Add eval and parallel dataset #4651
Conversation
|
|
||
| def _preprocess_transcript(transcript, token_to_index): | ||
| """Process transcript as label features.""" | ||
| return featurizer.compute_label_feature(transcript, token_to_index) |
There was a problem hiding this comment.
Usually we should avoid the one line wrap function with same parameter.
| def compute_label_feature(text, token_to_idx): | ||
| """Convert string to a list of integers.""" | ||
| tokens = list(text.strip().lower()) | ||
| feats = [token_to_idx[token] for token in tokens] |
| Arguments: | ||
| labels (string): mapping from integers to characters. | ||
| blank_index (int, optional): index for the blank '_' character. |
There was a problem hiding this comment.
The default value is different from the one in comments. Also why call underscore with blank?
| labels (string): mapping from integers to characters. | ||
| blank_index (int, optional): index for the blank '_' character. | ||
| Defaults to 0. | ||
| space_index (int, optional): index for the space ' ' character. |
| space_index (int, optional): index for the space ' ' character. | ||
| Defaults to 28. | ||
| """ | ||
| # e.g. labels = "[a-z]' _" |
There was a problem hiding this comment.
Do u need a param for the single quote index?
| for i, char in enumerate(sequence): | ||
| if char != self.int_to_char[self.blank_index]: | ||
| # if this char is a repetition and remove_repetitions=true, | ||
| # skip. |
There was a problem hiding this comment.
should wrap this into the previous line.
| space. Option to remove repetitions (e.g. 'abbca' -> 'abca'). | ||
| Arguments: | ||
| sequences: list of 1-d array of integers |
|
|
||
| def process_string(self, remove_repetitions, sequence): | ||
| """Process each given sequence.""" | ||
| seq_string = '' |
There was a problem hiding this comment.
In case your input string is very long, you should use list to hold individual char, and join them at the end with "". Python string is immutable object. concat string, means allocate new string and delete the old one.
| target_strings = self.process_strings( | ||
| target_strings, remove_repetitions=True) | ||
| wer = 0 | ||
| for i in xrange(len(decoded_strings)): |
There was a problem hiding this comment.
How about just:
for (decoded_string, target_string) in zip(decoded_strings, target_strings):
wer += self.wer(decoded_string, target_string) / float(len(target_string.split()))
| target_strings = self.process_strings( | ||
| target_strings, remove_repetitions=True) | ||
| cer = 0 | ||
| for i in xrange(len(decoded_strings)): |
| strings = [] | ||
| for x in xrange(len(sequences)): | ||
| seq_len = sizes[x] if sizes is not None else len(sequences[x]) | ||
| string = self._convert_to_string(sequences[x], seq_len) |
There was a problem hiding this comment.
Nooooo. str is a builtin type. result_string?
| seq_string += char | ||
| return seq_string | ||
|
|
||
| def wer(self, output, target): |
There was a problem hiding this comment.
In general acronyms as function names is discouraged.
|
|
||
| return distance.edit_distance(''.join(new_output), ''.join(new_target)) | ||
|
|
||
| def cer(self, output, target): |
|
|
||
| # Use multiprocessing for feature/label extraction | ||
| num_cores = multiprocessing.cpu_count() | ||
| pool = multiprocessing.Pool(processes=num_cores) |
There was a problem hiding this comment.
One trick I learned recently is that contextlib can let you use a context manager in 2 & 3.
with contextlib.closing(multiprocessing.Pool(processes=num_cores)):
pool.map(...)
4 participants
Hi All,
We add the model evaluation part (Haoliang works on it), and also apply multiprocessing to dataset in this PR. Let us know if you have any comments. Thank you!