本文整理抱抱脸transformers库在模型训练过程中常见实用的python代码片段,并附带大模型在线推理预测python代码,干货满满。如果对你有帮助,还请点赞关注转发~
本文目录
基础配置
安装必备的python包
!pip install transformers datasets seqeval sacrebleu evaluate accelerate==0.19.0 sentencepiece loralib peft import transformers import torch import datasets import peft print("transformers vision: %s"%(transformers.__version__)) # transformers vision:4.29.2 print("torch vision: %s"%(torch.__version__)) # torch vision: 1.13.0 print(f"Accelerate version: {accelerate.__version__}") # Accelerate version: 0.19.0 print("datasets version: %s"%(datasets.__version__)) # datasets version: 2.1.0 print(f"PEFT version: {peft.__version__}") # peft version: 0.3.0
设置超参数&&随机数
import os import gc import ctypes import torch import random import numpy as np os.environ['WANDB_DISABLED'] = 'true' os.environ["TOKENIZERS_PARALLELISM"] = "false" # 释放不必要的显存和内存 def clean_memory(): gc.collect() ctypes.CDLL("libc.so.6").malloc_trim(0) torch.cuda.empty_cache() from huggingface_hub import login login(token= 'your_token') import wandb wandb.login(key= 'your_key') wandb.init(project="trl_imdb_positive") def seed_everything(seed): random.seed(seed) os.environ['PYTHONHASHSEED'] = str(seed) np.random.seed(seed) torch.manual_seed(seed) if torch.cuda.is_available(): torch.cuda.manual_seed(seed) torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = True torch.backends.cudnn.enabled = True # 自定义超参数 class Args: model_path = "hfl/chinese-bert-wwm-ext" max_seq_len = 128 ratio = 0.8 device = torch.device("cuda" if torch.cuda.is_available else "cpu") train_batch_size = 32 dev_batch_size = 32 weight_decay = 0.01 epochs = 1 learning_rate = 3e-5 eval_step = 100 prompt = "情感是[MASK][MASK]。" seed = 2024 args = Args() seed_everything(args.seed)
tokenizer使用方法汇总
基础用法
from transformers import BertForMaskedLM, BertTokenizer, BertForSequenceClassification, BertConfig, AdamW from transformers import pipeline tokenizer = BertTokenizer.from_pretrained(args.model_path) sentence = "It is a very beautiful book." tokens=tokenizer.tokenize(sentence) print(tokens) # ['it', 'is', 'a', 'very', 'beautiful', 'book', '.'] tokenizer.convert_tokens_to_ids(tokens) #[8233, 8310, 143, 11785, 13106, 9106, 119]
encode用法
token_samples_c=tokenizer.encode(text=sentence,add_special_tokens=True) token_samples_c # [101, 8233, 8310, 143, 11785, 13106, 9106, 119, 102] print(tokenizer.all_special_ids,tokenizer.all_special_tokens) #[100, 102, 0, 101, 103], ['[UNK]', '[SEP]', '[PAD]', '[CLS]', '[MASK]']
encode_plus用法
token_samples_d=tokenizer.encode_plus(text=sentence,max_length=15,return_tensors='pt',add_special_tokens=True, padding="max_length", truncation="longest_first", return_attention_mask=True, return_token_type_ids=True )#方式4 返回一个字典,包含id,type,mask,无须手动添加CLS与SEP,可以指定返回类型与长度,add_special_tokens默认为True token_samples_d # {'input_ids': tensor([[ 101, 8233, 8310, 143, 11785, 13106, 9106, 119, 102, 0, # 0, 0, 0, 0, 0]]), # 'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]), # 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0]])}
获得词对应的index
vocab = tokenizer.vocab#查看词典 label2ind = {} for label_name in label2id: zz=[vocab[label] for label in label_name] label2ind[label_name]=zz print(label2ind) # {'故事': [3125, 752], '文化': [3152, 1265], '娱乐': [2031, 727], '体育': [860, 5509], '财经': [6568, 5307], '房产': [2791, 772], # '汽车': [3749, 6756], '教育': [3136, 5509], '科技': [4906, 2825], '军事': [1092, 752], '旅游': [3180, 3952], '国际': [1744, 7354], # '股票': [5500, 4873], '三农': [676, 1093], '游戏': [3952, 2767]}
获得特殊tokens的id
print(tokenizer.unk_token, tokenizer.convert_tokens_to_ids(tokenizer.unk_token))# [UNK] 100 print(tokenizer.sep_token, tokenizer.convert_tokens_to_ids(tokenizer.sep_token))# [SEP] 102 print(tokenizer.pad_token, tokenizer.convert_tokens_to_ids(tokenizer.pad_token))# [PAD] 0 print(tokenizer.cls_token, tokenizer.convert_tokens_to_ids(tokenizer.cls_token))# [CLS] 101 print(tokenizer.mask_token, tokenizer.convert_tokens_to_ids(tokenizer.mask_token)) # [MASK] 103 print(tokenizer.all_special_ids)#打印special_ids # [100, 102, 0, 101, 103] print(tokenizer.all_special_tokens)#打印special_tokens # ['[UNK]', '[SEP]', '[PAD]', '[CLS]', '[MASK]']
