ValueError: Feature not in features dictionary

问题

I am attempting to write a simple deep machine learning model using TensorFlow. I'm using a toy dataset I made up in Excel just to get the model working and accepting data. My code is as follows:

import pandas as pd
import numpy as np
import tensorflow as tf

raw_data = np.genfromtxt('ai/mock-data.csv', delimiter=',', dtype=str)
my_data = np.delete(raw_data, (0), axis=0) #deletes the first row, axis=0 indicates row, axis=1 indicates column
my_data = np.delete(my_data, (0), axis=1) #deletes the first column

policy_state = tf.feature_column.categorical_column_with_vocabulary_list('policy_state', [
    'AL', 'CA', 'MI'
])

modern_classic_ind = tf.feature_column.categorical_column_with_vocabulary_list('modern_classic_ind', [
    '0', '1'
])

h_plus_ind = tf.feature_column.categorical_column_with_vocabulary_list('h_plus_ind', [
    '0', '1'
])

retention_ind = tf.feature_column.categorical_column_with_vocabulary_list('retention_ind', [
    '0', '1'
])

feature_columns = [
    tf.feature_column.indicator_column(policy_state),
    tf.feature_column.indicator_column(modern_classic_ind),
    tf.feature_column.indicator_column(h_plus_ind)
]
classifier = tf.estimator.DNNClassifier(feature_columns=feature_columns,
                                      hidden_units=[10, 20, 10],
                                      n_classes=3,
                                      model_dir="/tmp/ret_model")

train_input_fn = tf.estimator.inputs.numpy_input_fn(
x={"x": np.array(my_data[:, 0:3], dtype=str)},
y=np.array(np.array(my_data[:, 3], dtype=str)),
num_epochs=None,
shuffle=True)

classifier.train(input_fn=train_input_fn, steps=2000)

Unfortunately, I am getting the following error. I have tried trimming the labels off the csv file versus leaving them, naming the feature columns different things, and changing the type of the numpy array. The error persists.

ValueError: Feature h_plus_ind is not in features dictionary.

If I remove h_plus_ind, it simply throws the error on a different column.

回答1:

When using tf.feature_columns, the data you feed in your input_fn should have the same keys as the feature columns previously created. So, the x of your train_input_fn should be a dictionary, with keys named after the feature_columns.

A mock example :

x = {"policy_state": np.array(['AL','AL','AL','AL','AL']),
     "modern_classic_ind": np.array(['0','0','0','0','0']),
     "h_plus_ind": np.array(['0','0','0','0','0']),}

On the side :

This great article from the developers google blog could be a great read, as it introduces a new way to create input_fn directly from a csv file with the tf.Dataset API. It has a better memory management, and avoid loading all the dataset into memory.

回答2:

I have the same problem but when i checked the names of my columns of database, there was a little mistake in the name of column. Check out your column's names.

回答3:

I have faced the same problem.In my case, the target variable was also fed to the features dictionary.I removed it from features dictionary and it worked.

回答4:

If you've reached this page because of TF serving, another possibility is that the keys in the dictionary passed as serving_input_fn do not correspond to the ones in your model, just double-check the dict.

来源：https://stackoverflow.com/questions/47635507/valueerror-feature-not-in-features-dictionary

标签

python

numpy

dictionary

tensorflow