section6_tf.ipynb

Kernel: Python 3

Entraîner un modèle de langage causal de zéro (TensorFlow)

Ici nous entraînons un modèle à générer du code Python. Le Python utilisant des fonctions basées sur des mots anglais, nous gardons un gpt-2 anglais dans l'optique d'obtenir de meilleures performances que ce que l'on pourrait s'attendre en utilisant un gpt-2 en français.

Installez les bibliothèques 🤗 Datasets et 🤗 Transformers pour exécuter ce notebook.

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

Vous aurez besoin de configurer git, adaptez votre email et votre nom dans la cellule suivante.

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

Vous devrez également être connecté au Hub d'Hugging Face. Exécutez ce qui suit et entrez vos informations d'identification.

from huggingface_hub import notebook_login

notebook_login()

def any_keyword_in_string(string, keywords):
    for keyword in keywords:
        if keyword in string:
            return True
    return False

filters = ["pandas", "sklearn", "matplotlib", "seaborn"]
example_1 = "import numpy as np"
example_2 = "import pandas as pd"

print(
    any_keyword_in_string(example_1, filters), any_keyword_in_string(example_2, filters)
)

from collections import defaultdict
from tqdm import tqdm
from datasets import Dataset


def filter_streaming_dataset(dataset, filters):
    filtered_dict = defaultdict(list)
    total = 0
    for sample in tqdm(iter(dataset)):
        total += 1
        if any_keyword_in_string(sample["content"], filters):
            for k, v in sample.items():
                filtered_dict[k].append(v)
    print(f"{len(filtered_dict['content'])/total:.2%} of data after filtering.")
    return Dataset.from_dict(filtered_dict)

# Cette cellule prendra beaucoup de temps à s'exécuter, donc vous devriez la sauter et aller à la suivante !
from datasets import load_dataset

split = "train"  # "valid"
filters = ["pandas", "sklearn", "matplotlib", "seaborn"]

data = load_dataset(f"transformersbook/codeparrot-{split}", split=split, streaming=True)
filtered_data = filter_streaming_dataset(data, filters)

from datasets import load_dataset, DatasetDict

ds_train = load_dataset("huggingface-course/codeparrot-ds-train", split="train")
ds_valid = load_dataset("huggingface-course/codeparrot-ds-valid", split="validation")

raw_datasets = DatasetDict(
    {
        "train": ds_train,  # .shuffle().select(range(50000)),
        "valid": ds_valid,  # .shuffle().select(range(500))
    }
)

raw_datasets

for key in raw_datasets["train"][0]:
    print(f"{key.upper()}: {raw_datasets['train'][0][key][:200]}")

from transformers import AutoTokenizer

context_length = 128
tokenizer = AutoTokenizer.from_pretrained("huggingface-course/code-search-net-tokenizer")

outputs = tokenizer(
    raw_datasets["train"][:2]["content"],
    truncation=True,
    max_length=context_length,
    return_overflowing_tokens=True,
    return_length=True,
)

print(f"Input IDs length: {len(outputs['input_ids'])}")
print(f"Input chunk lengths: {(outputs['length'])}")
print(f"Chunk mapping: {outputs['overflow_to_sample_mapping']}")

def tokenize(element):
    outputs = tokenizer(
        element["content"],
        truncation=True,
        max_length=context_length,
        return_overflowing_tokens=True,
        return_length=True,
    )
    input_batch = []
    for length, input_ids in zip(outputs["length"], outputs["input_ids"]):
        if length == context_length:
            input_batch.append(input_ids)
    return {"input_ids": input_batch}


tokenized_datasets = raw_datasets.map(
    tokenize, batched=True, remove_columns=raw_datasets["train"].column_names
)
tokenized_datasets

from transformers import AutoTokenizer, TFGPT2LMHeadModel, AutoConfig

config = AutoConfig.from_pretrained(
    "gpt2",
    vocab_size=len(tokenizer),
    n_ctx=context_length,
    bos_token_id=tokenizer.bos_token_id,
    eos_token_id=tokenizer.eos_token_id,
)

model = TFGPT2LMHeadModel(config)
model(model.dummy_inputs)  # Construit le modèle
model.summary()

from transformers import DataCollatorForLanguageModeling

tokenizer.pad_token = tokenizer.eos_token
data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False, return_tensors="tf")

out = data_collator([tokenized_datasets["train"][i] for i in range(5)])
for key in out:
    print(f"{key} shape: {out[key].shape}")

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(
    tokenized_dataset["valid"],
    collate_fn=data_collator,
    shuffle=False,
    batch_size=32,
)

from huggingface_hub import notebook_login

notebook_login()

from transformers import create_optimizer
import tensorflow as tf

num_train_steps = len(tf_train_dataset)
optimizer, schedule = create_optimizer(
    init_lr=5e-5,
    num_warmup_steps=1_000,
    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="codeparrot-ds", tokenizer=tokenizer)

model.fit(tf_train_dataset, validation_data=tf_eval_dataset, callbacks=[callback])

from transformers import pipeline

course_model = TFGPT2LMHeadModel.from_pretrained("huggingface-course/codeparrot-ds")
course_tokenizer = AutoTokenizer.from_pretrained("huggingface-course/codeparrot-ds")
pipe = pipeline(
    "text-generation", model=course_model, tokenizer=course_tokenizer, device=0
)

txt = """\
# create some data
x = np.random.randn(100)
y = np.random.randn(100)

# create scatter plot with x, y
"""
print(pipe(txt, num_return_sequences=1)[0]["generated_text"])

txt = """\
# create some data
x = np.random.randn(100)
y = np.random.randn(100)

# create dataframe from x and y
"""
print(pipe(txt, num_return_sequences=1)[0]["generated_text"])

txt = """\
# dataframe with profession, income and name
df = pd.DataFrame({'profession': x, 'income':y, 'name': z})

# calculate the mean income per profession
"""
print(pipe(txt, num_return_sequences=1)[0]["generated_text"])

txt = """
# import random forest regressor from scikit-learn
from sklearn.ensemble import RandomForestRegressor

# fit random forest model with 300 estimators on X, y:
"""
print(pipe(txt, num_return_sequences=1)[0]["generated_text"])