P-Tuning¶
1.P-Tuning简述¶
P-Tuning(论文:GPT Understands, Too),该方法将 Prompt 转换为可以学习的 Embedding 层,并用MLP+LSTM的方式来对Prompt Embedding进行一层处理。
相比Prefix Tuning,P-Tuning加入的可微的virtual token,但仅限于输入层,没有在每一层都加;另外,virtual token的位置也不一定是前缀,插入的位置是可选的。这里的出发点实际是把传统人工设计模版中的真实token替换成可微的virtual token。
经过预训练的LM的词嵌入已经变得高度离散,如果随机初始化virtual token,容易优化到局部最优值,而这些virtual token理论是应该有相关关联的。因此,作者通过实验发现用一个提示编码器(即用一个LSTM+MLP去编码这些virtual token以后,再输入到模型)来编码会收敛更快,效果更好。
2.微调实战¶
2.1 引入库¶
from transformers import AutoModelForCausalLM
from peft import (
get_peft_config,
get_peft_model,
get_peft_model_state_dict,
set_peft_model_state_dict,
PeftType,
TaskType,
PromptEncoderConfig,
)
import torch
from datasets import load_dataset
import os
from transformers import AutoTokenizer
from torch.utils.data import DataLoader
from transformers import default_data_collator, get_linear_schedule_with_warmup
from tqdm import tqdm
from datasets import load_dataset
2.2 创建 P-Tuning 微调方法对应的配置¶
P-tuning 使用提示编码器(PromptEncoder)来优化提示参数,因此,需要使用如下几个参数初始化 PromptEncoderConfig:
task_type
:训练的任务类型,如:序列分类(SEQ_CLS),因果语言建模(CAUSAL_LM)等。num_virtual_tokens
:虚拟token的数量,换句话说就是提示(prompt)。encoder_hidden_size
:编码器的隐藏大小,用于优化提示参数。encoder_reparameterization_type
:指定如何重新参数化提示编码器,可选项有:MLP 或 LSTM,默认值为 MLP。
当使用 LSTM 时, 提示编码器结构如下:
(prompt_encoder): ModuleDict(
(default): PromptEncoder(
(embedding): Embedding(20, 1024)
(lstm_head): LSTM(1024, 128, num_layers=2, batch_first=True, bidirectional=True)
(mlp_head): Sequential(
(0): Linear(in_features=256, out_features=256, bias=True)
(1): ReLU()
(2): Linear(in_features=256, out_features=1024, bias=True)
)
)
)
当使用 MLP 时, 提示编码器结构如下:
(prompt_encoder): ModuleDict(
(default): PromptEncoder(
(embedding): Embedding(20, 1024)
(mlp_head): Sequential(
(0): Linear(in_features=1024, out_features=128, bias=True)
(1): ReLU()
(2): Linear(in_features=128, out_features=128, bias=True)
(3): ReLU()
(4): Linear(in_features=128, out_features=1024, bias=True)
)
)
)
PEFT 中的 P-tuning 的提示编码器是基于英伟达的NeMo库中 prompt_encoder.py 进行的重构,源码如下所示。
class PromptEncoder(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.token_dim = config.token_dim
self.input_size = self.token_dim
self.output_size = self.token_dim
self.hidden_size = config.encoder_hidden_size
self.total_virtual_tokens = config.num_virtual_tokens * config.num_transformer_submodules
self.encoder_type = config.encoder_reparameterization_type
# 初始化 embedding 层
self.embedding = torch.nn.Embedding(self.total_virtual_tokens, self.token_dim)
if not config.inference_mode:
# 根据PromptEncoder重参数化类型初始化相应的lstm和mlp
if self.encoder_type == PromptEncoderReparameterizationType.LSTM:
lstm_dropout = config.encoder_dropout
num_layers = config.encoder_num_layers
# LSTM
self.lstm_head = torch.nn.LSTM(
input_size=self.input_size,
hidden_size=self.hidden_size,
num_layers=num_layers,
dropout=lstm_dropout,
bidirectional=True,
batch_first=True,
)
self.mlp_head = torch.nn.Sequential(
torch.nn.Linear(self.hidden_size * 2, self.hidden_size * 2),
torch.nn.ReLU(),
torch.nn.Linear(self.hidden_size * 2, self.output_size),
)
elif self.encoder_type == PromptEncoderReparameterizationType.MLP:
warnings.warn(
f"for {self.encoder_type}, the `encoder_num_layers` is ignored. Exactly 2 MLP layers are used."