datasets类用法介绍
处理本地和远程服务器上的数据集
## 处理本地数据 !wget https://github.com/crux82/squad-it/raw/master/SQuAD_it-train.json.gz !wget https://github.com/crux82/squad-it/raw/master/SQuAD_it-test.json.gz !gzip -dkv SQuAD_it-*.json.gz data_files = {"train": "SQuAD_it-train.json", "test": "SQuAD_it-test.json"} squad_it_dataset = load_dataset("json", data_files=data_files, field="data") squad_it_dataset ## 处理线上数据 url = "https://github.com/crux82/squad-it/raw/master/" data_files = { "train": url + "SQuAD_it-train.json.gz", "test": url + "SQuAD_it-test.json.gz", } squad_it_dataset = load_dataset("json", data_files=data_files, field="data") squad_it_dataset !wget "https://archive.ics.uci.edu/ml/machine-learning-databases/00462/drugsCom_raw.zip" !unzip drugsCom_raw.zip from datasets import load_dataset data_files = {"train": "drugsComTrain_raw.tsv", "test": "drugsComTest_raw.tsv"} # \t is the tab character in Python drug_dataset = load_dataset("csv", data_files=data_files, delimiter="\t") drug_dataset
基于dataset类常见的操作汇总
`## 查看明细 drug_sample = drug_dataset["train"].shuffle(seed=42).select(range(1000)) drug_sample[:3] ## 重命名列 drug_dataset = drug_dataset.rename_column( original_column_name="Unnamed: 0", new_column_name="patient_id") drug_dataset ## filter和map函数 def lowercase_condition(example): return {"condition": example["condition"].lower()} drug_dataset = drug_dataset.filter(lambda x: x["condition"] is not None) ## 过滤为空的值 drug_dataset = drug_dataset.map(lowercase_condition) ## 创建新列 def compute_review_length(example): return {"review_length": len(example["review"].split())} drug_dataset = drug_dataset.map(compute_review_length) ## sort函数 drug_dataset["train"].sort("review_length")[:3] drug_dataset = drug_dataset.filter(lambda x: x["review_length"] > 30) print(drug_dataset.num_rows, drug_dataset.num_columns) # DatasetDict({ # train: Dataset({ # features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount', 'review_length'], # num_rows: 138514 # }) # test: Dataset({ # features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount', 'review_length'], # num_rows: 46108 # }) # }) import html ## 性能加速 new_drug_dataset = drug_dataset.map(lambda x: {"review": [html.unescape(o) for o in x["review"]]}, batched=True) new_drug_dataset = drug_dataset.map(lambda x: {"review": [html.unescape(o) for o in x["review"]]}, batched=True, num_proc=4)) ## 移除多余的列 def tokenize_and_split(examples): return tokenizer( examples["review"], truncation=True, max_length=128, return_overflowing_tokens=True, ) tokenized_dataset = drug_dataset.map( tokenize_and_split, batched=True, remove_columns=drug_dataset["train"].column_names ) tokenized_dataset ## 数据集划分 drug_dataset_clean = drug_dataset["train"].train_test_split(train_size=0.8, seed=42) drug_dataset_clean["validation"] = drug_dataset_clean.pop("test") # Add the "test" set to our `DatasetDict` drug_dataset_clean["test"] = drug_dataset["test"] drug_dataset_clean `
保存数据集
drug_dataset_clean.save_to_disk("drug-reviews") ## .to_csv() , to_json() from datasets import load_from_disk drug_dataset_reloaded = load_from_disk("drug-reviews")
在huaggingface中创建自己的数据集通过代码上传
from huggingface_hub import list_datasets all_datasets = list_datasets() from huggingface_hub import login login(token=your_token) from huggingface_hub import create_repo repo_url = create_repo(name="github-issues", repo_type="dataset") from huggingface_hub import Repository repo = Repository(local_dir="github-issues", clone_from=repo_url) !cp issues-datasets-with-hf-doc-builder.jsonl github-issues/ ## issues-datasets-with-hf-doc-builder.jsonl 需要上传的数据 repo.lfs_track("*.jsonl") repo.push_to_hub()
