CLI (v2) MLtable YAML 架构

适用于：Azure CLI ml 扩展 v2（当前）

可以从 https://azuremlschemas.azureedge.net/latest/MLTable.schema.json 找到源 JSON 架构。

如何创作 MLTable 文件

本文仅提供了有关 MLTable YAML 架构的信息。 有关 MLTable 的详细信息，包括

• MLTable 文件创作
• MLTable 项目创建
• 在 Pandas 和 Spark 中的使用
• 端到端示例

请访问在 Azure 机器学习中使用表。

YAML 语法

转换

读取转换

其他转换

示例

MLTable 的使用示例。 从以下位置查找更多示例：

• 在 Azure 机器学习中使用表
• 示例 GitHub 存储库

快速入门

本快速入门从公共 https 服务器读取著名的 iris 数据集。 若要继续，必须将 MLTable 文件放在一个文件夹中。 首先，使用以下命令创建该文件夹和 MLTable 文件：

mkdir ./iris
cd ./iris
touch ./MLTable

接下来，将此内容置于 MLTable 文件中：

$schema: https://azuremlschemas.azureedge.net/latest/MLTable.schema.json

type: mltable
paths:
    - file: https://azuremlexamples.blob.core.chinacloudapi.cn/datasets/iris.csv

transformations:
    - read_delimited:
        delimiter: ','
        header: all_files_same_headers
        include_path_column: true

然后，可以使用以下方法具体化到 Pandas 中：

import mltable

tbl = mltable.load("./iris")
df = tbl.to_pandas_dataframe()

确保数据包含名为 Path 的新列。 此列包含 https://azuremlexamples.blob.core.chinacloudapi.cn/datasets/iris.csv 数据路径。

此 CLI 可以创建数据资产：

az ml data create --name iris-from-https --version 1 --type mltable --path ./iris

包含 MLTable 的文件夹会自动上传到云存储（默认的 Azure 机器学习数据存储）。

带分隔符的文本文件

$schema: https://azuremlschemas.azureedge.net/latest/MLTable.schema.json
type: mltable

# Supported paths include:
# local: ./<path>
# blob: wasbs://<container_name>@<account_name>.blob.core.chinacloudapi.cn/<path>
# Public http(s) server: https://<url>
# ADLS gen2: abfss://<file_system>@<account_name>.dfs.core.chinacloudapi.cn/<path>/
# Datastore: azureml://subscriptions/<subid>/resourcegroups/<rg>/workspaces/<ws>/datastores/<datastore_name>/paths/<path>

paths:
  - file: abfss://<file_system>@<account_name>.dfs.core.chinacloudapi.cn/<path>/ # a specific file on ADLS
  # additional options
  # - folder: ./<folder> a specific folder
  # - pattern: ./*.csv # glob all the csv files in a folder

transformations:
    - read_delimited:
        encoding: ascii
        header: all_files_same_headers
        delimiter: ","
        include_path_column: true
        empty_as_string: false
    - keep_columns: [col1, col2, col3, col4, col5, col6, col7]
    # or you can drop_columns...
    # - drop_columns: [col1, col2, col3, col4, col5, col6, col7]
    - convert_column_types:
        - columns: col1
          column_type: int
        - columns: col2
          column_type:
            datetime:
                formats:
                    - "%d/%m/%Y"
        - columns: [col1, col2, col3] 
          column_type:
            boolean:
                mismatch_as: error
                true_values: ["yes", "true", "1"]
                false_values: ["no", "false", "0"]
      - filter: 'col("col1") > 32 and col("col7") == "a_string"'
      # create a column called timestamp with the values extracted from the folder information
      - extract_columns_from_partition_format: {timestamp:yyyy/MM/dd}
      - skip: 10
      - take_random_sample:
          probability: 0.50
          seed: 1394
      # or you can take the first n records
      # - take: 200

Parquet

$schema: https://azuremlschemas.azureedge.net/latest/MLTable.schema.json
type: mltable

# Supported paths include:
# local: ./<path>
# blob: wasbs://<container_name>@<account_name>.blob.core.chinacloudapi.cn/<path>
# Public http(s) server: https://<url>
# ADLS gen2: abfss://<file_system>@<account_name>.dfs.core.chinacloudapi.cn/<path>/
# Datastore: azureml://subscriptions/<subid>/resourcegroups/<rg>/workspaces/<ws>/datastores/<datastore_name>/paths/<path>

paths:
  - pattern: azureml://subscriptions/<subid>/resourcegroups/<rg>/workspaces/<ws>/datastores/<datastore_name>/paths/<path>/*.parquet
  
transformations:
  - read_parquet:
        include_path_column: false
  - filter: 'col("temperature") > 32 and col("location") == "UK"'
  - skip: 1000 # skip first 1000 rows
  # create a column called timestamp with the values extracted from the folder information
  - extract_columns_from_partition_format: {timestamp:yyyy/MM/dd}

Delta Lake

$schema: https://azuremlschemas.azureedge.net/latest/MLTable.schema.json
type: mltable

# Supported paths include:
# local: ./<path>
# blob: wasbs://<container_name>@<account_name>.blob.core.chinacloudapi.cn/<path>
# Public http(s) server: https://<url>
# ADLS gen2: abfss://<file_system>@<account_name>.dfs.core.chinacloudapi.cn/<path>/
# Datastore: azureml://subscriptions/<subid>/resourcegroups/<rg>/workspaces/<ws>/datastores/<datastore_name>/paths/<path>

paths:
- folder: abfss://<file_system>@<account_name>.dfs.core.chinacloudapi.cn/<path>/

# NOTE: for read_delta_lake, you are *required* to provide either
# timestamp_as_of OR version_as_of.
# timestamp should be in RFC-3339/ISO-8601 format (for example:
# "2022-10-01T00:00:00Z", "2022-10-01T00:00:00+08:00",
# "2022-10-01T01:30:00-08:00")
# To get the latest, set the timestamp_as_of at a future point (for example: '2999-08-26T00:00:00Z')

transformations:
 - read_delta_lake:
      timestamp_as_of: '2022-08-26T00:00:00Z'
      # alternative:
      # version_as_of: 1   

JSON

$schema: https://azuremlschemas.azureedge.net/latest/MLTable.schema.json
paths:
  - file: ./order_invalid.jsonl
transformations:
  - read_json_lines:
        encoding: utf8
        invalid_lines: drop
        include_path_column: false

后续步骤

• 安装并使用 CLI (v2)
• 在 Azure 机器学习中使用表