aszer/bet
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

格式化代码

Browse Source
This commit is contained in:
aszerW
2025-10-25 16:39:27 +08:00
parent 4e32725279
commit eb057c236e
1 changed files with 30 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
pinnacle_experiments.py
View File

@@ -5,11 +5,13 @@ from sklearn.linear_model import LogisticRegression
from sklearn.metrics import log_loss from sklearn.metrics import log_loss
def compute_metrics(df: pd.DataFrame, def compute_metrics(
df: pd.DataFrame,
n_bins: int = 10, n_bins: int = 10,
bin_strategy: str = 'uniform', # 'uniform' or 'quantile' bin_strategy: str = "uniform", # 'uniform' or 'quantile'
include_draws: bool = True, include_draws: bool = True,
eps: float = 1e-6) -> dict: eps: float = 1e-6,
) -> dict:
""" """
计算预测评估指标并拟合校准关系。 计算预测评估指标并拟合校准关系。
@@ -25,14 +27,14 @@ def compute_metrics(df: pd.DataFrame,
""" """
# 处理 refunded # 处理 refunded
if include_draws: if include_draws:
mask = df['res'].isin(['won', 'refunded', 'lost']) mask = df["res"].isin(["won", "refunded", "lost"])
else: else:
mask = df['res'].isin(['won', 'lost']) mask = df["res"].isin(["won", "lost"])
df = df[mask].copy() df = df[mask].copy()
# 标签: won=1, others=0 (包括 refunded) # 标签: won=1, others=0 (包括 refunded)
y = df['res'].map({'won': 1, 'refunded': 0, 'lost': 0}).astype(int).values y = df["res"].map({"won": 1, "refunded": 0, "lost": 0}).astype(int).values
p = df['win_prob'].astype(float).values p = df["win_prob"].astype(float).values
# 裁剪概率以保证数值稳定 # 裁剪概率以保证数值稳定
p_clip = np.clip(p, eps, 1 - eps) p_clip = np.clip(p, eps, 1 - eps)
@@ -48,7 +50,7 @@ def compute_metrics(df: pd.DataFrame,
brier = float(np.mean((p_clip - y) ** 2)) brier = float(np.mean((p_clip - y) ** 2))
# ECE 计算（支持 uniform 或 quantile） # ECE 计算（支持 uniform 或 quantile）
if bin_strategy == 'quantile': if bin_strategy == "quantile":
# quantile bin edges # quantile bin edges
try: try:
edges = np.unique(np.percentile(p_clip, np.linspace(0, 100, n_bins + 1))) edges = np.unique(np.percentile(p_clip, np.linspace(0, 100, n_bins + 1)))
@@ -57,7 +59,9 @@ def compute_metrics(df: pd.DataFrame,
bin_idxs = np.minimum((p_clip * n_bins).astype(int), n_bins - 1) bin_idxs = np.minimum((p_clip * n_bins).astype(int), n_bins - 1)
else: else:
# searchsorted to assign bins # searchsorted to assign bins
bin_idxs = np.clip(np.searchsorted(edges, p_clip, side='right') - 1, 0, len(edges) - 2) bin_idxs = np.clip(
np.searchsorted(edges, p_clip, side="right") - 1, 0, len(edges) - 2
)
except Exception: except Exception:
bin_idxs = np.minimum((p_clip * n_bins).astype(int), n_bins - 1) bin_idxs = np.minimum((p_clip * n_bins).astype(int), n_bins - 1)
else: else:
@@ -70,39 +74,40 @@ def compute_metrics(df: pd.DataFrame,
idx = bin_idxs == b idx = bin_idxs == b
count = int(idx.sum()) count = int(idx.sum())
if count == 0: if count == 0:
bin_stats.append({'count': 0, 'mean_pred': float('nan'), 'emp_freq': float('nan')}) bin_stats.append(
{"count": 0, "mean_pred": float("nan"), "emp_freq": float("nan")}
)
continue continue
mean_pred = float(p_clip[idx].mean()) mean_pred = float(p_clip[idx].mean())
emp_freq = float(y[idx].mean()) emp_freq = float(y[idx].mean())
ece += abs(mean_pred - emp_freq) * count ece += abs(mean_pred - emp_freq) * count
bin_stats.append({'count': count, 'mean_pred': mean_pred, 'emp_freq': emp_freq}) bin_stats.append({"count": count, "mean_pred": mean_pred, "emp_freq": emp_freq})
ece = float(ece / total) if total > 0 else float('nan') ece = float(ece / total) if total > 0 else float("nan")
# accuracy # accuracy
acc = float(np.mean((p_clip >= 0.5) == (y == 1))) acc = float(np.mean((p_clip >= 0.5) == (y == 1)))
# 校准拟合: 使用 LogisticRegression 拟合 logit(E[y]) = alpha + beta * logit(p) # 校准拟合: 使用 LogisticRegression 拟合 logit(E[y]) = alpha + beta * logit(p)
X = sp_logit(p_clip).reshape(-1, 1) X = sp_logit(p_clip).reshape(-1, 1)
clf = LogisticRegression(C=1e6, solver='lbfgs', max_iter=200) clf = LogisticRegression(C=1e6, solver="lbfgs", max_iter=200)
clf.fit(X, y) clf.fit(X, y)
alpha = float(clf.intercept_[0]) alpha = float(clf.intercept_[0])
beta = float(clf.coef_[0][0]) beta = float(clf.coef_[0][0])
return { return {
'logloss': logloss, "logloss": logloss,
'brier': brier, "brier": brier,
'ece': ece, "ece": ece,
'accuracy': acc, "accuracy": acc,
'reg_alpha': alpha, "reg_alpha": alpha,
'reg_beta': beta, "reg_beta": beta,
# 'ece_bins': bin_stats, # 'ece_bins': bin_stats,
'n_samples': int(total) "n_samples": int(total),
} }
if __name__ == "__main__":
if __name__ == '__main__':
df = pd.read_feather("data/p_res.feather") df = pd.read_feather("data/p_res.feather")
df['win_prob'] = df['power_p'] df["win_prob"] = df["power_p"]
res = compute_metrics(df) res = compute_metrics(df)
print(res) print(res)