Tensorflow One Hot Encoder?
Does tensorflow have something similar to scikit learn's one hot encoder for processing categorical data? Would using a placeholder of tf.string behave as categorical data?
I realize I can manually pre-process the data before sending it to tensorflow, but having it built in is very convenient.
Answer
As of TensorFlow 0.8, there is now a native one-hot op, tf.one_hot that can convert a set of sparse labels to a dense one-hot representation. This is in addition to tf.nn.sparse_softmax_cross_entropy_with_logits, which can in some cases let you compute the cross entropy directly on the sparse labels instead of converting them to one-hot.
Previous answer, in case you want to do it the old way: @Salvador's answer is correct - there (used to be) no native op to do it. Instead of doing it in numpy, though, you can do it natively in tensorflow using the sparse-to-dense operators:
num_labels = 10
# label_batch is a tensor of numeric labels to process
# 0 <= label < num_labels
sparse_labels = tf.reshape(label_batch, [-1, 1])
derived_size = tf.shape(label_batch)[0]
indices = tf.reshape(tf.range(0, derived_size, 1), [-1, 1])
concated = tf.concat(1, [indices, sparse_labels])
outshape = tf.pack([derived_size, num_labels])
labels = tf.sparse_to_dense(concated, outshape, 1.0, 0.0)
The output, labels, is a one-hot matrix of batch_size x num_labels.
Note also that as of 2016-02-12 (which I assume will eventually be part of a 0.7 release), TensorFlow also has the tf.nn.sparse_softmax_cross_entropy_with_logits op, which in some cases can let you do training without needing to convert to a one-hot encoding.
Edited to add: At the end, you may need to explicitly set the shape of labels. The shape inference doesn't recognize the size of the num_labels component. If you don't need a dynamic batch size with derived_size, this can be simplified.
Edited 2016-02-12 to change the assignment of outshape per comment below.