Merlin has used incorrectly theano.scan function in gating.py. The bug has been confirmed by my local code.
Theano's scan function will loop through the first dimension. While the batch input has 3 dimensions (num_batches, num_timesteps, num_dimensions). A dimension transposing operation should be applied before scan to make the batch input has a (num_timesteps, num_batches, num_dimensions) shape. After scan function, you should transpose them back.
Merlin has used incorrectly
theano.scanfunction in gating.py. The bug has been confirmed by my local code.Theano's
scanfunction will loop through the first dimension. While the batch input has 3 dimensions(num_batches, num_timesteps, num_dimensions). A dimension transposing operation should be applied beforescanto make the batch input has a(num_timesteps, num_batches, num_dimensions)shape. Afterscanfunction, you should transpose them back.Attention, I only checked gating.py file.