Data format for BertForSequenceClassification with num_labels > 2

Hi,
I have a multilabel task (num_labels=8) and I want to use BertForSequenceClassification using Trainer to train the model.

But I get the following error:

ValueError: Expected input batch_size (8) to match target batch_size (64).

I assume that the problem is the data format of the labels. Currently, my label is a 8-dim list (e.g., [1,0,0,0,0,1,0,0]).

What is the right format for the label data?

Here my code:

class EmotionDataset(torch.utils.data.Dataset):
def init(self, encodings, labels):
self.encodings = encodings
self.labels = labels

def __getitem__(self, idx):
    item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}
    item['labels'] = torch.tensor(self.labels[idx])
    return item

def __len__(self):
    return len(self.labels)

MODEL_NAME = ‘dbmdz/bert-base-german-uncased’

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = BertForSequenceClassification.from_pretrained(MODEL_NAME, num_labels=8)

tokenize data

dataset_train = Dataset.from_pandas(df_train)
train_encodings = tokenizer(dataset_train['text], truncation=True, padding=True)
train_dataset = EmotionDataset(train_encodings, dataset_train['label])

training_args = TrainingArguments(
output_dir=’./results’, # output directory
num_train_epochs=1, # total # of training epochs
per_device_train_batch_size=8, # batch size per device during training
per_device_eval_batch_size=32, # batch size for evaluation
warmup_steps=500, # number of warmup steps for learning rate scheduler
weight_decay=0.01, # strength of weight decay
logging_dir=’./logs’, # directory for storing logs
)

trainer = Trainer(
model=model, # the instantiated :hugs: Transformers model to be trained
args=training_args, # training arguments, defined above
train_dataset=train_dataset, # training dataset
eval_dataset=test_dataset # evaluation dataset
)

_ = trainer.train()
trainer.evaluate()

Thanks,
Max

Hi @maxpower, I think the format of your dataset is fine but I think you have to change the model’s loss function to use a sigmoid instead of a softmax on the logits (i.e. BCEWithLogitsLoss). You can see a skeleton + hacky Colab in this thread: Fine-Tune for MultiClass or MultiLabel-MultiClass - #8 by lewtun

1 Like

Perfect, it works. Thanks so much!

1 Like

FYI I just posted a more elegant solution in the thread that just subclasses Trainer and overrides the compute_loss function (you can see it in action in the Colab notebook too :slight_smile:)

Hi lewtun,
Thanks for your help so far.
But I’m having issues getting it to work for multiclass classification.
The custommetric in the notebook only works for multilabel classification.
Is there anything I need to do please?