from transformers import pipeline
classifier = pipeline("sentiment-analysis") # 情感分析
result = classifier("I've been waiting for a HuggingFace course my whole life.")
print(result)

from transformers import pipeline
generator = pipeline("text-generation")     # 文本生成
results = generator("In this course, we will teach you how to")
print(results)

# Building the config
config = BertConfig()
# Building the model from the config,使用随机值对其进行初始化
model = BertModel(config)
print(config)
BertConfig {
  [...]
  "hidden_size": 768,               # 定义了hidden状态向量的大小
  "intermediate_size": 3072,
  "max_position_embeddings": 512,
  "num_attention_heads": 12,
  "num_hidden_layers": 12,          # 定义了Transformer模型的层数
  [...]
}
# 加载已经训练过的Transformers模型
model = BertModel.from_pretrained("bert-base-cased")

# 加载与保存分词器
from transformers import BertTokenizer

tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
tokenizer.save_pretrained("./models/bert-base-cased/")
# 加载与保存模型
from transformers import AutoModel
# 所有存储在 HuggingFace Model Hub 上的模型都可以通过 Model.from_pretrained() 来加载权重，参数可以是 checkpoint 的名称，也可以是本地路径（预先下载的模型目录）
model = AutoModel.from_pretrained("bert-base-cased")
model.save_pretrained("./models/bert-base-cased/") # 保存模型

inputs = tokenizer(["来到美丽的大自然，我们发现"], return_tensors="pt")
# {'input_ids': tensor([[    1, 68846, 68881, 67701, 67668, 98899, 91935]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1]])}

gen_kwargs = {"max_length": 128, "top_p": 0.8, "temperature": 0.8, "do_sample": True, "repetition_penalty": 1.1}
output = model.generate(**inputs, **gen_kwargs)
# decode the new tokens
output = tokenizer.decode(output[0].tolist(), skip_special_tokens=True)
print(output)

from datasets import load_dataset
raw_datasets = load_dataset("glue", "mrpc")
print(raw_datasets)
# 包含训练集、验证集和测试集。每一个集合都包含几个列(sentence1, sentence2, label, and idx)以及一个代表行数的变量
#DatasetDict({
#    train: Dataset({
#        features: ['sentence1', 'sentence2', 'label', 'idx'],
#        num_rows: 3668
#    })
#    validation: Dataset({
#        features: ['sentence1', 'sentence2', 'label', 'idx'],
#        num_rows: 408
#    })
#    test: Dataset({
#        features: ['sentence1', 'sentence2', 'label', 'idx'],
#        num_rows: 1725
#    })
#})

# example 是一个dict，对应数据集的每个元素，并返回一个包含input_ids、attention_mask 和token_type_ids为key的新dict
# 在机器学习任务中，一个example通常定义为模型的输入（也成为特征集合）
def tokenize_function(example):
    return tokenizer(example["sentence1"], example["sentence2"], truncation=True)
tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)
print(raw_datasets)
#DatasetDict({
#    train: Dataset({
#        features: ['attention_mask', 'idx', 'input_ids', 'label', 'sentence1', 'sentence2', 'token_type_ids'],
#        num_rows: 3668
#    })
#    validation: Dataset({
#        features: ['attention_mask', 'idx', 'input_ids', 'label', 'sentence1', 'sentence2', 'token_type_ids'],
#        num_rows: 408
#    })
#    test: Dataset({
#        features: ['attention_mask', 'idx', 'input_ids', 'label', 'sentence1', 'sentence2', 'token_type_ids'],
#        num_rows: 1725
#    })
#})

from datasets import load_dataset
from transformers import AutoTokenizer, DataCollatorWithPadding

raw_datasets = load_dataset("glue", "mrpc")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
def tokenize_function(example):
    return tokenizer(example["sentence1"], example["sentence2"], truncation=True)
tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

from datasets import load_dataset
#{
#    "data": [
#        {
#            "title": "Terremoto del Sichuan del 2008",
#            "paragraphs": [{...}]
#        }
#    ],
#     "version": "1.1"
#}
squad_it_dataset = load_dataset("json", data_files="SQuAD_it-train.json", field="data")
print(squad_it_dataset) # 加载本地文件会创建一个带有train的DatasetDict 对象
# DatasetDict({
#    train: Dataset({
#        features: ['title', 'paragraphs'],
#        num_rows: 442
#    })
#})
data_files = {"train": "SQuAD_it-train.json", "test": "SQuAD_it-test.json"}
squad_it_dataset = load_dataset("json", data_files=data_files, field="data")
print(squad_it_dataset) # 包括 train 和 test 的 DatasetDict 对象
#DatasetDict({
#    train: Dataset({
#        features: ['title', 'paragraphs'],
#        num_rows: 442
#    })
#    test: Dataset({
#        features: ['title', 'paragraphs'],
#        num_rows: 48
#    })
#})

