Reputation: 1318
I tried to transform the inputs of official tutorial about seq2seq model from placeholder to tfrecords but failed mainly because if I write all samples to a single tfrecord file I can't group samples with similar length into a same bucket, do I need to start four queues(the length of buckets) to realize it? any advice and solution is welcomed
Upvotes: 1
Views: 335
Reputation: 126184
The newly added tf.contrib.training.bucket_by_sequence_length()
function adds support for creating multiple queues to group by bucketed sequence length, and combining the full batches in a single queue.
Upvotes: 1