python - 安装TensorForestEstimator时TensorFlow崩溃

我正在尝试使用表示7个特征和7个标签的数字浮点数据来适合TensorForestEstimator模型。即，features和labels的形状均为(484876, 7)。我在num_classes=7中适当设置了num_features=7和ForestHParams。数据格式如下:

f1       f2     f3    f4      f5    f6    f7   l1       l2       l3       l4       l5       l6       l7
39000.0  120.0  65.0  1000.0  25.0  0.69  3.94 39000.0  39959.0  42099.0  46153.0  49969.0  54127.0  55911.0
32000.0  185.0  65.0  1000.0  75.0  0.46  2.19 32000.0  37813.0  43074.0  48528.0  54273.0  60885.0  63810.0
30000.0  185.0  65.0  1000.0  25.0  0.41  1.80 30000.0  32481.0  35409.0  39145.0  42750.0  46678.0  48595.0

调用fit()时，Python崩溃，并显示以下消息:



这是启用tf.logging.set_verbosity('INFO')时的输出:

INFO:tensorflow:training graph for tree: 0
INFO:tensorflow:training graph for tree: 1
...
INFO:tensorflow:training graph for tree: 9998
INFO:tensorflow:training graph for tree: 9999
INFO:tensorflow:Create CheckpointSaverHook.
2017-07-26 10:25:30.908894: F tensorflow/contrib/tensor_forest/kernels/count_extremely_random_stats_op.cc:404]
Check failed: column < num_classes_ (39001 vs. 8)

Process finished with exit code 134 (interrupted by signal 6: SIGABRT)

我不确定这个错误是什么意思，因为num_classes=7而不是8，并且由于功能和标签的形状是(484876, 7)，所以它真的没有意义，我不知道39001的来源。

这是要重现的代码:

import numpy as np
import pandas as pd
import os

def get_training_data():
    training_file = "data.txt"
    data = pd.read_csv(training_file, sep='\t')

    X = np.array(data.drop('Result', axis=1), dtype=np.float32)

    y = []
    for e in data.ResultStr:
        y.append(list(np.array(str(e).replace('[', '').replace(']', '').split(','))))

    y = np.array(y, dtype=np.float32)

    features = tf.constant(X)
    labels = tf.constant(y)

    return features, labels

hyperparameters = ForestHParams(
    num_trees=100,
    max_nodes=10000,
    bagging_fraction=1.0,
    num_splits_to_consider=0,
    feature_bagging_fraction=1.0,
    max_fertile_nodes=0,
    split_after_samples=250,
    min_split_samples=5,
    valid_leaf_threshold=1,
    dominate_method='bootstrap',
    dominate_fraction=0.99,
    # All parameters above are default
    num_classes=7,
    num_features=7
)

estimator = TensorForestEstimator(
    params=hyperparameters,
    # All parameters below are default
    device_assigner=None,
    model_dir=None,
    graph_builder_class=RandomForestGraphs,
    config=None,
    weights_name=None,
    keys_name=None,
    feature_engineering_fn=None,
    early_stopping_rounds=100,
    num_trainers=1,
    trainer_id=0,
    report_feature_importances=False,
    local_eval=False
)

estimator.fit(
    input_fn=lambda: get_training_data(),
    max_steps=100,
    monitors=[
        TensorForestLossHook(
            early_stopping_rounds=30
        )
    ]
)

如果我用SKCompat包装它，它也不起作用，会发生相同的错误。崩溃的原因是什么？

最佳答案

需要在regression=True中指定ForestHParams，因为默认情况下TensorForestEstimator假定它用于解决分类问题，该分类问题只能输出一个值。

在估计器初始化时会创建一个隐式的num_outputs变量，如果未指定1，则将其设置为regression。如果指定了regression，则通常会保存num_outputs = num_classes和检查点。