NNI工具包使用

·第一步：下载安装NNI，参照https://github.com/microsoft/nni
·第二步：找到baseline中的超参数，并改为通过NNI传入超参
·第三步：在项目中新建search space.json，config.yml·第四步：配置search space.json，将上一步找到的超参配置搜索范围
·第五步：在训练文件中加入上报指标
·第六步：配置config.ym
·第七步：运行NNI1，并进入webui查看是否成功运行
·第八步：等待·第九步：分析

nni的启动

nnictl create -(xxxx.yml)[这是创建的配置文件]

 

第一步 定义搜索空间

 

不同的数据增强

不同的优化器

{
  "optimizer":{"_type":"choice", "_value":["Adam", "Adamax", "Adagrad", "RMSprop", "Adagrad"]},
  "Transpose":{"_type":"choice", "_value":[0.3, 0.4, 0.5]},
  "HorizontalFlip":{"_type":"choice", "_value":[0.3, 0.4, 0.5]},
  "VerticalFlip":{"_type":"choice", "_value":[0.3, 0.4, 0.5]},
  "ShiftScaleRotate":{"_type":"choice", "_value":[0.3, 0.4, 0.5]},
  "hue_shift_limit":{"_type":"choice", "_value":[0.2, 0.3, 0.4]},
  "sat_shift_limit":{"_type":"choice", "_value":[0.2, 0.3, 0.4]},
  "val_shift_limit":{"_type":"choice", "_value":[0.2, 0.3, 0.4]},
  "HueSaturationValue":{"_type":"choice", "_value":[0.3, 0.4, 0.5]}
}

在没有nni的代码上加nni

try:
        tuner_params = nni.get_next_parameter()
        optimizer_type = tuner_params['optimizer']

def get_train_transforms(data_aug_param):
    # return Compose([
    #     RandomResizedCrop(CFG['img_size'], CFG['img_size']),
    #     Transpose(p=0.5),
    #     HorizontalFlip(p=0.5),
    #     VerticalFlip(p=0.5),
    #     ShiftScaleRotate(p=0.5),
    #     HueSaturationValue(hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5),
    #     RandomBrightnessContrast(brightness_limit=(-0.1, 0.1), contrast_limit=(-0.1, 0.1), p=0.5),
    #     Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),
    #     CoarseDropout(p=0.5),
    #     Cutout(p=0.5),
    #     ToTensorV2(p=1.0),
    # ], p=1.)
    return Compose([
        RandomResizedCrop(CFG['img_size'], CFG['img_size']),
        Transpose(p=data_aug_param['Transpose']),
        HorizontalFlip(p=data_aug_param['HorizontalFlip']),
        VerticalFlip(p=data_aug_param['VerticalFlip']),
        ShiftScaleRotate(p=data_aug_param['ShiftScaleRotate']),
        HueSaturationValue(hue_shift_limit=data_aug_param['hue_shift_limit'], sat_shift_limit=data_aug_param['sat_shift_limit'], val_shift_limit=data_aug_param['val_shift_limit'], p=data_aug_param['HueSaturationValue']),
        RandomBrightnessContrast(brightness_limit=(-0.1, 0.1), contrast_limit=(-0.1, 0.1), p=0.5),
        Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),
        CoarseDropout(p=0.5),
        Cutout(p=0.5),
        ToTensorV2(p=1.0),
    ], p=1.)

上报中间精度和最终指标

 

设置配置文件

 

authorName: default
experimentName: cldc
trialConcurrency: 1
maxExecDuration: 24h
maxTrialNum: 50
#choice: local, remote, pai
trainingServicePlatform: local
searchSpacePath: search_space.json
#choice: true, false
useAnnotation: false
tuner:
  #choice: TPE, Random, Anneal, Evolution, BatchTuner, MetisTuner, GPTuner
  #SMAC (SMAC should be installed through nnictl)
  builtinTunerName: TPE
  classArgs:
    #choice: maximize, minimize
    optimize_mode: maximize
trial:
  command: python train_nni.py  #训练用的代码
  codeDir: .
  gpuNum: 1    #gpu数量，一定记得改
localConfig:
  useActiveGpu: true

分析可视化结果