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(checkpoint)
sequences = ["I've been waiting for a HuggingFace course my whole life.", "So have I!"]
tokens = tokenizer(sequences, padding=True, truncation=True, return_tensors="pt")
print(tokens)
#{
#    'input_ids': tensor([
#        [  101,  1045,  1005,  2310,  2042,  3403,  2005,  1037, 17662, 12172,2607,  2026,  2878,  2166,  1012,   102],
#        [  101,  2061,  2031,  1045,   999,   102,     0,     0,     0,     0, 0,     0,     0,     0,     0,     0]
#    ]),     
#    'attention_mask': tensor([
#        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
#        [1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
#    ])
#}
output = model(**tokens)
print(output)
# SequenceClassifierOutput(
#    loss=None, 
#    logits=tensor([[-1.5607,  1.6123],[-3.6183,  3.9137]], grad_fn=<AddmmBackward0>), 
#    hidden_states=None, 
#    attentions=None
#)

class Module:
	__call__ : Callable[..., Any] = _call_impl
	def _call_impl(self, *args, **kwargs):
        forward_call = (self._slow_forward if torch._C._get_tracing_state() else self.forward)
       	...
     	result = forward_call(*args, **kwargs)  # dict 可以作为kwargs参数传入
     	... # 涉及到动态图的构建
     	return result

class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin, PeftAdapterMixin):
	def save_pretrained(...)
	@classmethod
    def from_pretrained(...)
class GenerationMixin:
   def generate(inputs,...):> Union[GenerateOutput, torch.LongTensor]:
        ...
        outputs = self(model_input_ids,attention_mask=model_attn,...)
        ...             
class ChatGLMPreTrainedModel(TorchModel, PreTrainedModel):
    ...
class ChatGLMModel(ChatGLMPreTrainedModel):
	def forward(self,input_ids,attention_mask,...):
        ...
class GPT2PreTrainedModel(PreTrainedModel):
    ...
class GPT2LMHeadModel(GPT2PreTrainedModel):
    def __init__(self, config):
        ...
        self.transformer = GPT2Model(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        ...
    def forward(self,input_ids,attention_mask,labels,...):
       transformer_outputs = self.transformer(input_ids,attention_mask,...)
       hidden_states = transformer_outputs[0]
       lm_logits = self.lm_head(hidden_states)
       loss = None
       if labels is not None:
           labels = labels.to(lm_logits.device) # move labels to correct device to enable model parallelism
           shift_logits = lm_logits[..., :-1, :].contiguous() # Shift so that tokens < n predict n
           shift_labels = labels[..., 1:].contiguous()
           loss_fct = CrossEntropyLoss()
           loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
        return ((loss,) + output) if loss is not None else output     

def generate(self,
        prompt_tokens: List[List[int]],  # 输入的提示
        max_gen_len: int,  # 最大生成长度
        temperature: float = 0.6,  # 影响生成文本的随机性
        top_p: float = 0.9,  # 用于决定采样过程中保留的 token 集合的概率阈值
        logprobs: bool = False,  # 是否返回每个 token 的对数概率
        echo: bool = False,  # 是否返回输入的提示
) -> Tuple[List[List[int]], Optional[List[List[float]]]]:
    # ---------------------------初始化长度为 total_len tokens张量，并填充 pad_id----------------------------------
    params = self.model.params
    bsz = len(prompt_tokens)
    assert bsz <= params.max_batch_size, (bsz, params.max_batch_size)

    min_prompt_len = min(len(t) for t in prompt_tokens)
    max_prompt_len = max(len(t) for t in prompt_tokens)
    assert max_prompt_len <= params.max_seq_len
    total_len = min(params.max_seq_len, max_gen_len + max_prompt_len)

    pad_id = self.tokenizer.pad_id
    tokens = torch.full((bsz, total_len), pad_id, dtype=torch.long, device="cuda")
    # 将prompt_tokens中的token复制到tokens张量中。
    for k, t in enumerate(prompt_tokens):
        tokens[k, : len(t)] = torch.tensor(t, dtype=torch.long, device="cuda")
    if logprobs: # 是否返回每个 token 的对数概率
        # 创建一个与tokens相同形状的token_logprobs张量，并用0填充
        token_logprobs = torch.zeros_like(tokens, dtype=torch.float)

