Réponses aux questions (TensorFlow)

!pip install datasets transformers[sentencepiece]
!apt install git-lfs

!git config --global user.email "[email protected]"
!git config --global user.name "Your Name"

from huggingface_hub import notebook_login

notebook_login()

from datasets import load_dataset
raw_datasets = load_dataset("piaf")

# piaf n'ayant pas de jeu de données de validation, nous en créons un
raw_datasets = raw_datasets['train']
raw_datasets = raw_datasets.train_test_split(test_size=0.2, shuffle=True)

raw_datasets

print("Context: ", raw_datasets["train"][0]["context"])
print("Question: ", raw_datasets["train"][0]["question"])
print("Answer: ", raw_datasets["train"][0]["answers"])

raw_datasets["train"].filter(lambda x: len(x["answers"]["text"]) != 1)

print(raw_datasets["test"][0]["answers"])
print(raw_datasets["test"][2]["answers"])

print(raw_datasets["test"][2]["context"])
print(raw_datasets["test"][2]["question"])

from transformers import AutoTokenizer

model_checkpoint = "camembert-base"
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)

tokenizer.is_fast

context = raw_datasets["train"][0]["context"]
question = raw_datasets["train"][0]["question"]

inputs = tokenizer(question, context)
tokenizer.decode(inputs["input_ids"])

inputs = tokenizer(
    question,
    context,
    max_length=100,
    truncation="only_second",
    stride=50,
    return_overflowing_tokens=True,
)

for ids in inputs["input_ids"]:
    print(tokenizer.decode(ids))

inputs = tokenizer(
    question,
    context,
    max_length=100,
    truncation="only_second",
    stride=50,
    return_overflowing_tokens=True,
    return_offsets_mapping=True,
)
inputs.keys()

inputs["overflow_to_sample_mapping"]

inputs = tokenizer(
    raw_datasets["train"][2:6]["question"],
    raw_datasets["train"][2:6]["context"],
    max_length=100,
    truncation="only_second",
    stride=50,
    return_overflowing_tokens=True,
    return_offsets_mapping=True,
)

print(f"The 4 examples gave {len(inputs['input_ids'])} features.")
print(f"Here is where each comes from: {inputs['overflow_to_sample_mapping']}.")

answers = raw_datasets["train"][2:6]["answers"]
start_positions = []
end_positions = []

for i, offset in enumerate(inputs["offset_mapping"]):
    sample_idx = inputs["overflow_to_sample_mapping"][i]
    answer = answers[sample_idx]
    start_char = answer["answer_start"][0]
    end_char = answer["answer_start"][0] + len(answer["text"][0])
    sequence_ids = inputs.sequence_ids(i)

    # Trouver le début et la fin du contexte
    idx = 0
    while sequence_ids[idx] != 1:
        idx += 1
    context_start = idx
    while sequence_ids[idx] == 1:
        idx += 1
    context_end = idx - 1

    # Si la réponse n'est pas entièrement dans le contexte, l'étiquette est (0, 0)
    if offset[context_start][0] > start_char or offset[context_end][1] < end_char:
        start_positions.append(0)
        end_positions.append(0)
    else:
        # Sinon, ce sont les positions de début et de fin du token
        idx = context_start
        while idx <= context_end and offset[idx][0] <= start_char:
            idx += 1
        start_positions.append(idx - 1)

        idx = context_end
        while idx >= context_start and offset[idx][1] >= end_char:
            idx -= 1
        end_positions.append(idx + 1)

start_positions, end_positions

idx = 0
sample_idx = inputs["overflow_to_sample_mapping"][idx]
answer = answers[sample_idx]["text"][0]

start = start_positions[idx]
end = end_positions[idx]
labeled_answer = tokenizer.decode(inputs["input_ids"][idx][start : end + 1])

print(f"Theoretical answer: {answer}, labels give: {labeled_answer}")

idx = 4
sample_idx = inputs["overflow_to_sample_mapping"][idx]
answer = answers[sample_idx]["text"][0]

decoded_example = tokenizer.decode(inputs["input_ids"][idx])
print(f"Theoretical answer: {answer}, decoded example: {decoded_example}")

max_length = 384
stride = 128


def preprocess_training_examples(examples):
    questions = [q.strip() for q in examples["question"]]
    inputs = tokenizer(
        questions,
        examples["context"],
        max_length=max_length,
        truncation="only_second",
        stride=stride,
        return_overflowing_tokens=True,
        return_offsets_mapping=True,
        padding="max_length",
    )

    offset_mapping = inputs.pop("offset_mapping")
    sample_map = inputs.pop("overflow_to_sample_mapping")
    answers = examples["answers"]
    start_positions = []
    end_positions = []

    for i, offset in enumerate(offset_mapping):
        sample_idx = sample_map[i]
        answer = answers[sample_idx]
        start_char = answer["answer_start"][0]
        end_char = answer["answer_start"][0] + len(answer["text"][0])
        sequence_ids = inputs.sequence_ids(i)

