Tensorflow SKCompat converting float32 values in P

Question

I am using the following Pandas dataframe as the training input for an SKCompat estimator:

>>> training_data.info()
<class 'pandas.core.frame.DataFrame'>
Int64Index: 8709 entries, 4396 to 1889
Data columns (total 8 columns):
season        8709 non-null int64
holiday       8709 non-null int64
workingday    8709 non-null int64
weather       8709 non-null int64
temp          8709 non-null float32
atemp         8709 non-null float32
humidity      8709 non-null int64
windspeed     8709 non-null float32

At some point in the tensorflow code it passes the dataframe through the function:

tensorflow.contrib.learn.python.learn.learn_io.pandas_io.extract_pandas_data.

This seems to lose the dtype information and go back to float64

>>> x_training = extract_pandas_data(x_training)
>>> x_training.dtype
{dtype} float64

further on I then get the following exception, as the floats have been converted to float64:

TypeError: Input 'input_data' of 'TreePredictions' Op has type float64 that does not match expected type of float32.

I have seen a few examples of people using tf.cast to get around this issue, but I don't understand how to apply for my use case. What do I need to do to this Pandas DataFrame to make it work with the TensorForestEstimator?

Many thanks,

Mark

---

Code example, with "tf.cast" fix:

def stackoverflow_example(x_training: pd.DataFrame, y_training: pd.DataFrame):

    params = tensor_forest.ForestHParams(
        num_classes=1, num_features=5,
        num_trees=10, max_nodes=1000)

    graph_builder_class = tensor_forest.TrainingLossForest

    est = estimator.SKCompat(random_forest.TensorForestEstimator(
        params, graph_builder_class=graph_builder_class))

    x_training = tf.cast(x_training.drop('datetime', 1), tf.float32)

    est.fit(x_training, y_training, batch_size=1000)

this code returns the following exception with the cast:

ValueError: Inputs cannot be tensors. Please provide input_fn.