)
layers = [
torch.nn.Linear(self.input_size, self.hidden_size),
torch.nn.ReLU(),
torch.nn.Linear(self.hidden_size, self.hidden_size),
torch.nn.ReLU(),
torch.nn.Linear(self.hidden_size, self.output_size),
]
self.mlp_head = torch.nn.Sequential(*layers)
else:
raise ValueError("Prompt encoder type not recognized. Please use one of MLP (recommended) or LSTM.")
def forward(self, indices):
input_embeds = self.embedding(indices)
if self.encoder_type == PromptEncoderReparameterizationType.LSTM:
output_embeds = self.mlp_head(self.lstm_head(input_embeds)[0])
elif self.encoder_type == PromptEncoderReparameterizationType.MLP:
output_embeds = self.mlp_head(input_embeds)
else:
raise ValueError("Prompt encoder type not recognized. Please use one of MLP (recommended) or LSTM.")
return output_embeds
device = "cuda"
model_name_or_path = "/data/nfs/llm/model/bloomz-560m"
tokenizer_name_or_path = "/data/nfs/llm/model/bloomz-560m"
# P-Tuning 配置类 PromptEncoderConfig
peft_config = PromptEncoderConfig(
task_type=TaskType.CAUSAL_LM,
num_virtual_tokens=20,
encoder_hidden_size=128
)
dataset_name = "twitter_complaints"
checkpoint_name = f"{dataset_name}_{model_name_or_path}_{peft_config.peft_type}_{peft_config.task_type}_v1.pt".replace("/", "_")
text_column = "Tweet text"
label_column = "text_label"
max_length = 64
lr = 3e-2
num_epochs = 10
batch_size = 8
/home/guodong.li/virtual-venv/peft-venv-py310-cu117/lib/python3.10/site-packages/tqdm/auto.py:21:TqdmWarning:IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html from .autonotebook import tqdm as notebook_tqdm
[2023-07-19 19:02:06,848] [INFO] [real_accelerator.py:133:get_accelerator] Setting ds_accelerator to cuda (auto detect)
from datasets import load_dataset
dataset = load_dataset("ought/raft", dataset_name)
# dataset = load_dataset("/home/guodong.li/data/peft/raft/raft.py", dataset_name, cache_dir="/home/guodong.li/data/peft/data")
classes = [k.replace("_", " ") for k in dataset["train"].features["Label"].names]
print(classes)
dataset = dataset.map(
lambda x: {"text_label": [classes[label] for label in x["Label"]]},
batched=True,
num_proc=1,
)
print(dataset)
dataset["train"][0]
Found cached dataset raft (/home/guodong.li/data/peft/data/raft/twitter_complaints/1.1.0/79c4de1312c1e3730043f7db07179c914f48403101f7124e2fe336f6f54d9f84) 100%|██████████| 2/2 [00:00<00:00, 450.13it/s] Loading cached processed dataset at /home/guodong.li/data/peft/data/raft/twitter_complaints/1.1.0/79c4de1312c1e3730043f7db07179c914f48403101f7124e2fe336f6f54d9f84/cache-0e20fff6b1d898ca.arrow Loading cached processed dataset at /home/guodong.li/data/peft/data/raft/twitter_complaints/1.1.0/79c4de1312c1e3730043f7db07179c914f48403101f7124e2fe336f6f54d9f84/cache-8d14a62b8a688c19.arrow
['Unlabeled', 'complaint', 'no complaint'] DatasetDict({ train:Dataset({ features:['Tweet text', 'ID', 'Label', 'text_label'], num_rows:50 }) test:Dataset({ features:['Tweet text', 'ID', 'Label', 'text_label'], num_rows:3399 }) })
{'Tweet text':'@HMRCcustomers No this is my first job', 'ID':0, 'Label':2, 'text_label':'no complaint'}
# data preprocessing
# padding_side = "left"
# tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, padding_side=padding_side)
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
if tokenizer.pad_token_id is None:
tokenizer.pad_token_id = tokenizer.eos_token_id
target_max_length = max([len(tokenizer(class_label)["input_ids"]) for class_label in classes])
print("target_max_length:", target_max_length)
def preprocess_function(examples):
batch_size = len(examples[text_column])
inputs = [f"{text_column} :{x} Label :" for x in examples[text_column]]
targets = [str(x) for x in examples[label_column]]
model_inputs = tokenizer(inputs)
labels = tokenizer(targets)
for i in range(batch_size):
sample_input_ids = model_inputs["input_ids"][i]
label_input_ids = labels["input_ids"][i] + [tokenizer.pad_token_id]
# print(i, sample_input_ids, label_input_ids)
model_inputs["input_ids"][i] = sample_input_ids + label_input_ids
labels["input_ids"][i] = [-100] * len(sample_input_ids) + label_input_ids
model_inputs["attention_mask"][i] = [1] * len(model_inputs["input_ids"][i])
