine-Tuning Training

Only train.py can be run on Windows when MobileNetV2 is used for fine-tuning training. You can run the shell script run_train.sh and input parameters on Linux when MobileNetV2 is used for fine-tuning training.

The Windows system outputs information to an interactive command line. When running run_train.sh on the Linux system, use &> <log_file_path> at the end of the command line to write the standard output and error output to the log file. After the fine-tuning is successful, training starts. The training time and loss of each epoch are continuously written into the ./train/rank*/log*.log file. If the fine-tuning fails, an error message is recorded in the preceding log file.

Loading Training on CPU

  • Set the number of nodes.

    Currently, train.py supports only a single processor. You do not need to adjust the number of processors. When the run_train.sh file is run, a single CPU is used by default. The number of CPUs cannot be changed.

  • Start incremental training.

    Example 1: Use the python file to call a CPU.

    # Windows or Linux with Python
    python train.py --platform CPU --dataset_path <TRAIN_DATASET_PATH>  --pretrain_ckpt ./pretrain_checkpoint/mobilenetv2_cpu_gpu.ckpt --freeze_layer backbone
    

    Example 2: Use the shell file to call a CPU.

    # Linux with Shell
    sh run_train.sh CPU <TRAIN_DATASET_PATH> ../pretrain_checkpoint/mobilenetV2_cpu_gpu.ckpt backbone
    

Loading Training on GPU

  • Set the number of nodes.

    Currently, train.py supports only a single processor. You do not need to adjust the number of nodes. When running the run_train.sh file, set <nproc_per_node> to the number of GPUs and <visible_devices> to IDs of available processors, that is, GPU IDs. You can select one or more device IDs and separate them with commas (,).

  • Start incremental training.

    • Example 1: Use the python file to call a GPU.

      # Windows or Linux with Python
      python train.py --platform GPU --dataset_path <TRAIN_DATASET_PATH> --pretrain_ckpt ./pretrain_checkpoint/mobilenetv2_cpu_gpu.ckpt --freeze_layer backbone
      
    • Example 2: Use the shell script to call a GPU whose device ID is 0.

      # Linux with Shell
      sh run_train.sh GPU 1 0 <TRAIN_DATASET_PATH> ../pretrain_checkpoint/mobilenetv2_cpu_gpu.ckpt backbone
      
    • Example 3: Use the shell script to call eight GPUs whose device IDs are 0,1,2,3,4,5,6,7.

      # Linux with Shell
      sh run_train.sh GPU 8 0,1,2,3,4,5,6,7 <TRAIN_DATASET_PATH> ../pretrain_checkpoint/mobilenetv2_cpu_gpu.ckpt backbone
      

Loading Training on Ascend AI Processor

  • Set the number of nodes.

    Currently, train.py supports only a single processor. You do not need to adjust the number of nodes. When running the run_train.sh file, set <nproc_per_node> to the number of Ascend AI Processors and <visible_devices> to IDs of available processors, that is, Ascend AI Processor IDs. You can select one or more device IDs from 0 to 7 on an 8-device server and separate them with commas (,). Currently, the number of Ascend AI Processors can only be set to 1 or 8.

  • Start incremental training.

    • Example 1: Use the python file to call an Ascend AI Processor.

      # Windows or Linux with Python
      python train.py --platform Ascend --dataset_path <TRAIN_DATASET_PATH>  --pretrain_ckpt  ./pretrain_checkpoint mobilenetv2_ascend.ckpt --freeze_layer backbone
      
    • Example 2: Use the shell script to call an Ascend AI Processor whose device ID is 0.

      # Linux with Shell
      sh run_train.sh Ascend 1 0 ~/rank_table.json <TRAIN_DATASET_PATH> ../pretrain_checkpoint/mobilenetv2_ascend.ckpt backbone
      
    • Example 3: Use the shell script to call eight Ascend AI Processors whose device IDs are 0,1,2,3,4,5,6,7.

      # Linux with Shell
      sh run_train.sh Ascend 8 0,1,2,3,4,5,6,7 ~/rank_table.json <TRAIN_DATASET_PATH> ../pretrain_checkpoint/mobilenetv2_ascend.ckpt backbone
      

Fine-Tuning Training Result

  • View the running result.

    • When running the python file, view the output information in the interactive command line. After running the shell script on Linux, run the cat ./train/rank0/log0.log command to view the output information. The output is as follows:

      train args: Namespace(dataset_path='./dataset/train', platform='CPU', \
      pretrain_ckpt='./pretrain_checkpoint/mobilenetv2_cpu_gpu.ckpt', freeze_layer='backbone')
      cfg: {'num_classes': 26, 'image_height': 224, 'image_width': 224, 'batch_size': 150, \
      'epoch_size': 200, 'warmup_epochs': 0, 'lr_max': 0.03, 'lr_end': 0.03, 'momentum': 0.9, \
      'weight_decay': 4e-05, 'label_smooth': 0.1, 'loss_scale': 1024, 'save_checkpoint': True, \
      'save_checkpoint_epochs': 1, 'keep_checkpoint_max': 20, 'save_checkpoint_path': './', \
      'platform': 'CPU'}
      Processing batch: 16: 100%|███████████████████████████████████████████ █████████████████████| 16/16 [00:00<?, ?it/s]
      epoch[200], iter[16] cost: 256.030, per step time: 256.030, avg loss: 1.775total cos 7.2574 s
      
  • Check the saved checkpoint files.

    • On Windows, run the dir checkpoint command to view the saved model files.

      dir ckpt_0
      2020//0814 11:20        267,727 mobilenetv2_1.ckpt
      2020//0814 11:21        267,727 mobilenetv2_10.ckpt
      2020//0814 11:21        267,727 mobilenetv2_11.ckpt
      ...
      2020//0814 11:21        267,727 mobilenetv2_7.ckpt
      2020//0814 11:21        267,727 mobilenetv2_8.ckpt
      2020//0814 11:21        267,727 mobilenetv2_9.ckpt
      
    • On Linux, run the ls ./checkpoint command to view the saved model files.

      ls ./ckpt_0/
      mobilenetv2_1.ckpt  mobilenetv2_2.ckpt
      mobilenetv2_3.ckpt  mobilenetv2_4.ckpt
      ...
      

Validating the Fine-Tuning Training Model

Validating the Model

Set mandatory parameters when using the validation set to test model performance. The default value of --platform is Ascend. You can set it to CPU or GPU. Finally, the standard output and error output are displayed in the interactive command line or written to the eval.log file.

# Windows/Linux with Python
python eval.py --platform CPU --dataset_path <VAL_DATASET_PATH> --pretrain_ckpt ./ckpt_0/mobilenetv2_15.ckpt

# Linux with Shell
sh run_eval.sh CPU <VAL_DATASET_PATH> ../ckpt_0/mobilenetv2_15.ckpt

Validation Result

When the python file is run, the validation result is output in the interactive command line. The shell script writes the information to ./eval.log. You need to run the cat ./eval.log command to view the information. The result is as follows:

result:{'acc': 0.9466666666666666666667}
pretrain_ckpt = ./ckpt_0/mobilenetv2_15.ckpt