Noticing Slow testing steps in --utf8 mode

Archives DeepSpeech
, , ,
#1

Hi,
I am training the model with common voice Arabic data - 12Hrs.

With following parameters:

!python3 DeepSpeech.py \
--train_files data/ar/clips/train.csv \
--dev_files data/ar/clips/dev.csv \
--test_files data/ar/clips/test.csv \
--train_batch_size 1 \
--dev_batch_size 1 \
--test_batch_size 1 \
--summary_dir logs/ \
--export_dir model_1024/ \
--checkpoint_dir checkpoint_dir \
--epochs 20 \
--utf8 \
--learning_rate 0.0001 \
--dropout_rate 0.00 \
--n_hidden 1024 \
--scorer ' '

I am noticing very slow progress with steps while testing.
Please advise, I have tested with normal mode (without --utf8). And seems fast while testing.

#2

Use higher batch sizes for training if your CPU/GPU can do that: 4 or 8

Use dropout of 0.25 or higher. Very important if nothing is moving.

And 12 hrs is very few data, don’t expect much

#3

UTF-8 increases the alphabet size to 256, and since you’re not passing a scorer, it also can’t trim out of vocabulary beams during decoding, so the decoding process gets slower. If you build a scorer it should counteract any slowness you’re seeing.

#4

The out of vocabulary trimming thing is also made worse by UTF-8 because without a scorer the decoder does not trim invalid UTF-8 sequences, so it’s spending time doing useless work.

#5

So even with scorer the utf-8 mode will be slower than normal mode?

#6

No. There’s no universal rule as it depends on what you’re comparing it to. Alphabet mode can be super slow with very large alphabets, for example. With a scorer it should be just as fast in most cases, and much faster if you have a large alphabet.

In general, unless you have a specific issue with alphabet mode that can be fixed by using UTF-8 mode, you should stick to using an alphabet. Some examples of these issues:

  1. My alphabet is too large and makes the model big and slow.
  2. I want to easily train on several languages at once.
  3. I want to easily transfer from one language to another including the final layer.
1 Like
#7

@reuben Thank you for these suggestion. :slight_smile: