[Question]: Multi-Task Learning with use_all_task

[EmbedCrafter]

Question

How can I correctly train two tasks simultaneously on a single corpus using the parameter use_all_task=True? When I attempted to train two models together on one corpus, I encountered a RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn.

multitask_dataset = CONLL_03_DUTCH()
tasks = ['ner', 'pos']
model_1 = initialize_tagger(multitask_dataset, shared_embedding, tasks[0])
model_2 = initialize_tagger(multitask_dataset, shared_embedding, tasks[1])
multitask_model = MultitaskModel([model_1, model_2], use_all_tasks=True, task_ids=tasks)
trainer = ModelTrainer(multitask_model, multitask_dataset)
trainer.fine_tune('resources/taggers/sota-ner-flert',
learning_rate=5.0e-6,
max_epochs=20)
╭─────────────────────────────────────── Traceback (most recent call last) ───────────────────────────────────────╮
│ in :60 │
│ │
│ ❱ 60 trainer.fine_tune('resources/taggers/sota-ner-flert', │
│ │
│ /pyzr/active_venv/lib/python3.10/site-packages/flair/trainers/trainer.py:253 in fine_tune │
│ │
│ ❱ 253 │ │ return self.train_custom( │
│ │
│ /pyzr/active_venv/lib/python3.10/site-packages/flair/trainers/trainer.py:606 in train_custom │
│ │
│ ❱ 606 │ │ │ │ │ │ │ self._backward(scaler.scale(loss)) │
│ │
│ /pyzr/active_venv/lib/python3.10/site-packages/flair/trainers/trainer.py:124 in _backward │
│ │
│ ❱ 124 │ │ loss.backward() │
│ │
│ /pyzr/active_venv/lib/python3.10/site-packages/torch/_tensor.py:487 in backward │
│ │
│ ❱ 487 │ │ torch.autograd.backward( │
│ │
│ /pyzr/active_venv/lib/python3.10/site-packages/torch/autograd/init.py:200 in backward │
│ │
│ ❱ 200 │ Variable._execution_engine.run_backward( # Calls into the C++ engine to run the bac │
╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

[fkdosilovic]

If you look at the docs, you'll see you need to explicitly add a special multitask_id label type. Here's a minimal working example using the CONLL_03_DUTCH dataset:

from flair.datasets import CONLL_03_DUTCH
from flair.embeddings import TransformerWordEmbeddings
from flair.models import SequenceTagger, MultitaskModel
from flair.trainers import ModelTrainer

corpus = CONLL_03_DUTCH()

# Adding the special label type to make use_all_tasks=True work.
for corpus_split in (corpus.train, corpus.dev, corpus.test):
    for sent in corpus_split:
        sent.add_label("multitask_id", "task_ner")
        sent.add_label("multitask_id", "task_pos")
        
# Preparing the label dictionary for each task.
ner_label_dict = corpus.make_label_dictionary("ner")
pos_label_dict = corpus.make_label_dictionary("pos")

# Preparing the shared embeddings and tagger for each task.
shared_embeddings = TransformerWordEmbeddings(
    model="FacebookAI/roberta-base",
    layers="-1",
    subtoken_pooling="first",
    fine_tune=True,
    use_context=False,
)

ner_tagger_model = SequenceTagger(
    hidden_size=256,
    embeddings=shared_embeddings,
    tag_dictionary=ner_label_dict,
    tag_type="ner",
    use_crf=False,
    use_rnn=False,
    reproject_embeddings=False,
)

pos_tagger_model = SequenceTagger(
    hidden_size=256,
    embeddings=shared_embeddings,
    tag_dictionary=pos_label_dict,
    tag_type="pos",
    use_crf=False,
    use_rnn=False,
    reproject_embeddings=False,
)

# Preparing the multitask model.
multitask_model = MultitaskModel(
    models=[ner_tagger_model, pos_tagger_model],
    task_ids=["task_ner", "task_pos"],
    use_all_tasks=True,
)

# Train the multitask model.
trainer = ModelTrainer(model=multitask_model, corpus=corpus)
trainer.fine_tune(
    base_path="ner_pos_mt_model_v1",
    warmup_fraction=0.06,
    learning_rate=2.0e-5,
    mini_batch_size=32,
    train_with_dev=False,
    train_with_test=False,
    max_epochs=3,
)