        # Trouver le début et la fin du contexte
        idx = 0
        while sequence_ids[idx] != 1:
            idx += 1
        context_start = idx
        while sequence_ids[idx] == 1:
            idx += 1
        context_end = idx - 1

        # Si la réponse n'est pas entièrement dans le contexte, l'étiquette est (0, 0)
        if offset[context_start][0] > start_char or offset[context_end][1] < end_char:
            start_positions.append(0)
            end_positions.append(0)
        else:
            # Sinon, ce sont les positions de début et de fin du token
            idx = context_start
            while idx <= context_end and offset[idx][0] <= start_char:
                idx += 1
            start_positions.append(idx - 1)

            idx = context_end
            while idx >= context_start and offset[idx][1] >= end_char:
                idx -= 1
            end_positions.append(idx + 1)

    inputs["start_positions"] = start_positions
    inputs["end_positions"] = end_positions
    return inputs

train_dataset = raw_datasets["train"].map(
    preprocess_training_examples,
    batched=True,
    remove_columns=raw_datasets["train"].column_names,
)
len(raw_datasets["train"]), len(train_dataset)

def preprocess_validation_examples(examples):
    questions = [q.strip() for q in examples["question"]]
    inputs = tokenizer(
        questions,
        examples["context"],
        max_length=max_length,
        truncation="only_second",
        stride=stride,
        return_overflowing_tokens=True,
        return_offsets_mapping=True,
        padding="max_length",
    )

    sample_map = inputs.pop("overflow_to_sample_mapping")
    example_ids = []

    for i in range(len(inputs["input_ids"])):
        sample_idx = sample_map[i]
        example_ids.append(examples["id"][sample_idx])

        sequence_ids = inputs.sequence_ids(i)
        offset = inputs["offset_mapping"][i]
        inputs["offset_mapping"][i] = [
            o if sequence_ids[k] == 1 else None for k, o in enumerate(offset)
        ]

    inputs["example_id"] = example_ids
    return inputs

validation_dataset = raw_datasets["test"].map(
    preprocess_validation_examples,
    batched=True,
    remove_columns=raw_datasets["test"].column_names,
)
len(raw_datasets["test"]), len(validation_dataset)

small_eval_set = raw_datasets["test"].select(range(100))
trained_checkpoint = "etalab-ia/camembert-base-squadFR-fquad-piaf"

tokenizer = AutoTokenizer.from_pretrained(trained_checkpoint)
eval_set = small_eval_set.map(
    preprocess_validation_examples,
    batched=True,
    remove_columns=raw_datasets["test"].column_names,
)

import tensorflow as tf
from transformers import TFAutoModelForQuestionAnswering

eval_set_for_model = eval_set.remove_columns(["example_id", "offset_mapping"])
eval_set_for_model.set_format("numpy")

batch = {k: eval_set_for_model[k] for k in eval_set_for_model.column_names}
trained_model = TFAutoModelForQuestionAnswering.from_pretrained(trained_checkpoint)

outputs = trained_model(**batch)

start_logits = outputs.start_logits.numpy()
end_logits = outputs.end_logits.numpy()

import collections

example_to_features = collections.defaultdict(list)
for idx, feature in enumerate(eval_set):
    example_to_features[feature["example_id"]].append(idx)

import numpy as np

n_best = 20
max_answer_length = 30
predicted_answers = []

for example in small_eval_set:
    example_id = example["id"]
    context = example["context"]
    answers = []

    for feature_index in example_to_features[example_id]:
        start_logit = start_logits[feature_index]
        end_logit = end_logits[feature_index]
        offsets = eval_set["offset_mapping"][feature_index]

        start_indexes = np.argsort(start_logit)[-1 : -n_best - 1 : -1].tolist()
        end_indexes = np.argsort(end_logit)[-1 : -n_best - 1 : -1].tolist()
        for start_index in start_indexes:
            for end_index in end_indexes:
                # Ignorez les réponses qui ne sont pas entièrement dans le contexte
                if offsets[start_index] is None or offsets[end_index] is None:
                    continue
                # Ignorer les réponses dont la longueur est soit < 0 soit > max_answer_length
                if (
                    end_index < start_index
                    or end_index - start_index + 1 > max_answer_length
                ):
                    continue

                answers.append(
                    {
                        "text": context[offsets[start_index][0] : offsets[end_index][1]],
                        "logit_score": start_logit[start_index] + end_logit[end_index],
                    }
                )

    best_answer = max(answers, key=lambda x: x["logit_score"])
    predicted_answers.append({"id": example_id, "prediction_text": best_answer["text"]})

from datasets import load_metric

metric = load_metric("squad")

theoretical_answers = [
    {"id": ex["id"], "answers": ex["answers"]} for ex in small_eval_set
]

print(predicted_answers[0])
print(theoretical_answers[0])

metric.compute(predictions=predicted_answers, references=theoretical_answers)

from tqdm.auto import tqdm


def compute_metrics(start_logits, end_logits, features, examples):
    example_to_features = collections.defaultdict(list)
    for idx, feature in enumerate(features):
        example_to_features[feature["example_id"]].append(idx)

    predicted_answers = []
    for example in tqdm(examples):
        example_id = example["id"]
        context = example["context"]
        answers = []

