def _pop_op(everything, accum, everything_max = None, everything_min = None, word = None, aword = None, one_step=False, use_noise=True): rval = proj_h[0](accum[0], one_step=one_step, use_noise=use_noise) for si in xrange(1,state['decoder_stack']): rval += proj_h[si](accum[si], one_step=one_step, use_noise=use_noise) if state['mult_out']: rval = rval * everything else: rval = rval + everything if aword and state['avg_word']: wcode = aword if one_step: if state['mult_out']: rval = rval * wcode else: rval = rval + wcode else: if not isinstance(wcode, TT.TensorVariable): wcode = wcode.out shape = wcode.shape rshape = rval.shape rval = rval.reshape([rshape[0]/shape[0], shape[0], rshape[1]]) wcode = wcode.dimshuffle('x', 0, 1) if state['mult_out']: rval = rval * wcode else: rval = rval + wcode rval = rval.reshape(rshape) if word and state['bigram']: if one_step: if state['mult_out']: rval *= proj_word(emb_t(word, use_noise=use_noise), one_step=one_step, use_noise=use_noise) else: rval += proj_word(emb_t(word, use_noise=use_noise), one_step=one_step, use_noise=use_noise) else: if isinstance(word, TT.TensorVariable): shape = word.shape ndim = word.ndim else: shape = word.shape ndim = word.out.ndim pword = proj_word(emb_t(word, use_noise=use_noise), one_step=one_step, use_noise=use_noise) shape_pword = pword.shape if ndim == 1: pword = Shift()(pword.reshape([shape[0], 1, outdim])) else: pword = Shift()(pword.reshape([shape[0], shape[1], outdim])) if state['mult_out']: rval *= pword.reshape(shape_pword) else: rval += pword.reshape(shape_pword) if state['deep_out']: rval = drop_layer(act_layer(rval), use_noise=use_noise) return rval
def _pop_op(everything, accum, everything_max=None, everything_min=None, word=None, aword=None, one_step=False, use_noise=True): rval = proj_h[0](accum[0], one_step=one_step, use_noise=use_noise) for si in xrange(1, state['decoder_stack']): rval += proj_h[si](accum[si], one_step=one_step, use_noise=use_noise) if state['mult_out']: rval = rval * everything else: rval = rval + everything if aword and state['avg_word']: wcode = aword if one_step: if state['mult_out']: rval = rval * wcode else: rval = rval + wcode else: if not isinstance(wcode, TT.TensorVariable): wcode = wcode.out shape = wcode.shape rshape = rval.shape rval = rval.reshape( [rshape[0] / shape[0], shape[0], rshape[1]]) wcode = wcode.dimshuffle('x', 0, 1) if state['mult_out']: rval = rval * wcode else: rval = rval + wcode rval = rval.reshape(rshape) if word and state['bigram']: if one_step: if state['mult_out']: rval *= proj_word(emb_t(word, use_noise=use_noise), one_step=one_step, use_noise=use_noise) else: rval += proj_word(emb_t(word, use_noise=use_noise), one_step=one_step, use_noise=use_noise) else: if isinstance(word, TT.TensorVariable): shape = word.shape ndim = word.ndim else: shape = word.shape ndim = word.out.ndim pword = proj_word(emb_t(word, use_noise=use_noise), one_step=one_step, use_noise=use_noise) shape_pword = pword.shape if ndim == 1: pword = Shift()(pword.reshape([shape[0], 1, outdim])) else: pword = Shift()(pword.reshape([shape[0], shape[1], outdim])) if state['mult_out']: rval *= pword.reshape(shape_pword) else: rval += pword.reshape(shape_pword) if state['deep_out']: rval = drop_layer(act_layer(rval), use_noise=use_noise) return rval