# print(model_inputs)
for i in range(batch_size):
sample_input_ids = model_inputs["input_ids"][i]
label_input_ids = labels["input_ids"][i]
model_inputs["input_ids"][i] = [tokenizer.pad_token_id] * (
max_length - len(sample_input_ids)
) + sample_input_ids
model_inputs["attention_mask"][i] = [0] * (max_length - len(sample_input_ids)) + model_inputs[
"attention_mask"
][i]
labels["input_ids"][i] = [-100] * (max_length - len(sample_input_ids)) + label_input_ids
model_inputs["input_ids"][i] = torch.tensor(model_inputs["input_ids"][i][:max_length])
model_inputs["attention_mask"][i] = torch.tensor(model_inputs["attention_mask"][i][:max_length])
labels["input_ids"][i] = torch.tensor(labels["input_ids"][i][:max_length])
model_inputs["labels"] = labels["input_ids"]
return model_inputs
processed_datasets = dataset.map(
preprocess_function,
batched=True,
num_proc=1,
remove_columns=dataset["train"].column_names,
load_from_cache_file=False,
desc="Running tokenizer on dataset",
)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["train"]
train_dataloader = DataLoader(train_dataset, shuffle=True, collate_fn=default_data_collator, batch_size=batch_size, pin_memory=True)
eval_dataloader = DataLoader(eval_dataset, collate_fn=default_data_collator, batch_size=batch_size, pin_memory=True)
target_max_length:3
def test_preprocess_function(examples):
batch_size = len(examples[text_column])
inputs = [f"{text_column} :{x} Label :" for x in examples[text_column]]
model_inputs = tokenizer(inputs)
# print(model_inputs)
for i in range(batch_size):
sample_input_ids = model_inputs["input_ids"][i]
model_inputs["input_ids"][i] = [tokenizer.pad_token_id] * (max_length - len(sample_input_ids)) + sample_input_ids
model_inputs["attention_mask"][i] = [0] * (max_length - len(sample_input_ids)) + model_inputs["attention_mask"][i]
model_inputs["input_ids"][i] = torch.tensor(model_inputs["input_ids"][i][:max_length])
model_inputs["attention_mask"][i] = torch.tensor(model_inputs["attention_mask"][i][:max_length])
return model_inputs
test_dataset = dataset["test"].map(
test_preprocess_function,
batched=True,
num_proc=1,
remove_columns=dataset["train"].column_names,
load_from_cache_file=False,
desc="Running tokenizer on dataset",
)
test_dataloader = DataLoader(test_dataset, collate_fn=default_data_collator, batch_size=batch_size, pin_memory=True)
next(iter(test_dataloader))
{'input_ids':tensor([[ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 227985, 5484, 915, 2566, 74757, 64626, 12384, 44639, 613, 52282, 2670, 79920, 3344, 1002, 368, 17646, 14472, 8348, 664, 718, 4, 19036, 17, 31849, 17, 6312, 76, 44, 62470, 56, 91, 50, 14839, 21, 77658, 915, 210], [ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 227985, 5484, 915, 405, 187059, 2256, 664, 2550, 18833, 18607, 162467, 4, 1387, 6199, 3291, 23405, 613, 4657, 17082, 566, 3432, 368, 78851, 1185, 61273, 23181, 1553, 15596, 212, 116057, 77658, 915, 210], [ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 227985, 5484, 915, 39762, 2566, 22253, 6201, 75701, 15, 632, 718, 5840, 10006, 6201, 18881, 427, 3804, 19528, 267, 158974, 1320, 368, 10029, 632, 49666, 92, 34, 77658, 915, 210], [ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 227985, 5484, 915, 2566, 104565, 8695, 2089, 6140, 109676, 99579, 1369, 512, 368, 4570, 54, 632, 368, 1503, 241485, 132226, 15, 982, 727, 1152, 18100, 861, 32596, 77597, 168154, 1306, 132226, 4346, 87843, 17, 130462, 364, 32923, 89, 53, 8309, 20, 75, 77658, 915, 210], [ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 227985, 5484, 915, 2566, 14173, 2960, 29906, 387, 20706, 49337, 1369, 77658, 915, 210], [ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 227985, 5484, 915, 2566, 219553, 45736, 36876, 1713, 72, 707, 187205, 13002, 177324, 77658, 915, 210], [ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 227985, 5484, 915, 2566, 233938, 28518, 13716, 427, 28146, 1119, 17918, 17, 236706, 368, 214997, 7555, 48659, 5276, 21600, 343, 17, 51416, 22403, 318, 1531, 1306, 1130, 20934, 567, 101161, 184849, 87843, 17, 1594, 15231, 2052, 16642, 20, 7180, 80, 26, 77658, 915, 