Type mismatch error during operation on scores array

[brieucdandin]

TL;DR

When defining a pipeline's elements and testing it on a couple of sentences, the tests results are inconsistent. Some go smoothly, while some encounter the following type error:
TypeError: ufunc 'true_divide' output (typecode 'd') could not be coerced to provided output parameter (typecode 'l') according to the casting rule ''same_kind''
Some alternate depending on runs.

It appears it is due to a type mismatch when attempting an operation to obtain an array scores: one array is populated with ints and the other with floats.

Am I missing something? Is it just due to a package version issue?
The full stack of the error is at the bottom of the post.

Environment

To make the package work, I first had to go through a few ad-hoc tweaks that might explain the issue:

pip install aspect-based-sentiment-analysis
pip install --upgrade protobuf
python3 -m spacy download en

from transformers import BertTokenizer

Note Python 3.7 is used:

print(sys.version)

Output: 3.7.10 (default, Jun  4 2021, 14:48:32) 
[GCC 7.5.0]

Tests

I initially conducted the following tests:

tests_list = [{'text' : "We are great fans of Slack, but we wish the subscriptions were more accessible to small startups.",
               'aspects' : ['slack', 'price']
              },
              {'text' : "I love it overall, but the screen itself is terrible.",
               'aspects' : ['overall', 'screen']
              },
              {'text' : "I love it.",
               'aspects' : ['it']
              },
              {'text' : "I love it!",
               'aspects' : ['it']
              },
              {'text' : "I love it. :)",
               'aspects' : ['it']
              },
              {'text' : "I hate it.",
               'aspects' : ['it']
              },
              {'text' : "I love it: the sound is warm, the ambiance nice and the staff is great! The image is just OK, but who cares?",
               'aspects' : ['it', 'sound', 'ambiance', 'staff', 'image']
              },
             ]

Functional default pipeline

The preset pipeline works just fine.

nlp_load = absa.load()

for text_to_test in tests_list:
  text = text_to_test['text']
  aspects = text_to_test['aspects']
  features_sentiments_list = nlp_load(text, aspects=aspects)
  print(text)
  for feature_index in range(len(features_sentiments_list.subtasks)):
    aspect = aspects[feature_index]
    if aspect != '_':
      print(aspect, ':\t', features_sentiments_list.subtasks[aspect].sentiment)
  print()

returns:

We are great fans of Slack, but we wish the subscriptions were more accessible to small startups.
slack :	 Sentiment.positive
price :	 Sentiment.negative

I love it overall, but the screen itself is terrible.
overall :	 Sentiment.positive
screen :	 Sentiment.negative

I love it.
it :	 Sentiment.positive

I love it!
it :	 Sentiment.positive

I love it. :)
it :	 Sentiment.positive

I hate it.
it :	 Sentiment.negative

I love it: the sound is warm, the ambiance nice and the staff is great! The image is just OK, but who cares?
it :	 Sentiment.positive
sound :	 Sentiment.positive
ambiance :	 Sentiment.positive
staff :	 Sentiment.positive
image :	 Sentiment.neutral

Inconsistently dysfunctional pipeline defined by hand

Yet, with a pipeline in which the components are defined (the one from the README), we obtain the type mismatch error on certain tests and not others.

Pipeline definition:

model_name = 'absa/classifier-rest-0.2'
model = absa.BertABSClassifier.from_pretrained(model_name)
tokenizer = BertTokenizer.from_pretrained(model_name)

prof_name = 'absa/basic_reference_recognizer-rest-0.1'
recognizer = absa.aux_models.BasicReferenceRecognizer.from_pretrained(prof_name)
professor = absa.Professor(reference_recognizer=recognizer)

text_splitter = absa.sentencizer() # The English CNN model from SpaCy.
nlp_pipeline = absa.Pipeline(model, tokenizer, professor, text_splitter)

Functional test

text = "I like it overall, but the text is long. This is just a bunch of letters and words to fill in the gap and make sure the error is not due to the length of the text."
aspects = ['overall', 'text']
print(text, aspects, '\n')
nlp_pipeline(text, aspects=aspects)

returns:

I like it overall, but the text is long. This is just a bunch of letters and words to fill in the gap and make sure the error is not due to the length of the text. ['overall', 'text'] 

Out[17]: CompletedTask( ... )

Most of the tests in tests_list work like charms too (most of the time):

for text_to_test in tests_list[1:]: # TODO: Why does tests_list[0] return an error?
  text = text_to_test['text']
  aspects = text_to_test['aspects']
  print(text, aspects)
  features_sentiments_list = nlp_pipeline(text, aspects=aspects)
  for feature_index in range(len(features_sentiments_list.subtasks)):
    aspect = aspects[feature_index]
    if aspect != '_':
      print(aspect, ':\t', features_sentiments_list.subtasks[aspect].sentiment)
  print()

returns:

I love it overall, but the screen itself is terrible. ['overall', 'screen']
overall :	 Sentiment.positive
screen :	 Sentiment.negative

I love it. ['it']
it :	 Sentiment.positive

I love it! ['it']
it :	 Sentiment.positive

I love it. :) ['it']
it :	 Sentiment.positive

I hate it. ['it']
it :	 Sentiment.negative

I love it: the sound is warm, the ambiance nice and the staff is great! The image is just OK, but who cares? ['it', 'sound', 'ambiance', 'staff', 'image']
it :	 Sentiment.positive
sound :	 Sentiment.positive
ambiance :	 Sentiment.positive
staff :	 Sentiment.positive
image :	 Sentiment.neutral

Dysfunctional tests

text = "We are great fans of Slack, but we wish the subscriptions were more accessible to small startups."
aspects = ['slack', 'price']
print(text, aspects, '\n')
nlp_pipeline(text, aspects=aspects)

returns the aforementioned TypeError ' - cf. below for full stack.
The same goes with text set to tests_list[0]['text'], tests_list[1]['text'], 'yo' and "yo".

Which leaves that some tests (like with tests_list[1]) alternatively fail and pass on successive runs, with no apparent correlation.

Full error stack

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<command-1097563962837147> in <module>
      2 aspects = tests_list[0]['aspects']
      3 print(text, aspects, '\n')
----> 4 nlp_pipeline(text, aspects=aspects)

/local_disk0/.ephemeral_nfs/envs/pythonEnv-7b4c08c6-365d-4481-b4a5-c7c62c416c5b/lib/python3.7/site-packages/aspect_based_sentiment_analysis/pipelines.py in __call__(self, text, aspects)
    206         task = self.preprocess(text, aspects)
    207         predictions = self.transform(task.examples)
--> 208         completed_task = self.postprocess(task, predictions)
    209         return completed_task
    210 

/local_disk0/.ephemeral_nfs/envs/pythonEnv-7b4c08c6-365d-4481-b4a5-c7c62c416c5b/lib/python3.7/site-packages/aspect_based_sentiment_analysis/pipelines.py in postprocess(task, batch_examples)
    301             aspect, = {e.aspect for e in examples}
    302             scores = np.max([e.scores for e in examples], axis=0)
--> 303             scores /= np.linalg.norm(scores, ord=1)
    304             sentiment_id = np.argmax(scores).astype(int)
    305             aspect_document = CompletedSubTask(

TypeError: ufunc 'true_divide' output (typecode 'd') could not be coerced to provided output parameter (typecode 'l') according to the casting rule ''same_kind''

[xesaad]

Hi @brieucdandin! Unfortunately not an answer, but just wanted to share that I have also encountered the same issue and opened a PR in an attempt to fix it. It seems that the package author has been unresponsive for some time now though :(