CODEWITHSUNDEEP
tensorflow
Tomakko
Tomakko

Reputation: 33

Can i make a custom tf.estimator without using feature columns?

i.e. can i directly feed tensors with numerical data into the model_fn of a custom estimator?

Creating a numerical column and do a mapping from keys to values seems to be an overkill for me, who only works on image data.

Upvotes: 0

Views: 236

Answers (1)

xdurch0
xdurch0

Reputation: 10474

Yes. The features argument to your model_fn can simply be a tensor, or a dict mapping strings to tensors in the case of multiple inputs. This also means that your input_fn can simply return such objects.

Upvotes: 2

Related Questions