Hooks

详细API文档： Hook System

训练指标统计大盘（阿里内部功能）

参考文档：nebula-mltracker

# trainer: Trainer
# project_name: 指标统计项目名称
# experiment_name: 指标统计实验名称
# track_config: 指标统计其他配置

if os.environ['RANK'] == '0':
    ml_tracker_hook = MLTrackerHook(
        project_name,
        experiment_name,
        track_config,
    )
    trainer.add_hooks([ml_tracker_hook])

预测结果Trace到ODPS（阿里内部功能）

# trainer: Trainer
# fields: 预测结果列名
# fields = ["id", "preds", "labels"]
# types: 预测结果列类型
# types = ["string", "string", "string"]

trace_hook_config = {
    "access_id": "xxx",
    "access_key": "xxx",
    "end_point": "xxx",
    "project": "xxx",
    "table_name": "xxx",
    "partition": "ds1=xxx,ds2=xxx",
}
trace_hook = TraceToOdpsHook(
    config=trace_hook_config, fields=fields, types=types, worker_num=8
)
trainer.add_hooks([trace_hook])

Timeline分析

# trainer: Trainer
# output_dir: 输出目录

if int(os.environ.get("RANK", 0)) == 0:
    hooks = ProfilerHook(
        wait=1,
        warmup=249,
        active=1,
        repeat=2,
        output_dir=output_dir,
    )
    trainer.add_hook(hooks)