加载huggingface的公共数据集
remote_dataset = load_dataset("lewtun/github-issues", split="train") remote_dataset ## 加载公共数据集 from huggingface_hub import hf_hub_url from datasets import load_dataset data_files = hf_hub_url( repo_id="lewtun/github-issues", filename="datasets-issues-with-hf-doc-builder.jsonl", repo_type="dataset", ) issues_dataset = load_dataset("json", data_files=data_files, split="train") ## 保持文本向量 def get_embeddings(text_list): encoded_input = tokenizer( text_list, padding=True, truncation=True, return_tensors="pt" ) encoded_input = {k: v.to(device) for k, v in encoded_input.items()} model_output = model(**encoded_input) return cls_pooling(model_output) embedding = get_embeddings(comments_dataset["text"][0]) embedding.shape ##torch.Size([1, 768]) embeddings_dataset = comments_dataset.map( lambda x: {"embeddings": get_embeddings(x["text"]).detach().cpu().numpy()[0]} ) embeddings_dataset.add_faiss_index(column="embeddings")
加载数据集
自定义函数加载数据集-加载函数
def load_data(data, prompt, max_seq_len): return_data = [] # [(文本, 标签id)] for d in data: text = d[0] label = d[1] text = "".join(text.split(" ")).strip() +"," +prompt if len(text) > max_seq_len - 2: continue return_data.append((text, label)) return return_data class Collate: def __init__(self, tokenizer, max_seq_len): self.tokenizer = tokenizer self.max_seq_len = max_seq_len def collate_fn(self, batch): input_ids_all = [] token_type_ids_all = [] attention_mask_all = [] label_all = [] mask_pos_all = [] for data in batch: text = data[0] label = data[1] inputs = self.tokenizer.encode_plus(text=text, max_length=self.max_seq_len, padding="max_length", truncation="longest_first", return_attention_mask=True, return_token_type_ids=True) input_ids = inputs["input_ids"] mask_pos = [i for i, token_id in enumerate(input_ids) if token_id == self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)] mask_pos_all.append(mask_pos) token_type_ids = inputs["token_type_ids"] attention_mask = inputs["attention_mask"] input_ids_all.append(input_ids) token_type_ids_all.append(token_type_ids) attention_mask_all.append(attention_mask) label_all.append(label) input_ids_all = torch.tensor(input_ids_all, dtype=torch.long) token_type_ids_all = torch.tensor(token_type_ids_all, dtype=torch.long) attention_mask_all = torch.tensor(attention_mask_all, dtype=torch.long) mask_pos_all = torch.tensor(mask_pos_all, dtype=torch.long) label_all = torch.tensor(label_all, dtype=torch.long) return_data = { "input_ids": input_ids_all, "attention_mask": attention_mask_all, "token_type_ids": token_type_ids_all, "label": label_all, "mask_pos": mask_pos_all, } return return_data
使用原生pytorch函数的DataLoder函数来加载数据
collate = Collate(tokenizer, args.max_seq_len) train_loader = DataLoader(train_data, batch_size=args.train_batch_size, shuffle=True, num_workers=2, collate_fn=collate.collate_fn) total_step = len(train_loader) * args.epochs args.total_step = total_step dev_loader = DataLoader(dev_data, batch_size=args.dev_batch_size, shuffle=False, num_workers=2, collate_fn=collate.collate_fn) test_loader = dev_loader for step, batch_data in enumerate(train_loader): label = batch_data["label"] batch_size = label.size(0) input_ids = batch_data["input_ids"] mask_pos = batch_data["mask_pos"] token_type_ids = batch_data["token_type_ids"] attention_mask = batch_data["attention_mask"] print(input_ids[:2]) print(mask_pos[:2]) break # tensor([[ 101, 6206, 5440, 6407, 1044, 4312, 3614, 8043, 5739, 1744, 1187, 3441, # 8188, 2399, 837, 5320, 8024, 2110, 4495, 812, 9126, 8024, 6662, 782, # 679, 4007, 117, 712, 7579, 3221, 103, 103, 511, 102, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0], # [ 101, 872, 6371, 711, 686, 4518, 677, 1525, 702, 1062, 1385, 3297, # 1326, 2154, 8043, 117, 712, 7579, 3221, 103, 103, 511, 102, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, # 0, 0, 0, 0, 0, 0, 0, 0]]) # tensor([[30, 31], # [19, 20]])
