python - tensorflow gather across multiple dimentions -

gather(params, indices) following

output[i, ..., j, :, ... :] = params[indices[i, ..., j], :, ..., :] 

so if have 4-dimensional params , 2-dimensional indices, end having 5-dimensional array result

the question how do

output[i, ..., j, :, ... :] = params[indices[i, :], ..., indices[j, :], :, ..., :] 

so acts numpy's

output = params[indices[0], indices[1], .. , :]

(the #206 ticket on github regarding different issue: numpy-like api, not gathering in general)

one possible way use gather_nd, (as far understand) if want gather_nd on not dimensions, still have create indices them, e.g. if have 10-dimensional array , want index first 2 dimensions 2-dimensional array b, a[b[0], b[1], :] our indices matrix have have 11 columns (with 8 redundant).

--- old indices ----       new index 0 0 <all rows of length 8> 0 1 1 <all rows of length 8> 1 ... 

there's update on #206 @ebrevdo working on generalizing slicing.

meanwhile, flatten array, construct linear indices elements want, use gather, reshape back, done in another answer mrry. that's not worse in efficiency native implementation