download-from-swanlab-url

Skill: Download metric data from a SwanLab run URL

Goal

Given a SwanLab cloud URL of the form

https://swanlab.cn/@<username>/<project>/runs/<exp_cuid>/chart

fetch the per-step time-series data for one or more metrics (e.g. reward, entropy, response length) as a pandas.DataFrame for downstream plotting or analysis.

Trial-and-error log (read first; saves hours)

The shortcuts you might try do not work:

1. WebFetch / curl of the chart page — the page is a Vue SPA. The HTML body is just <div id="app"></div>; no chart data is embedded. Don't bother scraping HTML.

2. Direct REST probe (curl https://api.swanlab.cn/api/v1/runs/<cuid>) — returns 404 Not Found even with a valid api key. The public REST surface is not the v1 path; use the Python SDK.

3. swanlab verify says you're logged in — but the env may point to a private cloud. On this host:

   $ swanlab verify
   swanlab: You are logged into https://cloud-20.agent-matrix.com as fuqingxu
   

   This is not swanlab.cn. The shell exports SWANLAB_API_HOST and SWANLAB_WEB_HOST in ~/.bashrc, redirecting the SDK to a different deployment. Calling OpenApi(api_key=...) then fails with Login failed: 404 Not Found because /login/api_key only exists on swanlab.cn's API host.

4. Metric keys are NOT what the UI shows. The chart card titled "reward" is logged as critic/rewards/mean. Asking get_metrics(keys=['reward']) returns code 404. You must probe candidate keys (see the cheat sheet below).

5. get_summary to enumerate keys — does not work for non-cloned runs. When rootExpId / rootProId are None (i.e. the run was not cloned from another project), experiment.get_summary returns HTTP 400 "Bad Request". Skip it; probe candidate keys directly with get_metrics.

The working recipe

1. Prerequisites

• swanlab Python package (>= 0.7.4 confirmed).
• An api key for swanlab.cn. On this machine it is stored in ~/.swanlab/.netrc:

  machine https://api.swanlab.cn
      login https://swanlab.cn
      password <API_KEY>
  

  Read the password field from there if you don't have it on hand.

2. Parse the URL

import re

URL = "https://swanlab.cn/@binaryhusky/spy-game-rl/runs/zku3ujg2k3unvt61jbu0s/chart"
m = re.match(r"https?://swanlab\.cn/@([^/]+)/([^/]+)/runs/([^/]+)", URL)
username, project, exp_id = m.group(1), m.group(2), m.group(3)

3. Override env vars BEFORE importing swanlab

Critical: if your shell has SWANLAB_API_HOST / SWANLAB_WEB_HOST / SWANLAB_API_KEY pointing to a private cloud, the SDK silently uses them. Either run via env -i or unset them:

unset SWANLAB_API_KEY SWANLAB_API_HOST SWANLAB_WEB_HOST
SWANLAB_API_HOST=https://api.swanlab.cn/api \
SWANLAB_WEB_HOST=https://swanlab.cn \
python your_script.py

Note: SWANLAB_API_HOST for swanlab.cn ends with /api (the SDK default in swanlab/env.py); without that suffix login also returns 404.

4. Open the API and fetch metadata

import swanlab
api = swanlab.OpenApi(api_key="<API_KEY>")  # raises ValidationError if hosts wrong

exp = api.get_experiment(project=project, exp_id=exp_id, username=username)
assert exp.code == 200, exp.errmsg
print(exp.data.name, exp.data.state)
# exp.data.profile['config'] contains the full training config (verl-style nested dict)

5. Discover the right metric keys

The keys you ask for must match what was logged. For a verl-based AgentJet run, the working names are:

Other common ones: response_length/max, response_length/min, critic/score/mean. Probe defensively — get_metrics returns code=200 and N rows on hit, code=404, rows=0 on miss:

candidates = [
    "critic/rewards/mean", "critic/score/mean",
    "actor/entropy", "actor/entropy_loss", "actor/pg_loss",
    "response_length/mean", "response_length/max", "response_length/min",
]
for k in candidates:
    r = api.get_metrics(exp_id=exp_id, keys=k)
    print(f"{k:40s} code={r.code} rows={0 if r.data is None else len(r.data)}")

6. Fetch and save

keys = ["critic/rewards/mean", "actor/entropy", "response_length/mean"]
r = api.get_metrics(exp_id=exp_id, keys=keys)
df = r.data  # indexed by `step`; one column per key + one `<key>_timestamp` column
df.to_csv("metrics.csv")

The DataFrame layout is:

              actor/entropy  actor/entropy_timestamp  critic/rewards/mean  ...
step
1                  0.5569            1774003810000               0.7271
2                  0.5732            1774004325000               0.7589
...

The _timestamp columns are unix millis; usually you can drop them and plot against the step index.

End-to-end runnable snippet

"""Fetch reward/entropy/response_length curves from a swanlab.cn run URL."""
import os, re, sys
# Strip any private-cloud overrides BEFORE importing swanlab.
for v in ("SWANLAB_API_KEY", "SWANLAB_API_HOST", "SWANLAB_WEB_HOST"):
    os.environ.pop(v, None)
os.environ["SWANLAB_API_HOST"] = "https://api.swanlab.cn/api"
os.environ["SWANLAB_WEB_HOST"] = "https://swanlab.cn"

import swanlab

URL = sys.argv[1]
API_KEY = sys.argv[2]  # or read from ~/.swanlab/.netrc
m = re.match(r"https?://swanlab\.cn/@([^/]+)/([^/]+)/runs/([^/]+)", URL)
username, project, exp_id = m.groups()

api = swanlab.OpenApi(api_key=API_KEY)
keys = ["critic/rewards/mean", "actor/entropy", "response_length/mean"]
r = api.get_metrics(exp_id=exp_id, keys=keys)
assert r.code == 200, r.errmsg
df = r.data.rename(columns={
    "critic/rewards/mean": "reward",
    "actor/entropy": "entropy",
    "response_length/mean": "response_length",
})
df.to_csv("metrics.csv")
print(df.head())

Plot recipe (seaborn, optional)

import seaborn as sns, matplotlib.pyplot as plt, pandas as pd
df = pd.read_csv("metrics.csv").sort_values("step")
sns.set_theme(context="paper", style="whitegrid")
fig, axes = plt.subplots(1, 3, figsize=(11.5, 3.0))
palette = sns.color_palette("deep")
for ax, (col, title, c) in zip(axes, [
    ("reward", "Reward", palette[0]),
    ("entropy", "Policy Entropy", palette[3]),
    ("response_length", "Response Length", palette[2]),
]):
    sns.lineplot(data=df, x="step", y=col, ax=ax, color=c, linewidth=1.6)
    ax.set_title(title); ax.set_xlabel("Training step"); ax.set_ylabel("")
fig.tight_layout()
fig.savefig("curves.pdf", bbox_inches="tight")

Troubleshooting cheat sheet