210], [ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 227985, 5484, 915, 2566, 80, 2068, 479, 2566, 80, 1376, 878, 147587, 3904, 632, 368, 6084, 65673, 78851, 11736, 15527, 19082, 33151, 461, 17, 45575, 17887, 632, 5219, 14216, 68870, 5967, 1841, 4346, 87843, 17, 1594, 14512, 27, 71, 8184, 19, 290, 63748, 77658, 915, 210]]), 'attention_mask':tensor([[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, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 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, 0, 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, 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, 0, 0, 0, 0, 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, 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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 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, 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, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [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, 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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 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, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])}
2.3 调用 get_peft_model
方法包装基础的 Transformer
模型¶
model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
通过 print_trainable_parameters
方法可以查看可训练参数的数量(仅为300,288)以及占比(仅为0.05366%)。
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()
trainable params:300,288 || all params:559,514,880 || trainable%:0.05366935013417338
/home/guodong.li/code/peft-20230717/src/peft/tuners/p_tuning.py:146:UserWarning:for MLP, the `encoder_num_layers` is ignored. Exactly 2 MLP layers are used. warnings.warn(
model
PeftModelForCausalLM( (base_model):BloomForCausalLM( (transformer):BloomModel( (word_embeddings):Embedding(250880, 1024) (word_embeddings_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (h):ModuleList( (0):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (1):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (2):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (3):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (4):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (5):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (6):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (7):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (8):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (9):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (10):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (11):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (12):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (13):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (14):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (15):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (16):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (17):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (18):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (19):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (20):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (21):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (22):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) (23):BloomBlock( (input_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (self_attention):BloomAttention( (query_key_value):Linear(in_features=1024, out_features=3072, bias=True) (dense):Linear(in_features=1024, out_features=1024, bias=True) (attention_dropout):Dropout(p=0.0, inplace=False) ) (post_attention_layernorm):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (mlp):BloomMLP( (dense_h_to_4h):Linear(in_features=1024, out_features=4096, bias=True) (gelu_impl):BloomGelu() (dense_4h_to_h):Linear(in_features=4096, out_features=1024, bias=True) ) ) ) (ln_f):LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) (lm_head):Linear(in_features=1024, out_features=250880, bias=False) ) (prompt_encoder):ModuleDict( (default):PromptEncoder( (embedding):Embedding(20, 1024) (mlp_head):Sequential( (0):Linear(in_features=1024, out_features=128, bias=True) (1):ReLU() (2):Linear(in_features=128, out_features=128, bias=True) (3):ReLU() (4):Linear(in_features=128, out_features=1024, bias=True) ) ) ) (word_embeddings):Embedding(250880, 1024) )