加载通用数据集
from sklearn.model_selection import StratifiedKFold from sklearn.model_selection import train_test_split from datasets import Dataset, DatasetDict test = Dataset.from_pandas(test) disaster_tweets_test = DatasetDict() disaster_tweets_test['test'] = test data_y = train['labels'] data_x = train.drop(columns='labels') def get_train_val_test(df_train,df_val): tds = Dataset.from_pandas(df_train) vds = Dataset.from_pandas(df_val) disaster_tweets = DatasetDict() disaster_tweets['train'] = tds disaster_tweets['validation'] = vds return disaster_tweets from transformers import AutoTokenizer from transformers import ElectraTokenizer, ElectraForSequenceClassification,AdamW import torch max_length = 40 # 98(key+local+text) # model_ckpt ='bert-large-uncased'#'microsoft/deberta-v3-large' #'microsoft/deberta-v3-base'#'microsoft/deberta-v3-large'# 'microsoft/deberta-v3-base'#'xlm-roberta-base' tokenizer = AutoTokenizer.from_pretrained(model_ckpt)# 'xlm-roberta-base') # tokenizer = ElectraTokenizer.from_pretrained('google/electra-base-discriminator') def tokenize(batch): return tokenizer(batch['text'], max_length=max_length,padding=True, truncation=True) #batch['keyword'], disaster_tweets_test_encoded = disaster_tweets_test.map(tokenize, batched=True, batch_size=None) from sklearn.model_selection import StratifiedKFold Fold = StratifiedKFold(n_splits=n_fold, shuffle=True, random_state=seed_val) for n, (train_index, val_index) in enumerate(Fold.split(data_x, data_y)): train_pf = data_x.iloc[train_index,:] train_pf['label'] = data_y.iloc[train_index] val_pf = data_x.iloc[val_index,:] val_pf['label'] = data_y.iloc[val_index] disaster_tweets= get_train_val_test(train_pf,val_pf) disaster_tweets_encoded = disaster_tweets.map(tokenize, batched=True, batch_size=None)
搭建网络模型
搭建网络层
class multilabel_dropout(nn.Module): # Multisample Dropout 论文: https://arxiv.org/abs/1905.09788 def __init__(self, hidden_size, num_labels=2): super(multilabel_dropout, self).__init__() self.classifier = torch.nn.Linear(hidden_size, num_labels) def forward(self, out): return torch.mean(torch.stack([ self.classifier( torch.nn.Dropout(p)(out)) for p in np.linspace(0.1,0.5, 5)], dim=0) , dim=0) class MeanPooling(nn.Module): def __init__(self): super(MeanPooling, self).__init__() def forward(self, last_hidden_state, attention_mask): input_mask_expanded = attention_mask.unsqueeze(-1).expand(last_hidden_state.size()).float() sum_embeddings = torch.sum(last_hidden_state * input_mask_expanded, 1) sum_mask = input_mask_expanded.sum(1) sum_mask = torch.clamp(sum_mask, min=1e-9) mean_embeddings = sum_embeddings / sum_mask return mean_embeddings
搭建自定义的模型
from transformers import AutoModelForSequenceClassification,BertForSequenceClassification,AutoModel,AutoConfig from transformers.modeling_outputs import TokenClassifierOutput num_labels = len(class_names) class MyModel(nn.Module): def __init__(self, model_name,num_labels): super(MyModel, self).__init__() self.num_labels=num_labels self.model = AutoModel.from_pretrained(model_name) self.config = AutoConfig.from_pretrained(model_name) self.drop = nn.Dropout(p=0.2) self.pooler = MeanPooling() self.fc = nn.Linear(self.config.hidden_size,num_labels) self.multi_drop=multilabel_dropout(self.config.hidden_size,self.num_labels) def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, labels=None): outputs = self.model(input_ids=input_ids, attention_mask=attention_mask,output_hidden_states=False) out = self.pooler(outputs.last_hidden_state, attention_mask) #outputs.last_hidden_state[:,0,:] # # out = self.drop(out) # logits = self.fc(out) logits= self.multi_drop(out) loss = None if labels is not None: loss_fct = nn.CrossEntropyLoss() loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1)) return TokenClassifierOutput(loss=loss, logits=logits, hidden_states=outputs.hidden_states,attentions=outputs.attentions) model = MyModel(model_name=model_ckpt,num_labels=num_labels).to(device)
