Xc@`sydZddlmZmZmZddlZddlmZddlm Z ddl m Z ddl m Z mZddlmZdd lmZejd Zeeje jgd Zeeje jgd Zeeje jgd Zeeje jgdZeeje gdZeeje gdZdS(s This file implement specialization optimization that break the canonization form of the graph. Currently there is problem with the order of optimization and the definition of definition of canonized graph. Right now there is a canonization optimization phase that try to make all equivalent graph identical. This is not always the case, but it do many of the basic stuff canonical. We need to extend the definition of canonization to make this true more often. The problem this file indent to fix in the future is that in the "Equilibrium" specialization optimization phase, there is optimization that request that the graph is canonical, some other request that this is not true, and some other that break the canonicalization for some optimization. As we can't control the order of those optimization, there is case that some optimization requesting a canonical graph won't be applied as optimization that break the canonicalization form of the graph executed before. To fix this, we need to split the specialization phase into a phase where optimization can't break the canonicalization form and one where this is allowed. This is also needed for the stabilized optimization phase, but as it happen before the specialization phase, this cause less problem. Also, we should make the fgraph refuse optimization that break the canonization of the graph in the optimizations phases where the graph is supposed to be canonical. i(tabsolute_importtprint_functiontdivisionN(tgof(tCAReduce(tbasic(t DimShufflet Subtensor(tregister_uncanonicalize(tscalarstheano.tensor.optcC`st|jtjr|jj|}t|jdjdkrltt j ||j d}|dgSt|jdjdkrdtj |j d|gSndS(s> If we don't use the argmax, change it to a max only. iiN(t isinstancetoptTt MaxAndArgmaxt get_paramstlentoutputstclientsRtscaltmaximumtinputstNonet_argmax(tnodetaxistnew((s@/tmp/pip-build-X4mzal/theano/theano/tensor/opt_uncanonicalize.pytlocal_max_and_argmax2s cC`s|jtjkr|jdjr|jd}|jrt|jjtr|jjjtj kr|jjd}|jr|jjtjkrttj |jjj |jjdgSqnt S(s Change -(max(-x)) to min. This is tested in tensor/tests/test_basic.py:test_min_max. Notes ----- We don't need an opt that will do the reverse as by default the interface put only MaxAndArgmax into the graph. i( R R tnegRtownerR Rt scalar_opRRtminimumRtFalse(RtmaxR((s@/tmp/pip-build-X4mzal/theano/theano/tensor/opt_uncanonicalize.pytlocal_max_to_minBs"   )cC`st|jtjr|jd}|jrt|jjtr|jjj}d|j|jjdjt t |jjdj}||krt Stj |jjd|jdgSnt S(s If a dimshuffle is inside an alloc and only adds dimension to the left, remove it. Alloc(DimShuffle(x), ...) - > Alloc(x, ...) itxi(R"( R R R tAllocRRRt new_ordertndimttupletrangeRtalloc(Rtinput_R$texpected_new_order((s@/tmp/pip-build-X4mzal/theano/theano/tensor/opt_uncanonicalize.pytlocal_alloc_dimshuffle]s    'cC`st|jtjr|jd}|jrt|jjtr|jjj}d}x:|D]2}|dkrtq\q\||krtS|d7}q\Wtj |jjd|jdd|j dj gSntS(s If a dimshuffle is inside a reshape and does not change the order of dimensions, remove it. Reshape(Dimshuffle(x), shp) -> Reshape(x, shp) iR"iR%( R R R tReshapeRRRR$RtreshapeRR%(RR)R$toffsettdim((s@/tmp/pip-build-X4mzal/theano/theano/tensor/opt_uncanonicalize.pytlocal_reshape_dimshuffless      cC`st|jtr|jdjr|jd}t|jjtjr|jj}dt||j t t |j }||krt St||j }d|t |jjd}tj |jjd|gSnt S(s If an alloc is inside a dimshuffle which only adds dimension to the left, scrap the dimshuffle and adds 1 into the alloc dimshuffle{x, 0, 1}(alloc([3 4], 3, 2) => alloc([3 4], 1, 3, 2) iR"i(R"(i(R R RRRR R#R$RR%R&R'RR((RR)R$R*t nb_new_dimstnew_shape_input((s@/tmp/pip-build-X4mzal/theano/theano/tensor/opt_uncanonicalize.pytlocal_dimshuffle_allocs "    cC`st|jtr|jdjrd|jjkr8tS|jj}t|dkr|d}x+|dD]}||kstS|}qkWn|jd}t|jjtr|j }t t |j }x|D]}|j |qWtg|D]}||^qstSt |jjj}|jjdg} tjd} dddg} d} d} d}x|jjjD]}t|tr#| }| d7} xK| D]C}t||dk r| |jjd| g7} | d7} qqW|| krR| |krR| || <| | g7} qRq{| |jjd| g7} | d7} |d7}q{Wxt t|jjj| dj D]_}|||kr|x)t||kr|jtdqW|j| | j| q|q|Wt|| gSntS( sIf a subtensor is inside a dimshuffle which only drop broadcastable dimensions, scrap the dimshuffle and index the subtensor with 0 x[i:j, :, k:l].dimshuffle(0, 2) => x[i:j, 0, k:l] if x.broadcastable == (False, True, False) iR"itstarttstoptstepiN(R R RRRR$RRRt broadcastabletlistR'R%tremovetalltidx_listR tconstanttslicetgetattrRtappend(RR$tpast_dimR/R)R7t missing_dimstit new_idx_listt new_inputstzerotslice_attr_listtjtslice_itsubtensor_removed_dimstidxtpast_jt slice_attr((s@/tmp/pip-build-X4mzal/theano/theano/tensor/opt_uncanonicalize.pytlocal_dimshuffle_subtensors` "       #     ( t__doc__t __future__RRRtloggingttheanoRttheano.tensor.elemwiseRt theano.tensorRR RRttheano.tensor.optRR Rt getLoggert_loggertlocal_optimizerR RRR!R#R+R,R0R3RM(((s@/tmp/pip-build-X4mzal/theano/theano/tensor/opt_uncanonicalize.pyt s* !!!!