Can we merge a batch of text into one decoding graph?

[yaguanghu]

Hello, I'm learning using k2 to implement CTC.
In the "create_decoding_graph" function, it seems like merging a batch of label text into one decoding graph, is it reasonable?

snowfall/egs/librispeech/asr/simple_v1/train.py

Line 34 in 26398e8

def create_decoding_graph(texts, L, symbols):

def create_decoding_graph(texts, L, symbols):
    word_ids_list = []
    for text in texts:
        filter_text = [
            i if i in symbols._sym2id else '<UNK>' for i in text.split(' ')
        ]
        word_ids = [symbols.get(i) for i in filter_text]
        word_ids_list.append(word_ids)

And, what's the meaning of the following line, for ctc blank self loop?

snowfall/egs/librispeech/asr/simple_v1/train.py

Line 44 in 26398e8

decoding_graph = k2.add_epsilon_self_loops(decoding_graph)

[qindazhu]

Sure, definitely we need to do this, we also plan to cache those static decoding graph at the beginning of training.

[yaguanghu]

In my opinion, we should create an independent label graph for each utterance.That means we can get a list of decoding graph at the begining of training, or during training.
But according to the above code, it seems like merging those into one graph. Is my understanding right?

[danpovey]

The 'fsa' object actually supports an array of Fsa's. They are not merged, they are separate but stored in one data structure.

…

On Fri, Nov 20, 2020 at 2:14 PM yaguang ***@***.***> wrote: In my opinion, we should create an independent label graph for each utterance.That means we can get a list of decoding graph at the begining of training, or during training. But according to the above code, it seems like merging those into one graph. Is my understanding right? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#24 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLOYBEA3NTETJQBEA5FTSQYCMJANCNFSM4T4KRT3Q> .

[yaguanghu]

I see.
So what's the meaning of "add_epsilon_self_loops"?

snowfall/egs/librispeech/asr/simple_v1/train.py

Line 44 in 26398e8

decoding_graph = k2.add_epsilon_self_loops(decoding_graph)

[qindazhu]

We add self-loops with symbol zero (epsilon) here so it can match blank in CTC training. You may want to check the code intersect_dense_pruned below, we do CTC alignment with FSA intersection. Noted we view network output as an Fsa (dense_fsa_vec) as well

[yaguanghu]

How to deal with repeated characters?

[danpovey]

Repeated characters not allowed in the successful path (currently, anyway)

…

On Fri, Nov 20, 2020 at 3:02 PM yaguang ***@***.***> wrote: How to deal with repeated characters? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#24 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLOYWD3NXWTI3ZYOXPL3SQYH7TANCNFSM4T4KRT3Q> .

[yaguanghu]

I see.
Maybe i can try to implement a standard CTC with k2, is it worth doing?

[danpovey]

Sure! Would just require changes in the FSA topology.

…

On Fri, Nov 20, 2020 at 3:12 PM yaguang ***@***.***> wrote: I see. Maybe i can try to implement a standard CTC with k2, is it worth doing? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#24 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLO5OWFI46MU6AIXOV43SQYJHDANCNFSM4T4KRT3Q> .