shtoshni / g2p Goto Github PK

Excuse me！
I cannot find any operation for the UNK word.

Can you tell me the process method for UNK?

Thank you very much!

Best Regards!

Although CMUDict setup doesn't raise an exception, I tried it with other dataset and I believe there is a bug in seq2seq_model.py get_batch(self, data, bucket_id=None) method. Specifically I believe there is a case when the decoder_pad_size becomes < 0 when self.isTraining is false at the following code:

decoder_pad_size = max_len_target - (len(decoder_input) + 1)

When decoder_pad_size is negative the following error is raised:

decoder_inputs = np.asarray(decoder_inputs, dtype=np.int32).T
File "/home/XXXX/.local/share/virtualenvs/G2P-MASTER-mSKoJG47/lib/python2.7/site-packages/numpy/core/numeric.py", line 538, in asarray
    return array(a, dtype, copy=False, order=order)
ValueError: setting an array element with a sequence.

I believe the following is the culprit. It should be with + 1

seq_len_target[i] = decoder_size        # Original Code
seq_len_target[i] = decoder_size + 1    # Fixed Code

For your information, here are the conditions when ValueError occurs:

• decoder_size = 35
• max_len_target = 35
• self.isTraining = false
• len(decoder_input) = 35
• decoder_pad_size = -1

With above condition, the original code creates decoder_inputs with the shape of 256 [FLAGS.batch_size] x 36,
when it should be 256 [FLAGS.batch_size] x 37 ([data_utils.GO_ID] + decoder_input + [data_utils.EOS_ID] + [data_utils.PAD_ID]
* 0)

As an extra information, my G2P dataset is not English and has a Maximum input sequence length of 65 and Maximum output sequence length of 97. While the above fix of + 1 seem to do the trick (no more ValueError), should I be concerned with other parameters (e.g. _buckets = [(35, 35)] in data_utils.py? I read your comment regarding the bucket but the link you mention is broken: http://goo.gl/d8ybpl).

Buckets are useful to limit the amount of padding required since we use
minibatch processing. For more detail refer this: http://goo.gl/d8ybpl
Bucket sizes are specified as a list with each entry of the form -
(Max input sequence length, Max output sequence length)
Since this project is about Grapheme-to-Phoneme conversion, where the input
sequence is characters in a word and output sequence is phonemes in word
pronunciation, we use a single bucket to merely denote the max word length
and max pronunciation length

There is a link in the recent paper titled "Jointly Learning to Align and Convert Graphemes to Phonemes with Neural Attention Models" that connects to this repository, but it seems that the code was not committed.

Hi!

I'm trying to replicate the results from your paper for the CMUDict dataset, but I can't get it above 44% WER 12% PER.
You mention the cmusphinx version, but did you use this dataset unprocessed? In the paper it seems like the stress suffices are removed from the vowels, therefore making the task much easier.

Thanks!

Hey, I can't seem to find the splits you used to train on the CMU dictionary. Wondering where I might find them/if you can provide them.

Excuse me!
When will you release your code about "attention mechanism"?

Thank you very much!