搭建huggingface上通用模型
from transformers import AutoModelForSequenceClassification,BertForSequenceClassification,AutoModel,AutoConfig from transformers.modeling_outputs import TokenClassifierOutput num_labels = len(class_names) model = AutoModelForSequenceClassification\ .from_pretrained(model_ckpt, num_labels=num_labels)\ .to(device) model = BertForSequenceClassification.from_pretrained(model_ckpt, num_labels=2) model= model.to(device)
上传模型到仓库
## 下载模型权重 !sudo apt-get install git-lfs !git lfs install !git clone https://huggingface.co/THUDM/chatglm2-6b.git !GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/THUDM/chatglm2-6b # 下载非lfs文件 !ls {save_path} #使用python变量; save_path='demo.txt' ## 从国内下载模型 huggingface镜像 !pip install -U huggingface_hub !git clone https://github.com/LetheSec/HuggingFace-Download-Accelerator.git %cd HuggingFace-Download-Accelerator !python hf_download.py --model mistralai/Mixtral-8x7B-Instruct-v0.1 --save_dir ../hf_hub from huggingface_hub import snapshot_download snapshot_download(repo_id="baichuan-inc/Baichuan2-13B-Chat-4bits", local_dir="baichuan-inc/Baichuan2-13B-Chat-4bits") """ ## 上传大模型到huggingface中 1. git clone https://huggingface.co/username/model_name 2. git lfs install 3. 将待上传的文件移动到该目录下 4. huggingface-cli lfs-enable-largefiles ./ ## 超过5GB文件 5. git add . 6. git commit -m "commit from $USER" 7. git push """ ## 利用python代码上传大模型到huggingface中 save_path = "your_finetune_model_weight" from huggingface_hub import login login() #需要注册一个huggingface账户,在个人页面setting那里创建一个有write权限的access token from huggingface_hub import HfApi api = HfApi() #创建huggingface 模型库 repo_id = "shujunge/chatglm2_6b_helloword" api.create_repo(repo_id=repo_id) #上传模型可能需要等待10分钟左右~ api.upload_folder( folder_path=save_path, repo_id=repo_id, repo_type="model", #space, model, datasets ) #上传成功后可以删除本地模型 !rm -rf {save_path}
模型训练和配置
设置优化器
from transformers import BertForMaskedLM, BertTokenizer, BertForSequenceClassification, BertConfig, AdamW # 设置不同的学习率 def build_optimizer(self): no_decay = ['bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [ {'params': [p for n, p in self.model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': self.args.weight_decay}, {'params': [p for n, p in self.model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0} ] # optimizer = AdamW(model.parameters(), lr=learning_rate) optimizer = AdamW(optimizer_grouped_parameters, lr=self.args.learning_rate) return optimizer ## 设置不同的学习器 from transformers.optimization import Adafactor, AdafactorSchedule optimizer = Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True, lr=None) lr_scheduler = AdafactorSchedule(optimizer) trainer = Trainer(..., optimizers=(optimizer, lr_scheduler))
定义指标函数
import evaluate predictions = trainer.predict(tokenized_datasets["validation"]) print(predictions.predictions.shape, predictions.label_ids.shape) (408, 2) (408,) preds = np.argmax(predictions.predictions, axis=-1) metric = evaluate.load("glue", "mrpc") metric.compute(predictions=preds, references=predictions.label_ids) # {'accuracy': 0.8578431372549019, 'f1': 0.8996539792387542}
使用sklearn库搭建验证指标函数
from sklearn.metrics import accuracy_score, f1_score def compute_metrics(pred) -> dict: labels = pred.label_ids preds = pred.predictions.argmax(-1) f1 = f1_score(labels, preds, average='weighted') acc = accuracy_score(labels, preds) return {'accuracy': acc, 'f1': f1} def compute_metrics(eval_preds): metric = evaluate.load("glue", "mrpc") logits, labels = eval_preds predictions = np.argmax(logits, axis=-1) return metric.compute(predictions=predictions, references=labels) from datasets import load_metric def compute_metrics(eval_pred): metrics = ["accuracy", "recall", "precision", "f1"] #List of metrics to return metric={} for met in metrics: metric[met] = load_metric(met) logits, labels = eval_pred predictions = np.argmax(logits, axis=-1) metric_res={} for met in metrics: metric_res[met]=metric[met].compute(predictions=predictions, references=labels)[met] return metric_res