model.peft_config
{'default':PromptEncoderConfig(peft_type=<PeftType.P_TUNING:'P_TUNING'>, auto_mapping=None, base_model_name_or_path='/data/nfs/llm/model/bloomz-560m', revision=None, task_type=<TaskType.CAUSAL_LM:'CAUSAL_LM'>, inference_mode=False, num_virtual_tokens=20, token_dim=1024, num_transformer_submodules=1, num_attention_heads=16, num_layers=24, encoder_reparameterization_type=<PromptEncoderReparameterizationType.MLP:'MLP'>, encoder_hidden_size=128, encoder_num_layers=2, encoder_dropout=0.0)}
2.4 模型训练¶
模型训练的其余部分均无需更改,当模型训练完成之后,保存高效微调部分的模型权重以供模型推理即可。
# model
# optimizer and lr scheduler
optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
lr_scheduler = get_linear_schedule_with_warmup(
optimizer=optimizer,
num_warmup_steps=0,
num_training_steps=(len(train_dataloader) * num_epochs),
)
# training and evaluation
model = model.to(device)
for epoch in range(num_epochs):
model.train()
total_loss = 0
for step, batch in enumerate(tqdm(train_dataloader)):
batch = {k: v.to(device) for k, v in batch.items()}
# print(batch)
# print(batch["input_ids"].shape)
outputs = model(**batch)
loss = outputs.loss
total_loss += loss.detach().float()
loss.backward()
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
model.eval()
eval_loss = 0
eval_preds = []
for step, batch in enumerate(tqdm(eval_dataloader)):
batch = {k: v.to(device) for k, v in batch.items()}
with torch.no_grad():
outputs = model(**batch)
loss = outputs.loss
eval_loss += loss.detach().float()
eval_preds.extend(
tokenizer.batch_decode(torch.argmax(outputs.logits, -1).detach().cpu().numpy(), skip_special_tokens=True)
)
eval_epoch_loss = eval_loss / len(eval_dataloader)
eval_ppl = torch.exp(eval_epoch_loss)
train_epoch_loss = total_loss / len(train_dataloader)
train_ppl = torch.exp(train_epoch_loss)
print(f"{epoch=}:{train_ppl=} {train_epoch_loss=} {eval_ppl=} {eval_epoch_loss=}")
100%|██████████| 7/7 [00:01<00:00, 4.66it/s] 100%|██████████| 7/7 [00:00<00:00, 16.60it/s]
epoch=0:train_ppl=tensor(1.4019e+23, device='cuda:0') train_epoch_loss=tensor(53.2973, device='cuda:0') eval_ppl=tensor(1.5565e+22, device='cuda:0') eval_epoch_loss=tensor(51.0993, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 9.08it/s] 100%|██████████| 7/7 [00:00<00:00, 17.60it/s]
epoch=1:train_ppl=tensor(3.4324e+14, device='cuda:0') train_epoch_loss=tensor(33.4694, device='cuda:0') eval_ppl=tensor(694663.1250, device='cuda:0') eval_epoch_loss=tensor(13.4512, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 9.15it/s] 100%|██████████| 7/7 [00:00<00:00, 17.51it/s]
epoch=2:train_ppl=tensor(594353.6875, device='cuda:0') train_epoch_loss=tensor(13.2952, device='cuda:0') eval_ppl=tensor(450830.4062, device='cuda:0') eval_epoch_loss=tensor(13.0188, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 9.12it/s] 100%|██████████| 7/7 [00:00<00:00, 17.51it/s]
epoch=3:train_ppl=tensor(673112.8125, device='cuda:0') train_epoch_loss=tensor(13.4197, device='cuda:0') eval_ppl=tensor(385877.5938, device='cuda:0') eval_epoch_loss=tensor(12.8633, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 8.92it/s] 100%|██████████| 7/7 [00:00<00:00, 16.12it/s]
epoch=4:train_ppl=tensor(565632.5625, device='cuda:0') train_epoch_loss=tensor(13.2457, device='cuda:0') eval_ppl=tensor(309009., device='cuda:0') eval_epoch_loss=tensor(12.6411, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 8.93it/s] 100%|██████████| 7/7 [00:00<00:00, 17.41it/s]
epoch=5:train_ppl=tensor(428292.1250, device='cuda:0') train_epoch_loss=tensor(12.9676, device='cuda:0') eval_ppl=tensor(264157.9688, device='cuda:0') eval_epoch_loss=tensor(12.4843, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 9.06it/s] 100%|██████████| 7/7 [00:00<00:00, 17.46it/s]
epoch=6:train_ppl=tensor(378711.1875, device='cuda:0') train_epoch_loss=tensor(12.8445, device='cuda:0') eval_ppl=tensor(251440.4219, device='cuda:0') eval_epoch_loss=tensor(12.4350, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 9.17it/s] 100%|██████████| 7/7 [00:00<00:00, 17.47it/s]
