菜单
Bayesian parameter estimation with MCMC (emcee) and probabilistic programming (PyMC). Posterior distributions, corner plots, model evidence, convergence diagnostics. Use when you need full posterior distributions, not just point estimates.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
基于 SOC 职业分类
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/synthetic-sciences/openscience --skill bayesian-inference命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Diffusion-based molecular docking. Predict protein-ligand binding poses from PDB/SMILES, confidence scores, virtual screening, for structure-based drug design. Not for affinity prediction.
Fast inference and fine-tuning platform with serverless and on-demand GPU deployments. OpenAI-compatible API for chat completions, embeddings, function calling, vision, and structured output. Supports SFT, DPO, and RL fine-tuning. SOC2 + HIPAA compliant.
Serverless inference, fine-tuning, embeddings, image generation, and batch processing on 200+ open-source models via an OpenAI-compatible API. Use when you need fast, cost-effective access to open-source LLMs without managing infrastructure.
| name | bayesian-inference |
| description | Bayesian parameter estimation with MCMC (emcee) and probabilistic programming (PyMC). Posterior distributions, corner plots, model evidence, convergence diagnostics. Use when you need full posterior distributions, not just point estimates. |
| category | physics |
| version | 1.0.0 |
| author | Synthetic Sciences |
| license | MIT |
| tags | ["Bayesian","MCMC","Posterior","emcee","PyMC","Inference","Uncertainty","Physics"] |
| dependencies | ["emcee>=3.1.0","corner>=2.2.0","scipy>=1.11.0","numpy>=1.24.0","matplotlib>=3.7.0"] |
Full Bayesian parameter estimation using MCMC sampling. Get posterior distributions (not just best-fit values), compute credible intervals, compare models via Bayes factors, and diagnose sampler convergence.
curve_fit with error bars is sufficient (use physics-fitting)import numpy as np
import emcee
import corner
import matplotlib.pyplot as plt
# Model and data
def model(t, A, gamma, omega):
return A * np.exp(-gamma * t) * np.cos(omega * t)
t_data = np.linspace(0, 10, 50)
y_true = model(t_data, 2.0, 0.3, 1.5)
y_err = 0.1 * np.ones_like(t_data)
y_data = y_true + y_err * np.random.randn(len(t_data))
# Log-probability = log-prior + log-likelihood
def log_prior(theta):
A, gamma, omega = theta
if A < 0 or gamma < 0 or omega < 0:
return -np.inf
A > gamma > omega > :
-np.inf
():
A, gamma, omega = theta
y_model = model(t, A, gamma, omega)
chi2 = np.(((y - y_model) / yerr)**)
- * chi2 - np.(np.log(yerr)) - * (y) * np.log(*np.pi)
():
lp = log_prior(theta)
np.isfinite(lp):
-np.inf
lp + log_likelihood(theta, t, y, yerr)
ndim =
nwalkers =
p0 = np.array([, , ]) + * np.random.randn(nwalkers, ndim)
sampler = emcee.EnsembleSampler(nwalkers, ndim, log_probability,
args=(t_data, y_data, y_err))
sampler.run_mcmc(p0, , progress=)
flat_samples = sampler.get_chain(discard=, thin=, flat=)
()
labels = [r"$A$", r"$\gamma$", r"$\omega$"]
truths = [2.0, 0.3, 1.5]
fig = corner.corner(flat_samples, labels=labels, truths=truths,
quantiles=[0.16, 0.5, 0.84],
show_titles=True, title_kwargs={"fontsize": 12})
fig.savefig('corner_plot.png', dpi=150, bbox_inches='tight')
# Print parameter estimates
for i, label in enumerate(labels):
mcmc = np.percentile(flat_samples[:, i], [16, 50, 84])
q = np.diff(mcmc)
print(f"{label} = {mcmc[1]:.4f} (+{q[1]:.4f} / -{q[0]:.4f})")
# Trace plots (check mixing)
fig, axes = plt.subplots(ndim, 1, figsize=(10, 7), sharex=True)
chain = sampler.get_chain()
for i in range(ndim):
axes[i].plot(chain[:, :, i], alpha=0.3, linewidth=0.5)
axes[i].set_ylabel(labels[i])
axes[i].axvline(1000, color='red', linestyle='--', label='Burn-in')
axes[-1].set_xlabel('Step')
axes[0].legend(loc='upper right')
plt.savefig('trace_plots.png', dpi=150, bbox_inches='tight')
# Autocorrelation time (convergence diagnostic)
try:
tau = sampler.get_autocorr_time()
print(f"Autocorrelation times: {tau}")
print(f"Chain length / tau: {sampler.iteration / tau}")
print(f" (Should be > 50 for reliable results)")
except emcee.autocorr.AutocorrError:
print("Warning: chain too short for autocorrelation estimate")
# Gelman-Rubin diagnostic (R-hat)
chain_split = sampler.get_chain(discard=1000) # (steps, walkers, ndim)
n_steps, n_walkers, n_dim = chain_split.shape
# Between-chain variance
chain_means = chain_split.mean(axis=0) # (walkers, ndim)
B = n_steps * np.var(chain_means, axis=0, ddof=1)
# Within-chain variance
W = np.mean(np.var(chain_split, axis=0, ddof=1), axis=0)
# R-hat
var_est = (1 - 1/n_steps) * W + B / n_steps
R_hat = np.sqrt(var_est / W)
print(f"R-hat: {R_hat}")
print(f" (Should be < 1.01 for convergence)")
# Sample model predictions from the posterior
t_pred = np.linspace(0, 12, 200)
n_samples = 100
idx = np.random.randint(len(flat_samples), size=n_samples)
fig, ax = plt.subplots(figsize=(10, 6))
for i in idx:
y_pred = model(t_pred, *flat_samples[i])
ax.plot(t_pred, y_pred, 'r-', alpha=0.05)
ax.errorbar(t_data, y_data, yerr=y_err, fmt='ko', capsize=3, label='Data')
ax.plot(t_pred, model(t_pred, 2.0, 0.3, 1.5), 'b-', linewidth=2, label='True')
ax.set_xlabel('Time [s]')
ax.set_ylabel('y')
ax.set_title('Posterior Predictive Check')
ax.legend()
ax.grid(True, alpha=0.3)
plt.savefig('posterior_predictive.png', dpi=150, bbox_inches='tight')
# WARNING: Harmonic mean estimator is unreliable for high dimensions.
# For serious model comparison, use nested sampling (dynesty).
# Simple approach: compare max log-likelihood
log_likes_1 = sampler.get_log_prob(discard=1000, flat=True)
print(f"Model 1: max log(L) = {np.max(log_likes_1):.2f}")
print(f" mean log(L) = {np.mean(log_likes_1):.2f}")
# For proper Bayes factors, use dynesty:
# pip install dynesty
# from dynesty import NestedSampler
# sampler = NestedSampler(log_likelihood, prior_transform, ndim)
# sampler.run_nested()
# results = sampler.results
# print(f"log(Z) = {results.logz[-1]:.2f} ± {results.logzerr[-1]:.2f}")
| Symptom | Fix |
|---|---|
| Walkers stuck (low acceptance) | Prior too restrictive or likelihood too peaked — reparameterize |
| Very long autocorrelation | Correlated parameters — reparameterize (e.g., log-space) |
| Multimodal posterior | Use parallel tempering or nested sampling |
| Corner plot shows hard edges | Prior bounds too tight — widen them |
| Trace plots show drift | Not converged — run longer or improve initialization |