使用evaluate库搭建验证指标函数
import numpy as np import evaluate metric = evaluate.load("seqeval") def compute_metrics(eval_preds): logits, labels = eval_preds predictions = np.argmax(logits, axis=-1) # Remove ignored index (special tokens) and convert to labels true_labels = [[label_names[l] for l in label if l != -100] for label in labels] true_predictions = [ [label_names[p] for (p, l) in zip(prediction, label) if l != -100] for prediction, label in zip(predictions, labels) ] all_metrics = metric.compute(predictions=true_predictions, references=true_labels) return { "precision": all_metrics["overall_precision"], "recall": all_metrics["overall_recall"], "f1": all_metrics["overall_f1"], "accuracy": all_metrics["overall_accuracy"], }
自定义损失函数
class RegressionTrainer(Trainer): def compute_loss(self, model, inputs, return_outputs=False): labels = inputs.get("labels") outputs = model(**inputs) logits = outputs.get('logits') loss = torch.mean(torch.square(logits.squeeze() - labels.squeeze())) return (loss, outputs) if return_outputs else loss class CustomTrainer(Trainer): def compute_loss(self, model, inputs, return_outputs=False): labels = inputs.get("labels") outputs = model(**inputs) # forward pass logits = outputs.get('logits') # compute custom loss # Class weighting loss_fct = nn.CrossEntropyLoss() loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1)) return (loss, outputs) if return_outputs else loss def training_step(self, model: nn.Module, inputs: Dict[str, Union[torch.Tensor, Any]]) -> torch.Tensor: model.train() inputs = self._prepare_inputs(inputs) fgm = FGM(model,epsilon=1,emb_name='word_embeddings.') with self.autocast_smart_context_manager(): loss = self.compute_loss(model, inputs) loss = loss / self.args.gradient_accumulation_steps loss=self.scaler.scale(loss) ##半精度用torch自带的就行了 loss.backward() fgm.attack() # 在embedding上添加对抗扰动 loss_adv = self.compute_loss(model, inputs) loss_adv.backward() fgm.restore() return loss_adv.detach()
继承Trainer类显示学习率
class MyTrainer(Trainer): def log(self, logs): logs["learning_rate"] = self._get_learning_rate() super().log(logs)
配置Trainer类相关训练模型超参数
## 配置训练参数 logging_steps = len(disaster_tweets_encoded['train']) // batch_size model_name = f"{model_ckpt}-finetuned-disaster" training_args = TrainingArguments( seed=seed_val, report_to='none', output_dir=model_name, learning_rate=3e-5, per_device_train_batch_size=batch_size, per_device_eval_batch_size=batch_size, weight_decay=.01, disable_tqdm=False, logging_steps=logging_steps, log_level='error', load_best_model_at_end=True, save_total_limit = 1, fp16=True, # 按照epoch设置 num_train_epochs= 2, save_strategy="epoch", ## 与evaluation_strategy保存一致 evaluation_strategy='epoch', # 按照step设置训练参数 evaluation_strategy='steps', save_strategy= 'steps', save_steps=logging_steps, eval_steps=logging_steps, max_steps=logging_steps*n_epochs, # metric_for_best_model='f1', ##设置监控指标 如果这个设置了值,对于的greater_is_better则为True # greater_is_better=True, ## 保存最大或者最小 ##默认为False(值越小越好) push_to_hub=False, )
模型训练
trainer = MyTrainer( model=model, args=training_args, compute_metrics=compute_metrics, train_dataset=disaster_tweets_encoded['train'], eval_dataset=disaster_tweets_encoded['validation'], tokenizer=tokenizer, callbacks=[early_stop, mlc], ) trainer.train() #模型训练 trainer.evaluate() #模型验证 #模型预测 # trainer.save_model(f"{n}_fold_weights") proba_prediction_test = trainer.predict(disaster_tweets_test_encoded['test'])
模型的权重保存
self.model = BertForMaskedLM.from_pretrained(args.model_path) torch.save(self.model.state_dict(), "bert_prompt.pt") trainer = Trainer(args) model = BertForMaskedLM.from_pretrained(args.model_path) model.load_state_dict(torch.load(ckpt_path)) model.to(args.device) trainer.save_model("my_weight") tokenizer.save_pretrained('my_weight') trainer.push_to_hub(commit_message="Training complete", tags="summarization") ## 推送远程仓库
