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)
)

False True

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)

# 执行这个代码块需要非常长的时间,因此你可以跳过它,继续执行下一个!
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)

3.26% of data after filtering.

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

DatasetDict({
    train: Dataset({
        features: ['repo_name', 'path', 'copies', 'size', 'content', 'license'],
        num_rows: 606720
    })
    valid: Dataset({
        features: ['repo_name', 'path', 'copies', 'size', 'content', 'license'],
        num_rows: 3322
    })
})

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

'REPO_NAME: kmike/scikit-learn'
'PATH: sklearn/utils/__init__.py'
'COPIES: 3'
'SIZE: 10094'
'''CONTENT: """
The :mod:`sklearn.utils` module includes various utilites.
"""

from collections import Sequence

import numpy as np
from scipy.sparse import issparse
import warnings

from .murmurhash import murm
LICENSE: bsd-3-clause'''

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']}")

Input IDs length: 34
Input chunk lengths: [128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 117, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 41]
Chunk mapping: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]

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

DatasetDict({
    train: Dataset({
        features: ['input_ids'],
        num_rows: 16702061
    })
    valid: Dataset({
        features: ['input_ids'],
        num_rows: 93164
    })
})

from transformers import AutoTokenizer, GPT2LMHeadModel, 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 = GPT2LMHeadModel(config)
model_size = sum(t.numel() for t in model.parameters())
print(f"GPT-2 size: {model_size/1000**2:.1f}M parameters")

GPT-2 size: 124.2M parameters

from transformers import DataCollatorForLanguageModeling

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

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

input_ids shape: torch.Size([5, 128])
attention_mask shape: torch.Size([5, 128])
labels shape: torch.Size([5, 128])

from huggingface_hub import notebook_login

notebook_login()

huggingface-cli login

from transformers import Trainer, TrainingArguments

args = TrainingArguments(
    output_dir="codeparrot-ds",
    per_device_train_batch_size=32,
    per_device_eval_batch_size=32,
    evaluation_strategy="steps",
    eval_steps=5_000,
    logging_steps=5_000,
    gradient_accumulation_steps=8,
    num_train_epochs=1,
    weight_decay=0.1,
    warmup_steps=1_000,
    lr_scheduler_type="cosine",
    learning_rate=5e-4,
    save_steps=5_000,
    fp16=True,
    push_to_hub=True,
)

trainer = Trainer(
    model=model,
    tokenizer=tokenizer,
    args=args,
    data_collator=data_collator,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["valid"],
)

trainer.train()

trainer.push_to_hub()

import torch
from transformers import pipeline

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
pipe = pipeline(
    "text-generation", model="huggingface-course/codeparrot-ds", device=device
)

txt = """\
# 创建一些数据
x = np.random.randn(100)
y = np.random.randn(100)

# 使用 x,y 创建散点图
"""
print(pipe(txt, num_return_sequences=1)[0]["generated_text"])

# 创建一些数据
x = np.random.randn(100)
y = np.random.randn(100)

# 使用 x,y 创建散点图
plt.scatter(x, y)

# 创建散点

txt = """\
# 创建一些数据
x = np.random.randn(100)
y = np.random.randn(100)

# 从 x 和 y 创建 dataframe
"""
print(pipe(txt, num_return_sequences=1)[0]["generated_text"])

# 创建一些数据
x = np.random.randn(100)
y = np.random.randn(100)

# 从 x 和 y 创建 dataframe
df = pd.DataFrame({'x': x, 'y': y})
df.insert(0,'x', x)
for

txt = """\
# 有职业,收入和名字的 dataframe
df = pd.DataFrame({'profession': x, 'income':y, 'name': z})

# 计算每个职业的平均收入
"""
print(pipe(txt, num_return_sequences=1)[0]["generated_text"])

# 有职业,收入和名字的 dataframe
df = pd.DataFrame({'profession': x, 'income':y, 'name': z})

# 计算每个职业的平均收入
profession = df.groupby(['profession']).mean()

# 计算

txt = """
# 从 scikit-learn 导入随机森林回归器
from sklearn.ensemble import RandomForestRegressor

# 用 X, y 拟合带有 300 个估算器的随机森林模型:
"""
print(pipe(txt, num_return_sequences=1)[0]["generated_text"])

# 从 scikit-learn 导入随机森林回归器
from sklearn.ensemble import RandomForestRegressor

# 用 X, y 拟合带有 300 个估算器的随机森林模型:
rf = RandomForestRegressor(n_estimators=300, random_state=random_state, max_depth=3)
rf.fit(X, y)
rf

keytoken_ids = []
for keyword in [
    "plt",
    "pd",
    "sk",
    "fit",
    "predict",
    " plt",
    " pd",
    " sk",
    " fit",
    " predict",
    "testtest",
]:
    ids = tokenizer([keyword]).input_ids[0]
    if len(ids) == 1:
        keytoken_ids.append(ids[0])
    else:
        print(f"Keyword has not single token: {keyword}")

