How to convert a dict to tensors in tensorflow

Question

I am following the tutorial about tensorflow: https://www.tensorflow.org/tutorials/wide

There are lots of categorical features which have to be converted to sparse matrix with tf.feature_column.categorical_column_with_vocabulary_list().

BUT, I do not want to use the predefined Estimator,

m = tf.estimator.LinearClassifier(
    model_dir=model_dir, feature_columns=base_columns + crossed_columns)

I prefer to use a costumed NN model, with:

estimator = tf.contrib.learn.Estimator(model_fn=model)
estimator.fit(input_fn=input_fn(df, num_epochs=100, shuffle=True), \ 
              steps=100)

So in model(), there will be

 def model(features, labels, mode): 
    ...
    node = tf.add(tf.matmul(features, w), b)
    ...

Then, I got the errors like:

 TypeError: Failed to convert object of type <class 'dict'> to Tensor.
 Contents: {'education': <tf.Tensor
 'random_shuffle_queue_DequeueUpTo:1' shape=(?,) dtype=string>, 'age':
 <tf.Tensor 'random_shuffle_queue_DequeueUpTo:2' shape=(?,) dtype=float64> ... 

My question is how to convert the features to a tensor that can be used as input.

Hope I have described the question clear. Thank you in advance.

Solution

features is a dict of Tensor you can get a Tensor by like features['education'], but this Tensor still type of string, it still can't use tf.add(tf.matmul(features, w), b), you should process your string type feature into numberical feature with like tf.feature_column.categorical_column_with_vocabulary_list().

Update:

You can check offical dnn implementation , in def dnn_logit_fn part, it use feature_column_lib.input_layer to generate input layer from features and columns, and the columns is a list of tf.feature_columns.*.

When defining a tf.feature_columns.* such as tf.feature_column.categorical_column_with_vocabulary_list(), it accept string which must exists in features.keys() as first parameter, it connects a tensor from features to a feature_column to tell tf how to process the raw input (string) tensor into a feature tensor(numberical).