模型验证和指标计算
from sklearn.metrics import classification_report results = trainer.predict(disaster_tweets_test_encoded['test'])['predictions'] report=classification_report(test['labels'], np.argmax(results, axis=-1).tolist(), target_names=class_names)
添加记录训练曲线函数
from transformers import TrainerCallback class CustomCallback(TrainerCallback): def __init__(self, trainer) -> None: super().__init__() self._trainer = trainer def on_epoch_end(self, args, state, control, **kwargs): if control.should_evaluate: control_copy = deepcopy(control) self._trainer.evaluate(eval_dataset=self._trainer.train_dataset, metric_key_prefix="train") return control_copy trainer = Trainer( model=model, # the instantiated Transformers model to be trained args=training_args, # training arguments, defined above train_dataset=train_dataset, # training dataset eval_dataset=valid_dataset, # evaluation dataset compute_metrics=compute_metrics, # the callback that computes metrics of interest tokenizer=tokenizer ) trainer.add_callback(CustomCallback(trainer)) train = trainer.train() {‘train_loss’: 0.7159061431884766, ‘train_accuracy’: 0.4, ‘train_f1’: 0.5714285714285715, ‘train_runtime’: 6.2973, ‘train_samples_per_second’: 2.382, ‘train_steps_per_second’: 0.159, ‘epoch’: 1.0} {‘eval_loss’: 0.8529007434844971, ‘eval_accuracy’: 0.0, ‘eval_f1’: 0.0, ‘eval_runtime’: 2.0739, ‘eval_samples_per_second’: 0.964, ‘eval_steps_per_second’: 0.482, ‘epoch’: 1.0}
在线部署大模型进行部署推理
部署chatglm3_6b模型进行16位推理预测
环境配置
from IPython.display import clear_output !pip install transformers>=4.37.0 !pip install -q peft !pip install -q accelerate !pip install -q bitsandbytes import torch import transformers import peft import accelerate print("torch:",torch.__version__) #torch: 2.1.2 print("transformers:",transformers.__version__) #transformers: 4.37.0 print("peft:",peft.__version__) #peft: 0.8.2 print("accelerate:",accelerate.__version__) #accelerate: 0.26.1 !pip show bitsandbytes #Version: 0.42.0 # 验证是否支持nf4 from bitsandbytes.cuda_setup.main import CUDASetup setup = CUDASetup.get_instance() if setup.initialized != True: setup.run_cuda_setup() lib = setup.lib lib.cquantize_blockwise_fp16_nf4
在线编写代码部署chatglm3_6B实时推理
下面我将给大家利用transformers框架进行实操部署chatglm3以半精度(float16)进行在线推理。显卡采用kaggle:P100,16G;
#下载chatglm3-6b权重 from huggingface_hub import snapshot_download snapshot_download(repo_id="THUDM/chatglm3-6b", local_dir='THUDM/chatglm3-6b') import gc import ctypes import torch import random from transformers import AutoTokenizer,AutoConfig, AutoModel, BitsAndBytesConfig import torch # 释放不必要的显存和内存 def clean_memory(): gc.collect() ctypes.CDLL("libc.so.6").malloc_trim(0) torch.cuda.empty_cache() from transformers import AutoModelForCausalLM, AutoTokenizer model_name_or_path = '/kaggle/working/THUDM/chatglm3-6b' tokenizer = AutoTokenizer.from_pretrained( model_name_or_path, torch_dtype= torch.float16, trust_remote_code=True) model = AutoModel.from_pretrained(model_name_or_path,trust_remote_code=True).half().cuda() !nvidia-smi
```python %%time response,history= model.chat(tokenizer,query= '你好。你是谁,你能干什么',history=[]) print(response)
输出结果:
你好,我是 ChatGLM3-6B,是清华大学 KEG 实验室和智谱 AI 公司于 2023 年共同训练的语言模型。我的任务是针对用户的问题和要求提供适当的答复和支持。我能够回答各种问题,包括但不限于科学、技术、历史、文化、娱乐等方面。同时,我也能够提供一些基本的服务,如智能推荐、语言翻译、代码调试等。 CPU times: user 4.05 s, sys: 0 ns, total: 4.05 s Wall time: 4.05 s
对qwen1.5_7B模型进行4位推理预测
下面是编写python代码进行加载qwen1.5_7B大模型的按照4位精度加载推理案例。