'Keyword has not single token: testtest'

from torch.nn import CrossEntropyLoss
import torch


def keytoken_weighted_loss(inputs, logits, keytoken_ids, alpha=1.0):
    # 左移 tokens < n 预测 n
    shift_labels = inputs[..., 1:].contiguous()
    shift_logits = logits[..., :-1, :].contiguous()
    # 计算每一个token的loss
    loss_fct = CrossEntropyLoss(reduce=False)
    loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
    # 对于每个样本重新调整大小并平均
    loss_per_sample = loss.view(shift_logits.size(0), shift_logits.size(1)).mean(axis=1)
    # 计算并缩放权重
    weights = torch.stack([(inputs == kt).float() for kt in keytoken_ids]).sum(
        axis=[0, 2]
    )
    weights = alpha * (1.0 + weights)
    # 计算评价权重
    weighted_loss = (loss_per_sample * weights).mean()
    return weighted_loss

from torch.utils.data.dataloader import DataLoader

tokenized_datasets.set_format("torch")
train_dataloader = DataLoader(tokenized_datasets["train"], batch_size=32, shuffle=True)
eval_dataloader = DataLoader(tokenized_datasets["valid"], batch_size=32)

weight_decay = 0.1


def get_grouped_params(model, no_decay=["bias", "LayerNorm.weight"]):
    params_with_wd, params_without_wd = [], []
    for n, p in model.named_parameters():
        if any(nd in n for nd in no_decay):
            params_without_wd.append(p)
        else:
            params_with_wd.append(p)
    return [
        {"params": params_with_wd, "weight_decay": weight_decay},
        {"params": params_without_wd, "weight_decay": 0.0},
    ]

def evaluate():
    model.eval()
    losses = []
    for step, batch in enumerate(eval_dataloader):
        with torch.no_grad():
            outputs = model(batch["input_ids"], labels=batch["input_ids"])

        losses.append(accelerator.gather(outputs.loss))
    loss = torch.mean(torch.cat(losses))
    try:
        perplexity = torch.exp(loss)
    except OverflowError:
        perplexity = float("inf")
    return loss.item(), perplexity.item()

model = GPT2LMHeadModel(config)

from torch.optim import AdamW

optimizer = AdamW(get_grouped_params(model), lr=5e-4)

from accelerate import Accelerator

accelerator = Accelerator(fp16=True)

model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
    model, optimizer, train_dataloader, eval_dataloader
)

num_train_epochs = 1
num_update_steps_per_epoch = len(train_dataloader)
num_training_steps = num_train_epochs * num_update_steps_per_epoch

lr_scheduler = get_scheduler(
    name="linear",
    optimizer=optimizer,
    num_warmup_steps=1_000,
    num_training_steps=num_training_steps,
)

from huggingface_hub import Repository, get_full_repo_name

model_name = "codeparrot-ds-accelerate"
repo_name = get_full_repo_name(model_name)
repo_name

'sgugger/codeparrot-ds-accelerate'

output_dir = "codeparrot-ds-accelerate"
repo = Repository(output_dir, clone_from=repo_name)

evaluate()

(10.934126853942871, 56057.14453125)

from tqdm.notebook import tqdm

gradient_accumulation_steps = 8
eval_steps = 5_000

model.train()
completed_steps = 0
for epoch in range(num_train_epochs):
    for step, batch in tqdm(
        enumerate(train_dataloader, start=1), total=num_training_steps
    ):
        logits = model(batch["input_ids"]).logits
        loss = keytoken_weighted_loss(batch["input_ids"], logits, keytoken_ids)
        if step % 100 == 0:
            accelerator.print(
                {
                    "samples": step * samples_per_step,
                    "steps": completed_steps,
                    "loss/train": loss.item() * gradient_accumulation_steps,
                }
            )
        loss = loss / gradient_accumulation_steps
        accelerator.backward(loss)
        if step % gradient_accumulation_steps == 0:
            accelerator.clip_grad_norm_(model.parameters(), 1.0)
            optimizer.step()
            lr_scheduler.step()
            optimizer.zero_grad()
            completed_steps += 1
        if (step % (eval_steps * gradient_accumulation_steps)) == 0:
            eval_loss, perplexity = evaluate()
            accelerator.print({"loss/eval": eval_loss, "perplexity": perplexity})
            model.train()
            accelerator.wait_for_everyone()
            unwrapped_model = accelerator.unwrap_model(model)
            unwrapped_model.save_pretrained(output_dir, save_function=accelerator.save)
            if accelerator.is_main_process:
                tokenizer.save_pretrained(output_dir)
                repo.push_to_hub(
                    commit_message=f"Training in progress step {step}", blocking=False
                )