Yiming Wang

@vijayaditya I made a PR to your repo to add an LSTM's xconfig layer (w/o projection)

I compared the 2 fsas with `k2.is_rand_equivalent()` and it is False. H_inv is obtained by ```python hmm_vec = k2.create_fsa_vec(hmms) hmm_vec.labels = torch.where(hmm_vec.labels >= 0, hmm_vec.labels + 1, hmm_vec.labels) # temporarily...

> Not sure if I understand your problems correctly, but `k2.invert` will remove those epsilon labels when output them as `dest.aux_labels`, as here you invert two times (in the first...

I was assuming, during intersect and connect, all named tensor attributes (e.g. temp_labels, aux_labels) in a src Fsa will be mapped correctly the dest Fsa, so that `HL.temp_labels` has changed...

> As your L here is a very simple format > > ``` > FREETEXT freetext > HiXiaowen hixiaowen > NihaoWenwen nihaowenwen > SIL > ``` > > I think...

> mm, see my reply here [#550 (comment)](https://github.com/k2-fsa/k2/issues/550#issuecomment-751310481) Shouldn't temp_labels be changed according to arc_map during intersect and connect, just like aux_labels? I mean, I suppose https://github.com/k2-fsa/k2/blob/master/k2/python/k2/fsa_algo.py#L125 is already handling...

> HL = k2.invert(HL) I see. Yes, it works

BTW, I realized `remove_epsilons_iterative_tropical` is in tropical semiring, so after applying this to H for example, sumexp(scores) for all outgoing scores from a specific state will no longer be 1.0,...

> Yiming: I doubt we rely on the stochasticity. In particular I'm pretty sure we don't rely in an important way on any difference that could be resolved by pushing....

@danpovey in LF-MMI, I did ```python num_graph.scores = num_graph.scores.new_zeros(num_graph.scores.size()) num_graph = k2.connect(k2.intersect(num_graph, den_graph, treat_epsilons_specially=False)) ``` where I make sure `num_graph` and `den_graph` both have no label 0, but still tot_scores...