Predicting on trained model: errors

[ianbstewart]

When I train a model on a toy dataset and then try to predict coreference, I get a strange error related to the char_map of the CorefResult object. Full example below:

prepare data

from fastcoref import FCoref, LingMessCoref
model = FCoref(device='cpu')
texts = ['We are so happy to see you using our coref package. This package is very fast!']
preds = model.predict(texts=texts, output_file='train_file_with_clusters.jsonlines')

training

from fastcoref import TrainingArgs, CorefTrainer
args = TrainingArgs(
output_dir='fast_coref_fine_tune',
overwrite_output_dir=True,
model_name_or_path='distilroberta-base',
device='cpu',
epochs=1,
logging_steps=1,
eval_steps=1,
) # you can control other arguments such as learning head and others.

trainer = CorefTrainer(
args=args,
train_file='train_file_with_clusters.jsonlines',
dev_file='train_file_with_clusters.jsonlines',
)
trainer.train()
trainer.evaluate(test=True)

test model

out_dir = 'fast_coref_fine_tune/model/'
trained_model = FCoref(model_name_or_path=out_dir, device='cpu')
preds = trained_model.predict(
texts=['I went to the store and saw my friend, then I saw her later at the cinema'],
)
print(preds[0])

Error:

TypeError Traceback (most recent call last)
File :1
----> 1 print(preds[0])

File /local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.9/site-packages/fastcoref/modeling.py:63, in CorefResult.str(self)
61 else:
62 text_to_print = self.text
---> 63 return f'CorefResult(text="{text_to_print}", clusters={self.get_clusters()})'

File /local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.9/site-packages/fastcoref/modeling.py:41, in CorefResult.get_clusters(self, as_strings)
38 if not as_strings:
39 return [[self.char_map[mention][1] for mention in cluster] for cluster in self.clusters]
---> 41 return [[self.text[self.char_map[mention][1][0]:self.char_map[mention][1][1]] for mention in cluster]
42 for cluster in self.clusters]

File /local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.9/site-packages/fastcoref/modeling.py:41, in (.0)
38 if not as_strings:
39 return [[self.char_map[mention][1] for mention in cluster] for cluster in self.clusters]
---> 41 return [[self.text[self.char_map[mention][1][0]:self.char_map[mention][1][1]] for mention in cluster]
42 for cluster in self.clusters]

File /local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.9/site-packages/fastcoref/modeling.py:41, in (.0)
38 if not as_strings:
39 return [[self.char_map[mention][1] for mention in cluster] for cluster in self.clusters]
---> 41 return [[self.text[self.char_map[mention][1][0]:self.char_map[mention][1][1]] for mention in cluster]
42 for cluster in self.clusters]

TypeError: 'NoneType' object is not subscriptable

[shon-otmazgin]

hello @ianbstewart I was able to reproduce it. will debug it today/tmrw Thanks!

[shon-otmazgin]

OK I think I fixed it.

since the model just trained on 1 example, it failed to create spans that "really" coffering, some of the spans the model suggested were the entire text and even spans with a special token and there raised an Exception.

I released a new version to PyPi version 2.1.5

you can do pip install fastcoref==2.1.5

Let me know if it helped

[ianbstewart]

Yup, it works now thanks! New output from the same code above:

pred = CorefResult(text="I went to the store and saw my friend, then I saw ...", clusters=[])