The task is to detect whether a blood label is present in a text sequence using BERT for sequence classification pre-trained model.
class BloodDataset(Dataset):
"""MIMIC Blood dataset."""
def __init__(self, arff_file):
"""
Args:
arff_file (string): Path to the arff file with annotations.
"""
self.indices, self.contents, self.labels = read_arff(arff_file)
self.labels = torch.as_tensor(self.labels)
self.inputs = encode(self.contents)
self.input_ids = (self.inputs['input_ids'])
self.attention_mask = (self.inputs['attention_mask'])
def __len__(self):
return len(self.indices)
def __getitem__(self, idx):
if torch.is_tensor(idx):
idx = idx.tolist()
if idx in self.indices:
sample_index = self.indices.index(idx)
sample = {'index': idx,
'content': self.contents[sample_index],
'label': self.labels[sample_index],
'input_ids': self.input_ids[sample_index],
'attention_mask': self.attention_mask[sample_index]
}
return sample
else:
return "Sample not found!"
Tutorial from Huggingface proposes a trainer solution:
model = BertForSequenceClassification.from_pretrained(model_type)
training_args = TrainingArguments(
output_dir='./results', # output directory
logging_dir='./logs', # directory for storing logs
)
trainer = Trainer(
# the instantiated Transformers model to be trained
model=model,
args=training_args,
train_dataset=train_dataset, # training dataset
eval_dataset=test_dataset # evaluation dataset
)
return trainer
https://huggingface.co/transformers/training.html#trainer
train_dataset = preprocess.BloodDataset("test_blood.arff")
trainer = train.run(train_dataset, train_dataset)
trainer.train()
yields error:
/.local/lib/python3.6/site-packages/transformers/data/data_collator.py", line 38, in <listcomp>
features = [vars(f) for f in features]
TypeError: vars() argument must have __dict__ attribute
What's a proper dataset input for the trainer?