NeuralTuringMachine
NeuralTuringMachine copied to clipboard
Problem on associative recall
Using ntm as the default setting works fine on copy task, but it doesn't work on associative recall task.
I run for more than 20000 training steps with 5 hours and 1 GPU, and the sequence error doesn't drop.
Does associative recall task needs another setting?
What are you setting max_seq_len
and num_bits_per_vector
to? To replicate the results in the paper max_seq_len=6
(that's 6 items so 18 vectors) and num_bits_per_vector=6
I will try the above configuration.
By the way, I run the previous configuration for the whole weekends, and the training process reach 350000 training step, and the 'nan' loss occur. I have stopped the previous configuration.