    prev_pos = 0
    eos_reached = torch.tensor([False] * bsz, device="cuda")
    input_text_mask = tokens != pad_id
    # -------------------------------------------------------------

    for cur_pos in range(min_prompt_len, total_len):
        # 调用模型的forward方法获取logits
        logits = self.model.forward(tokens[:, prev_pos:cur_pos], prev_pos)
        if logprobs: # 是否返回每个 token 的对数概率
            # 计算token level的logprobs
            token_logprobs[:, prev_pos + 1: cur_pos + 1] = -F.cross_entropy(
                input=logits.transpose(1, 2),
                target=tokens[:, prev_pos + 1: cur_pos + 1],
                reduction="none",
                ignore_index=pad_id,
            )
        # 根据温度参数和top_p参数对logits进行softmax和采样，得到下一个token
        if temperature > 0:
            # sample_top_p函数对probs进行采样
            probs = torch.softmax(logits[:, -1] / temperature, dim=-1)
            next_token = sample_top_p(probs, top_p)
        else:
            # 将logits中概率最大的token作为下一个token。
            next_token = torch.argmax(logits[:, -1], dim=-1)

        next_token = next_token.reshape(-1)
        # only replace token if prompt has already been generated
        next_token = torch.where(
            input_text_mask[:, cur_pos], tokens[:, cur_pos], next_token
        )
        # tokens张量更新
        tokens[:, cur_pos] = next_token
        eos_reached |= (~input_text_mask[:, cur_pos]) & (
                next_token == self.tokenizer.eos_id
        )
        prev_pos = cur_pos
        # 检查是否已经生成了所有的eos token，如果是则停止生成
        if all(eos_reached):
            break

    if logprobs:
        # token_logprobs列表化
        token_logprobs = token_logprobs.tolist()
    out_tokens, out_logprobs = [], []
    for i, toks in enumerate(tokens.tolist()):
        # cut to max gen len
        # 对于 tokens 张量中的每一行（即每一个生成的序列），如果 echo 参数为假，则去掉提示部分
        start = 0 if echo else len(prompt_tokens[i])
        toks = toks[start: len(prompt_tokens[i]) + max_gen_len]
        probs = None
        if logprobs:
            probs = token_logprobs[i][start: len(prompt_tokens[i]) + max_gen_len]
        # cut to eos tok if any
        # 存在结束标记，则去掉结束标记之后的部分
        if self.tokenizer.eos_id in toks:
            eos_idx = toks.index(self.tokenizer.eos_id)
            toks = toks[:eos_idx]
            probs = probs[:eos_idx] if logprobs else None
        out_tokens.append(toks)
        out_logprobs.append(probs)
    # 返回生成的tokens和对数概率（如果logprobs参数为真）
    return (out_tokens, out_logprobs if logprobs else None)

from transformers import TrainingArguments
raw_datasets = load_dataset("glue", "mrpc")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
def tokenize_function(example):
    return tokenizer(example["sentence1"], example["sentence2"], truncation=True)
tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

training_args = TrainingArguments("test-trainer")
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)
trainer = Trainer(
    model,
    training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"],
    data_collator=data_collator,
    tokenizer=tokenizer,
)
trainer.train()

raw_datasets = load_dataset("glue", "mrpc")

# data preprocessing
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
def tokenize_function(example):
    return tokenizer(example["sentence1"], example["sentence2"], truncation=True)
tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

tokenized_datasets = tokenized_datasets.remove_columns(["sentence1", "sentence2", "idx"])
tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
tokenized_datasets.set_format("torch")
## tokenized_datasets column_names: ["attention_mask", "input_ids", "labels", "token_type_ids"]
train_dataloader = DataLoader(
    tokenized_datasets["train"], shuffle=True, batch_size=8, collate_fn=data_collator
)
eval_dataloader = DataLoader(
    tokenized_datasets["validation"], batch_size=8, collate_fn=data_collator
)

model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)
optimizer = AdamW(model.parameters(), lr=3e-5)
## put our model and our batches on GPU
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
model.to(device)
num_epochs = 3
num_training_steps = num_epochs * len(train_dataloader)
lr_scheduler = get_scheduler(   #  the learning rate scheduler 
    "linear",
    optimizer=optimizer,
    num_warmup_steps=0,
    num_training_steps=num_training_steps,
)
progress_bar = tqdm(range(num_training_steps))

model.train()
for epoch in range(num_epochs):
    for batch in train_dataloader:
        batch = {k: v.to(device) for k, v in batch.items()}
        outputs = model(**batch)
        loss = outputs.loss
        loss.backward()

