Using embedding_rnn_seq2seq in tensorflow

Question

The documentation shown in the official documentation seems to be jumping straight into attention models without showing how to use the basic seq2seq model. I'm attempting to translate from a certain date format into one standard method. Some examples are shown below:

[('7 07 13', '2013-07-07'),
 ('30 JULY 1977', '1977-07-30'),
 ('Tuesday, September 14, 1971', '1971-09-14'),
 ('18 09 88', '1988-09-18'),
 ('31, Aug 1986', '1986-08-31')]

where the second column is the output 'y'.

Here is how I plan on using a seq2seq model at a high level:

1. Embed both the output and input characters into 10 dimensional vectors.
2. Pad the input vectors so that they are all a fixed length (29 in my case, and the output is of length 11).
3. Pass those 29 characters into seq2seq so that they output 11 logits of size hidden_size
4. Use the average crossentropy loss (across time steps and batch size) to get a loss value.
5. Optimise this loss.

The model that I have so far is as follows:

import tensorflow.contrib.legacy_seq2seq as seq2seq

enc_inp = [tf.placeholder(tf.int32, shape=(None,)) for t in range(x_seq_length)]
labels = [tf.placeholder(tf.int32, shape=(None,)) for t in range(y_seq_length)]
is_train  = tf.placeholder(tf.bool)
weights = [tf.ones_like(labels_t, dtype=tf.float32) for labels_t in labels]

memory_dim = 32
embed_dim = 32
cell = tf.contrib.rnn.BasicLSTMCell(memory_dim)
dec_outputs, dec_memory = seq2seq.embedding_rnn_seq2seq(enc_inp, dec_inp, cell, 
                                                        len(char2numX), len(char2numY), 
                                                        embed_dim,
                                                        feed_previous = tf.logical_not(is_train))

loss = seq2seq.sequence_loss(dec_outputs, labels, weights)

Hoever, it is complaining that Lengths of logits, weights, and targets must be the same 29, 11, 11, because (I think) dec_outputs is of length 29, whereas I was hoping it would somehow be of length 11.

1. My question is given that I am translating from a sequence of length 29 to a sequence of length 11, how am I supposed to do this in tensorflow.
2. Also correct me if Im wrong, the input to these models is of size [time_steps, batch_size]? Notice that batch_size is the second argument, not first. Got this impression by a few tutorials I read.

Full code with data is available here.