Skip to content

API.md

Latest commit

 

History

History
226 lines (162 loc) · 3.75 KB

API.md

File metadata and controls

226 lines (162 loc) · 3.75 KB
Raw

DRL MuJoCo Web API

FastAPI 默认监听 0.0.0.0:8000。接口当前无鉴权，设计用于本机或 SSH 隧道访问，不应直接暴露到公网。

基础地址：

http://localhost:8000

除视频文件外，响应均为 JSON。启动类接口当前以 HTTP 200 返回业务状态，调用方必须检查 status 字段。

环境与指标

GET /api/environments

curl 'http://localhost:8000/api/environments'
{
  "environments": {
    "hopper": {
      "name": "Hopper-v5",
      "description": "...",
      "difficulty": "..."
    }
  },
  "active": "hopper"
}

GET /api/metrics/distributed?env=<env>

GET /api/metrics/single?env=<env>

env 可取 hopper、walker2d、halfcheetah，缺省为 hopper。返回 CSV 行转换后的 JSON 数组。

curl 'http://localhost:8000/api/metrics/distributed?env=hopper'
[
  {
    "step": 0,
    "elapsed_sec": 2.1,
    "total_steps": 16384.0,
    "sps": 7801.9,
    "episodes": 12,
    "avg_return": 18.2,
    "buffer_size": 16384.0,
    "loss": 0.31,
    "num_gpus": 1.0
  }
]

训练控制

POST /api/training/distributed/start?env=<env>

POST /api/training/single/start?env=<env>

请求无 JSON body：

curl -X POST 'http://localhost:8000/api/training/distributed/start?env=hopper'

成功：

{"status": "started"}

重复启动：

{"status": "already running"}

立即崩溃：

{
  "status": "error",
  "returncode": 1,
  "error_detail": "Traceback ..."
}

POST /api/training/stop?env=<env>

省略 env 会停止当前服务进程管理的全部训练。

curl -X POST 'http://localhost:8000/api/training/stop?env=hopper'
{"status": "stopped", "environments": ["hopper"]}

GET /api/training/status?env=<env>

{"env": "hopper", "running": true}

不传 env：

{"statuses": {"hopper": true, "walker2d": false, "halfcheetah": false}}

GET /api/training/logs?env=<env>&lines=<n>

返回内存中的最近训练输出。lines 缺省为 50。

{"env": "hopper", "lines": ["[Main] ..."], "total": 120}

GPU 扩展实验

GET /api/scaling/configs

扫描 config/scaling/*.yaml：

{
  "configs": {
    "hopper_gpu4": {
      "env": "hopper",
      "num_gpus": 4,
      "config_path": "config/scaling/hopper_gpu4.yaml",
      "metrics_path": "output/scaling/hopper_gpu4/metrics.csv"
    }
  }
}

GET /api/scaling/metrics?config_name=<name>

POST /api/scaling/start?config_name=<name>

curl -X POST 'http://localhost:8000/api/scaling/start?config_name=hopper_gpu4'
{
  "status": "started",
  "config": {"env": "hopper", "num_gpus": 4, "config_path": "..."}
}

GET /api/scaling/status

{"statuses": {"hopper_gpu4": true}}

GET /api/cluster/info

{
  "gpu_count": 1,
  "gpus": [{"index": 0, "name": "NVIDIA ...", "memory_gb": 24.0}],
  "ray_resources": {}
}

视频

POST /api/videos/generate?env=<env>

异步生成单机与分布式策略视频。

GET /api/videos/status?env=<env>

返回 idle、generating、completed 或 error 状态及错误信息。

GET /api/videos/distributed?env=<env>

GET /api/videos/single?env=<env>

成功时返回 MP4 文件；文件不存在时返回 JSON：

{"error": "Video not found"}

WebSocket

WS /ws/training

连接：

ws://localhost:8000/ws/training

服务端每秒推送当前活跃环境的指标：

{
  "type": "metrics",
  "env": "hopper",
  "distributed": [],
  "single": []
}

训练退出时推送：

{
  "type": "training_stopped",
  "env": "hopper",
  "returncode": 1,
  "error_detail": "可选错误详情"
}