我正在尝试实现一个带有自定义目标函数的lightGBM分类器。我的目标数据有四个类别，并且我的数据被分成自然组，每组有12个观测值。
自定义目标函数需要实现以下两点：
- 预测模型的输出必须是概率性的，并且每个观测值的概率总和必须为1。这也被称为softmax目标函数，实现起来相对简单
- 每个组内每个类别的概率总和必须为1。这在二项分类空间中已被实现，称为条件逻辑模型。
总结来说，对于每个组（在我这里是4个观测值），概率在每列和每行中的总和应为1。我编写了一个略显粗糙的函数来实现这一点，但当我尝试在python的xgb框架中运行我的自定义目标函数时，我得到了以下错误：
TypeError: cannot unpack non-iterable numpy.float64 object
我的完整代码如下：
import lightgbm as lgbimport numpy as npimport pandas as pddef standardiseProbs(preds, groupSize, eta = 0.1, maxIter = 100): # add groupId to preds dataframe n = preds.shape[0] if n % groupSize != 0: print('The selected group size paramter is not compatible with the data') preds['groupId'] = np.repeat(np.arange(0, int(n/groupSize)), groupSize) #initialise variables error = 10000 i = 0 # perform loop while error exceeds set threshold (subject to maxIter) while error > eta and i<maxIter: i += 1 # get sum of probabilities by game byGroup = preds.groupby('groupId')[0, 1, 2, 3].sum().reset_index() byGroup.columns = ['groupId', '0G', '1G', '2G', '3G'] if '3G' in list(preds.columns): preds = preds.drop(['3G', '2G', '1G', '0G'], axis=1) preds = preds.merge(byGroup, how='inner', on='groupId') # adjust probs to be consistent across a game for v in [1, 2, 3]: preds[v] = preds[v] / preds[str(v) + 'G'] preds[0] = (groupSize-3)* (preds[0] / preds['0G']) # sum probabilities by player preds['rowSum'] = preds[3] + preds[2] + preds[1] + preds[0] # adjust probs to be consistent across a player for v in [0, 1, 2, 3]: preds[v] = preds[v] / preds['rowSum'] # get sum of probabilities by game byGroup = preds.groupby('groupId')[0, 1, 2, 3].sum().reset_index() byGroup.columns = ['groupId', '0G', '1G', '2G', '3G'] # calc error errMat = abs(np.subtract(byGroup[['0G', '1G', '2G', '3G']].values, np.array([(groupSize-3), 1, 1, 1]))) error = sum(sum(errMat)) preds = preds[['groupId', 0, 1, 2, 3]] return predsdef condObjective(preds, train): labels = train.get_label() preds = pd.DataFrame(np.reshape(preds, (int(preds.shape[0]/4), 4), order='C'), columns=[0,1,2,3]) n = preds.shape[0] yy = np.zeros((n, 4)) yy[np.arange(n), labels] = 1 preds['matchId'] = np.repeat(np.arange(0, int(n/4)), 4) preds = preds[['matchId', 0, 1, 2, 3]] preds = standardiseProbs(preds, groupSize = 4, eta=0.001, maxIter=500) preds = preds[[0, 1, 2, 3]].values grad = (preds - yy).flatten() hess = (preds * (1. - preds)).flatten() return grad, hessdef mlogloss(preds, train): labels = train.get_label() preds = pd.DataFrame(np.reshape(preds, (int(preds.shape[0]/4), 4), order='C'), columns=[0,1,2,3]) n = preds.shape[0] yy = np.zeros((n, 4)) yy[np.arange(n), labels] = 1 preds['matchId'] = np.repeat(np.arange(0, int(n/4)), 4) preds = preds[['matchId', 0, 1, 2, 3]] preds = standardiseProbs(preds, groupSize = 4, eta=0.001, maxIter=500) preds = preds[[0, 1, 2, 3]].values loss = -(np.sum(yy*np.log(preds)+(1-yy)*np.log(1-preds))/n) return lossn, k = 880, 5xtrain = np.random.rand(n, k)ytrain = np.random.randint(low=0, high=2, size=n)ltrain = lgb.Dataset(xtrain, label=ytrain)xtest = np.random.rand(int(n/2), k)ytest = np.random.randint(low=0, high=2, size=int(n/2))ltest = lgb.Dataset(xtrain, label=ytrain)lgbmParams = {'boosting_type': 'gbdt', 'num_leaves': 250, 'max_depth': 3, 'min_data_in_leaf': 10, 'min_gain_to_split': 0.75, 'learning_rate': 0.01, 'subsample_for_bin': 120100, 'min_child_samples': 70, 'reg_alpha': 1.45, 'reg_lambda': 2.5, 'feature_fraction': 0.45, 'bagging_fraction': 0.55, 'is_unbalance': True, 'objective': 'multiclass', 'num_class': 4, 'metric': 'multi_logloss', 'verbose': 1}lgbmModel = lgb.train(lgbmParams, ltrain, valid_sets=ltest,fobj=condObjective, feval=mlogloss, num_boost_round=5000, early_stopping_rounds=100, verbose_eval=50)假设没有更好的方法来强制我的预测符合我设定的限制条件，我需要做什么才能使自定义目标函数正常工作？
回答：
这个错误的问题
-> 2380 eval_name, val, is_higher_better = feval_ret // this is the return of mlogloss 2381 ret.append((data_name, eval_name, val, is_higher_better)) 2382 return retTypeError: 'numpy.float64' object is not iterable来自函数 mlogloss()。因为你把它用作评估函数 feval=mlogloss，它应该返回三样东西：它的名字、值和一个指示更高值是否更好的布尔值。
def mlogloss(...):...return "my_loss_name", loss_value, False