        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()
        progress_bar.update(1)

config = transformers.GPT2Config.from_pretrained(MODEL_NAME)
config.do_sample = config.task_specific_params['text-generation']['do_sample']
config.max_length = config.task_specific_params['text-generation']['max_length']
# print(config)
model = transformers.GPT2LMHeadModel.from_pretrained(MODEL_NAME, config=config)
# Tokenizer
tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)
# Tokenize the input
enc = tokenizer(['The elf queen'], return_tensors='pt')
print('enc =', enc)
print(tokenizer.batch_decode(enc['input_ids']))

input_ids = enc['input_ids']
attention_mask = torch.ones(input_ids.shape, dtype=torch.int64)
# predicts the next token at each position. 也就是 input_ids = [v1,v2,v3] 输出为 [v20,v30,v4]]。 v20 是根据v1 生成的下一个token，大概率跟v2 不一样，v4 是根据v1,v2,v3 生成的。
out = model(input_ids=input_ids, attention_mask=attention_mask)
logits = out['logits']
# -1 在python list 里表示最后一个元素。
new_id = logits[:, -1, :].argmax(dim=1)
print(new_id)
print(tokenizer.batch_decode(new_id))

input_ids = enc['input_ids']
for i in range(20):
    attention_mask = torch.ones(input_ids.shape, dtype=torch.int64)
    logits = model(input_ids=input_ids,attention_mask=attention_mask)['logits']                    
    new_id = logits[:, -1, :].argmax(dim=1)    # Generate new ID
    input_ids = torch.cat([input_ids, new_id.unsqueeze(0)], dim=1)  # input_ids 加入新生成的字符

class MyDset(torch.utils.data.Dataset):
    """A custom dataset that serves 1024-token blocks as input_ids == labels"""
    def __init__(self, data: list[list[int]]):
        self.data = []
        for d in data:
            input_ids = torch.tensor(d, dtype=torch.int64)
            attention_mask = torch.ones(len(d), dtype=torch.int64)
            self.data.append({'input_ids': input_ids, 'attention_mask': attention_mask, 'labels': input_ids})

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

    def __getitem__(self, idx: int):
        return self.data[idx]
def break_text_to_pieces(text_path: str, tokenizer: transformers.PreTrainedTokenizer, block_len: int = 512) -> list[str]:
    """Read a file and convert it to tokenized blocks, edding <|endoftext|> to each block"""
    with open(text_path) as f:
        text = f.read()
    chunk_len0 = block_len - 1  # Leave space for a TOKEN_ENDOFTEXT
    tokens = tokenizer.encode(text) # 原文本直接弄，够粗暴
    blocks = []
    pos = 0
    while pos < len(tokens):
        chunk = tokens[pos: pos + chunk_len0]
        chunk.append(TOKEN_ENDOFTEXT)
        blocks.append(chunk)
        pos += chunk_len0

    if len(blocks[-1]) < block_len:
        del blocks[-1]

    return blocks
def train_val_split(data: list[str], ratio: float):
    n = len(data)
    assert n >= 2
    n_val = max(1, int(n * ratio))
    return data[n_val:], data[:n_val]
def prepare_dsets(text_path: str, tokenizer: transformers.PreTrainedTokenizer, block_len: int):
    """Read the text, prepare the datasets """
    data = break_text_to_pieces(text_path, tokenizer, block_len)
    data_train, data_val = train_val_split(data, 0.2)
    return MyDset(data_train), MyDset(data_val)

# Load model and tokenizer
model = transformers.GPT2LMHeadModel.from_pretrained(MODEL_NAME)
tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)
training_args = transformers.TrainingArguments(output_dir="idiot_save/", learning_rate=1e-3,...)
# 传给trainer 的必须是预处理好的dataset（包含input_ids 等column）
trainer = transformers.Trainer(model=model,args=training_args,train_dataset=dset_train,eval_dataset=dset_val)
trainer.train()
# Save the model if needed
model.save_pretrained('./trained_model/')
tokenizer.save_pretrained('./trained_model/')
# Now our model is trained, try the generation
text = 'Natural language understanding comprises a wide range of diverse tasks'
batch = tokenizer([text], return_tensors='pt')
for k, v in batch.items():
    batch[k] = v.to(DEVICE)
out = model.generate(input_ids=batch['input_ids'], attention_mask=batch['attention_mask'], max_length=20)
print('GENERATION=', tokenizer.batch_decode(out.cpu()))