        # Parcourir en boucle toutes les fonctionnalités associées à cet exemple
        for feature_index in example_to_features[example_id]:
            start_logit = start_logits[feature_index]
            end_logit = end_logits[feature_index]
            offsets = features[feature_index]["offset_mapping"]

            start_indexes = np.argsort(start_logit)[-1 : -n_best - 1 : -1].tolist()
            end_indexes = np.argsort(end_logit)[-1 : -n_best - 1 : -1].tolist()
            for start_index in start_indexes:
                for end_index in end_indexes:
                    # Ignorez les réponses qui ne sont pas entièrement dans le contexte
                    if offsets[start_index] is None or offsets[end_index] is None:
                        continue
                    # Sauter les réponses dont la longueur est soit < 0, soit > max_answer_length
                    if (
                        end_index < start_index
                        or end_index - start_index + 1 > max_answer_length
                    ):
                        continue

                    answer = {
                        "text": context[offsets[start_index][0] : offsets[end_index][1]],
                        "logit_score": start_logit[start_index] + end_logit[end_index],
                    }
                    answers.append(answer)

        # Sélectionnez la réponse avec le meilleur score
        if len(answers) > 0:
            best_answer = max(answers, key=lambda x: x["logit_score"])
            predicted_answers.append(
                {"id": example_id, "prediction_text": best_answer["text"]}
            )
        else:
            predicted_answers.append({"id": example_id, "prediction_text": ""})

    theoretical_answers = [{"id": ex["id"], "answers": ex["answers"]} for ex in examples]
    return metric.compute(predictions=predicted_answers, references=theoretical_answers)

compute_metrics(start_logits, end_logits, eval_set, small_eval_set)

model = TFAutoModelForQuestionAnswering.from_pretrained(model_checkpoint)

from transformers import DefaultDataCollator

data_collator = DefaultDataCollator(return_tensors="tf")

tf_train_dataset = model.prepare_tf_dataset(
    train_dataset,
    collate_fn=data_collator,
    shuffle=True,
    batch_size=16)

tf_eval_dataset = model.prepare_tf_dataset(
    validation_dataset,
    collate_fn=data_collator,
    shuffle=False,
    batch_size=16)

from transformers import create_optimizer
from transformers.keras_callbacks import PushToHubCallback
import tensorflow as tf

# Le nombre d'étapes d'entraînement est le nombre d'échantillons dans le jeu de données, divisé par la taille du batch puis multiplié
# par le nombre total d'époques. Notez que le jeu de données tf_train_dataset est ici un batch de données tf.data.Dataset,
# pas le jeu de données original Hugging Face, donc son len() est déjà num_samples // batch_size.
num_train_epochs = 3
num_train_steps = len(tf_train_dataset) * num_train_epochs
optimizer, schedule = create_optimizer(
    init_lr=2e-5,
    num_warmup_steps=0,
    num_train_steps=num_train_steps,
    weight_decay_rate=0.01,
)
model.compile(optimizer=optimizer)

# Entraîner en mixed-precision float16
tf.keras.mixed_precision.set_global_policy("mixed_float16")

from transformers.keras_callbacks import PushToHubCallback

callback = PushToHubCallback(output_dir="camembert-base-finetuned-piaf", tokenizer=tokenizer)

# Nous allons faire la validation après, donc pas de validation au milieu de l'entraînement
model.fit(tf_train_dataset, callbacks=[callback], epochs=num_train_epochs)

predictions = model.predict(tf_eval_dataset)
compute_metrics(
    predictions["start_logits"],
    predictions["end_logits"],
    validation_dataset,
    raw_datasets["test"],
)

from transformers import pipeline

# Remplacez par votre propre checkpoint
model_checkpoint = "huggingface-course/camembert-finetuned-piaf"
question_answerer = pipeline("question-answering", model=model_checkpoint)

context = """
🤗 Transformers est soutenu par les trois bibliothèques d'apprentissage profond les plus populaires - Jax, PyTorch et TensorFlow - avec une intégration transparente entre elles. Il est simple d'entraîner vos modèles avec l'une avant de les charger pour l'inférence avec l'autre.
"""
question = "Quelles sont les bibliothèques d'apprentissage profond derrière 🤗 Transformers ?"
question_answerer(question=question, context=context)