AutoML：训练模型以识别凹坑

我们该怎么办？

1-准备数据集

• _annotations.coco.json，一个包含images、categories和annotations的json文件
• 图像本身（.jpg文件）

• images：包含有关数据集图像的信息（id、文件名、大小等）
• categories：边界框类别的名称和id
• annotations：包含有关对象的信息，包括边界框坐标（在此数据集中，它们位于绝对坐标）、对象的图像id和类别id

from azureml.core import Workspace

workspace = Workspace.from_config() # 如果你在一个aml计算实例上运行notebook，这是可行的

default_datastore = workspace.get_default_datastore()
datastore_name = default_datastore.name

{
   "image_url":"AmlDatastore://data_directory/../Image_name.image_format",
   "image_details":{
      "format":"image_format",
      "width":"image_width",
      "height":"image_height"
   },
   "label":[
      {
         "label":"class_name_1",
         "topX":"xmin/width",
         "topY":"ymin/height",
         "bottomX":"xmax/width",
         "bottomY":"ymax/height",
         "isCrowd":"isCrowd"
      },
      {
         "label":"class_name_2",
         "topX":"xmin/width",
         "topY":"ymin/height",
         "bottomX":"xmax/width",
         "bottomY":"ymax/height",
         "isCrowd":"isCrowd"
      },
      "..."
   ]
}

# 从coco文件生成训练jsonl文件
!python coco2jsonl.py \
--input_coco_file_path "./potholeObjects/train/_annotations.coco.json" \
--output_dir "./potholeObjects/train" --output_file_name "train_pothole_from_coco.jsonl" \
--task_type "ObjectDetection" \
--base_url "AmlDatastore://{datastore_name}/potholeObjects/train/"

from azureml.core import Dataset
from azureml.data.datapath import DataPath
from azureml.data import DataType

# 上传文件到数据存储
Dataset.File.upload_directory(
    src_dir="./potholeObjects/train/", target=DataPath(default_datastore, "/potholeObjects/train"), show_progress=True
)

training_dataset_name = "potholeObjectesTrainingDataset"

# 创建数据集
training_dataset = Dataset.Tabular.from_json_lines_files(
        path=DataPath(default_datastore, "/potholeObjects/train/train_pothole_from_coco.jsonl"),
        set_column_types={"image_url": DataType.to_stream(default_datastore.workspace)},
)

# 在工作区中注册数据集
training_dataset = training_dataset.register(
        workspace=workspace, name=training_dataset_name
)

2.进行实验

from azureml.core import Experiment

experiment_name = "pothole-yolov5-model"
experiment = Experiment(workspace, name=experiment_name)

compute_target = workspace.compute_targets['gpu-computer'] # 给计算实例的名称

from azureml.automl.core.shared.constants import ImageTask
from azureml.train.automl import AutoMLImageConfig
from azureml.train.hyperdrive import GridParameterSampling, choice

automl_config_yolov5 = AutoMLImageConfig(
    task=ImageTask.IMAGE_OBJECT_DETECTION,
    compute_target=compute_target,
    training_data=training_dataset,
    validation_data=validation_dataset,
    hyperparameter_sampling=GridParameterSampling({"model_name": choice("yolov5")}),
    iterations=1,
)

automl_image_run = experiment.submit(automl_config_yolov5)

from azureml.automl.core.shared.constants import ImageTask
from azureml.train.automl import AutoMLImageConfig
from azureml.train.hyperdrive import BanditPolicy, RandomParameterSampling
from azureml.train.hyperdrive import choice, uniform

parameter_space = {
    "model": choice(
        {
            "model_name": choice("yolov5"),
            "learning_rate": uniform(0.0001, 0.01),
            "model_size": choice("small", "medium"),  # 模型相关
            #'img_size': choice(640, 704, 768), # 特定型号;可能需要有大内存的GPU
        },
        {
            "model_name": choice("fasterrcnn_resnet50_fpn"),
            "learning_rate": uniform(0.0001, 0.001),
            "optimizer": choice("sgd", "adam", "adamw"),
            "min_size": choice(600, 800),  # 模型相关
            #'warmup_cosine_lr_warmup_epochs': choice(0, 3),
        },
    ),
}

tuning_settings = {
    "iterations": 10,
    "max_concurrent_iterations": 2,
    "hyperparameter_sampling": RandomParameterSampling(parameter_space),
    "early_termination_policy": BanditPolicy(
        evaluation_interval=2, slack_factor=0.2, delay_evaluation=6
    ),
}

automl_image_config = AutoMLImageConfig(
    task=ImageTask.IMAGE_OBJECT_DETECTION,
    compute_target=compute_target,
    training_data=training_dataset,
    validation_data=validation_dataset,
    **tuning_settings,
)

3-将预测可视化

best_child_run = automl_image_run.get_best_child()

model_name = best_child_run.properties["model_name"]
model_output_path = best_child_run.properties["model_output_path"]

# 从最佳运行中注册模型
model = best_child_run.register_model(
    model_name=model_name, model_path=model_output_path
)

from azureml.contrib.automl.dnn.vision.common.model_export_utils import load_model, run_inference
from azureml.contrib.automl.dnn.vision.object_detection_yolo.writers.score import _score_with_model

TASK_TYPE = 'image-object-detection'
model_settings = {"img_size": 640, "model_size": "medium", "box_score_thresh": 0.1, "box_iou_thresh": 0.5}

model_wrapper = load_model(TASK_TYPE, 'model.pt', **model_settings)

sample_image = "./img-23_jpg.rf.e6aa0daf83e72ccbf1ea10eb6a6ab3bd.jpg"
with open(sample_image, 'rb') as f:
    bytes_img = f.read()
    
model_response = run_inference(model_wrapper, bytes_img, _score_with_model)

最后