Type mismatch error during operation on scores array #65
Description
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''