from IPython.display import clear_output # 切记版本要升级在4.37.0以上 !pip install transformers>=4.37.0 !pip install -q peft !pip install -q accelerate !pip install -q bitsandbytes import torch import transformers import peft import accelerate print("torch:",torch.__version__) #torch: 2.1.2 print("transformers:",transformers.__version__) #transformers: 4.37.0 print("peft:",peft.__version__) #peft: 0.8.2 print("accelerate:",accelerate.__version__) #accelerate: 0.26.1 !pip show bitsandbytes #Version: 0.42.0 # 搭建模型 from transformers import AutoModelForCausalLM, AutoTokenizer,BitsAndBytesConfig import torch device = "cuda" # the device to load the model onto bnb_config=BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16, bnb_4bit_use_double_quant=True, #QLoRA 设计的 Double Quantization bnb_4bit_quant_type="nf4", #QLoRA 设计的 Normal Float 4 量化数据类型 llm_int8_threshold=6.0, llm_int8_has_fp16_weight=False, ) """ 注意:Qwen1.5 最大的不同之处,在于 Qwen1.5 与 HuggingFace transformers 代码库的集成。从 4.37.0 版本开始,您可以直接使用 transformers 库原生代码,而不加载任何自定义代码(即指定trust_remote_code=True)来使用 Qwen1.5 """ model = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen1.5-7B-Chat", load_in_4bit=True, torch_dtype= torch.float16, quantization_config=bnb_config, ) tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-7B-Chat") model
%%time prompt = "什么是大模型?" messages = [ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": prompt} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) model_inputs = tokenizer([text], return_tensors="pt").to(device) gen_kwargs = { "max_new_tokens": 1024, "num_beams": 1, "do_sample": True, "top_p": 0.8, "temperature": 0.01, "top_k": 50, 'repetition_penalty':1 } generated_ids = model.generate( model_inputs.input_ids, **gen_kwargs ) generated_ids = [ output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) ] response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] print(response)
大模型通常指的是在人工智能领域中训练规模较大、参数较多的模型,这些模型通常具有更高的计算资源需求和训练时间,但也可以学习到更复杂的模式和知识,从而在各种任务上表现出更强的性能。 大模型可以指的是一系列模型,例如在自然语言处理领域,BERT、GPT-2、GPT-3等都是大型预训练模型,它们的参数量通常在数十亿到万亿级别。这些模型在训练时使用大量的未标注数据,然后通过微调(fine-tuning)的方式在特定任务上进行优化,以达到更好的效果。 大模型的优点是可以处理更复杂的任务和大规模的数据,但同时也带来了挑战,如需要大量的计算资源、训练时间长、过拟合等问题,以及可能的隐私和伦理问题。 CPU times: user 12.7 s, sys: 0 ns, total: 12.7 s Wall time: 12.7 s
qwen1.5_7B 4位精度显存占用情况
这是qwen1.5_7B模型在4位精度下显存占用情况。
!nvidia-smi
+---------------------------------------------------------------------------------------+ | NVIDIA-SMI 535.129.03 Driver Version: 535.129.03 CUDA Version: 12.2 | |-----------------------------------------+----------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+======================+======================| | 0 Tesla P100-PCIE-16GB Off | 00000000:00:04.0 Off | 0 | | N/A 40C P0 37W / 250W | 6892MiB / 16384MiB | 0% Default | | | | N/A | +-----------------------------------------+----------------------+----------------------+ +---------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=======================================================================================| +---------------------------------------------------------------------------------------+
对大模型进行监督指令微调SFT
====
今天给大家介绍大模型训练框架transformers库常见实用的python代码,具体函数详情可参考huggingface官网。大家可在实际模型训练过程中选择合适的方法来提供自己的效率。如果本文对你有帮助,还请你点赞在看转发。你的支持就是我创作的最大动力,关注下面公众号不迷路~
那么,如何系统的去学习大模型LLM?
我在一线互联网企业工作十余年里,指导过不少同行后辈。帮助很多人得到了学习和成长。
作为一名热心肠的互联网老兵,我意识到有很多经验和知识值得分享给大家,也可以通过我们的能力和经验解答大家在人工智能学习中的很多困惑,所以在工作繁忙的情况下还是坚持各种整理和分享。
但苦于知识传播途径有限,很多互联网行业朋友无法获得正确的资料得到学习提升,故此将并将重要的AI大模型资料包括AI大模型入门学习思维导图、精品AI大模型学习书籍手册、视频教程、实战学习等录播视频免费分享出来。
所有资料 ⚡️ ,朋友们如果有需要全套 《LLM大模型入门+进阶学习资源包》,扫码获取~
篇幅有限,部分资料如下:
👉LLM大模型学习指南+路线汇总👈
💥大模型入门要点,扫盲必看!
💥既然要系统的学习大模型,那么学习路线是必不可少的,这份路线能帮助你快速梳理知识,形成自己的体系。
路线图很大就不一一展示了 (文末领取)
👉大模型入门实战训练👈
💥光学理论是没用的,要学会跟着一起做,要动手实操,才能将自己的所学运用到实际当中去,这时候可以搞点实战案例来学习。
👉国内企业大模型落地应用案例👈
💥两本《中国大模型落地应用案例集》 收录了近两年151个优秀的大模型落地应用案例,这些案例覆盖了金融、医疗、教育、交通、制造等众多领域,无论是对于大模型技术的研究者,还是对于希望了解大模型技术在实际业务中如何应用的业内人士,都具有很高的参考价值。 (文末领取)
👉GitHub海量高星开源项目👈
💥收集整理了海量的开源项目,地址、代码、文档等等全都下载共享给大家一起学习!
👉LLM大模型学习视频👈
💥观看零基础学习书籍和视频,看书籍和视频学习是最快捷也是最有效果的方式,跟着视频中老师的思路,从基础到深入,还是很容易入门的。 (文末领取)
👉640份大模型行业报告(持续更新)👈
💥包含640份报告的合集,涵盖了AI大模型的理论研究、技术实现、行业应用等多个方面。无论您是科研人员、工程师,还是对AI大模型感兴趣的爱好者,这套报告合集都将为您提供宝贵的信息和启示。
👉获取方式:
这份完整版的大模型 LLM 学习资料已经上传CSDN,朋友们如果需要可以微信扫描下方CSDN官方认证二维码免费领取【保证100%免费
】
😝有需要的小伙伴,可以Vx扫描下方二维码免费领取🆓