Chinese recognition speed is too slow

[rchglev]

I applied this repository for recognition of Chinese.
When I recognize English text, its speed is fast.
But after making Chinese model, its speed decreased 100X.
I surprised.
Hanzi Number is about 6000 characters.

[rchglev]

Please help me

[Li-Ming-Fan]

hi, @rchglev

I don't know how you implemented your Chinese model, so I do not really know where the crux is.

In the beam_search_decoder function, I did not specify beam_width and used the default value (100),
as you can see from the function prototype:

tf.nn.ctc_beam_search_decoder(
inputs,
sequence_length,
beam_width=100,
top_paths=1,
merge_repeated=True
)

The beam_width value may have some impact on the speed.

RNN structure is notorious for its slow speed. If the input sequence is very long, then the calculation will be time-consuming.