epoch=7:train_ppl=tensor(289856.9375, device='cuda:0') train_epoch_loss=tensor(12.5771, device='cuda:0') eval_ppl=tensor(242582.3281, device='cuda:0') eval_epoch_loss=tensor(12.3991, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 9.20it/s] 100%|██████████| 7/7 [00:00<00:00, 15.88it/s]
epoch=8:train_ppl=tensor(348310.9688, device='cuda:0') train_epoch_loss=tensor(12.7609, device='cuda:0') eval_ppl=tensor(234650.7188, device='cuda:0') eval_epoch_loss=tensor(12.3659, device='cuda:0')
100%|██████████| 7/7 [00:00<00:00, 8.99it/s] 100%|██████████| 7/7 [00:00<00:00, 17.26it/s]
epoch=9:train_ppl=tensor(344919.1250, device='cuda:0') train_epoch_loss=tensor(12.7511, device='cuda:0') eval_ppl=tensor(231422.7031, device='cuda:0') eval_epoch_loss=tensor(12.3520, device='cuda:0')
# 模型评估
model.eval()
i = 16
inputs = tokenizer(f'{text_column} :{dataset["test"][i]["Tweet text"]} Label :', return_tensors="pt")
print(dataset["test"][i]["Tweet text"])
print(inputs)
with torch.no_grad():
inputs = {k: v.to(device) for k, v in inputs.items()}
outputs = model.generate(
input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"], max_new_tokens=10, eos_token_id=3
)
print(outputs)
print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True))
Hey @nytimes your link to cancel my subscription isn't working and nobody is answering the chat. Please don't play that kind of stupid game. {'input_ids':tensor([[227985, 5484, 915, 54078, 2566, 7782, 24502, 2632, 8989, 427, 36992, 2670, 140711, 21994, 10789, 530, 88399, 632, 183542, 368, 44799, 17, 29901, 5926, 7229, 861, 11596, 461, 78851, 14775, 17, 77658, 915, 210]]), '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, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])} tensor([[227985, 5484, 915, 54078, 2566, 7782, 24502, 2632, 8989, 427, 36992, 2670, 140711, 21994, 10789, 530, 88399, 632, 183542, 368, 44799, 17, 29901, 5926, 7229, 861, 11596, 461, 78851, 14775, 17, 77658, 915, 210, 2550, 2, 36, 17, 1387, 51216, 632, 7220, 2, 36]], device='cuda:0') ["Tweet text :Hey @nytimes your link to cancel my subscription isn't working and nobody is answering the chat. Please don't play that kind of stupid game. Label : #A. The weather is goodA"]
# saving model
peft_model_id = f"{model_name_or_path}_{peft_config.peft_type}_{peft_config.task_type}"
print("model_output:", peft_model_id)
model.save_pretrained(peft_model_id)
model_output:/data/nfs/llm/model/bloomz-560m_P_TUNING_CAUSAL_LM
输出的模型权重文件如下所示:
/data/nfs/llm/model/bloomz-560m_P_TUNING_CAUSAL_LM
├── [ 451] adapter_config.json
├── [ 81K] adapter_model.bin
└── [ 129] README.md
0 directories, 3 files
注意:这里只会保存经过训练的增量 PEFT 权重。其中,adapter_config.json 为 P-Tuning 配置文件;adapter_model.bin 为 P-Tuning 权重文件。
ckpt = f"{peft_model_id}/adapter_model.bin"
!du -h $ckpt
print("--------------")
!tree -h $peft_model_id
huggingface/tokenizers:The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks... To disable this warning, you can either: - Avoid using `tokenizers` before the fork if possible - Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false) 84K /data/nfs/llm/model/bloomz-560m_P_TUNING_CAUSAL_LM/adapter_model.bin -------------- huggingface/tokenizers:The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks... To disable this warning, you can either: - Avoid using `tokenizers` before the fork if possible - Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false) /data/nfs/llm/model/bloomz-560m_P_TUNING_CAUSAL_LM ├── [ 451] adapter_config.json ├── [ 81K] adapter_model.bin └── [ 129] README.md 0 directories, 3 files
2.5 加载微调后的权重文件进行推理¶
from peft import PeftModel, PeftConfig
peft_model_id = f"{model_name_or_path}_{peft_config.peft_type}_{peft_config.task_type}"
print("model_input:", peft_model_id)
config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path)
model = PeftModel.from_pretrained(model, peft_model_id)
model_input:/data/nfs/llm/model/bloomz-560m_P_TUNING_CAUSAL_LM
model.to(device)
model.eval()
i = 4
inputs = tokenizer(f'{text_column} :{dataset["test"][i]["Tweet text"]} Label :', return_tensors="pt")
print(dataset["test"][i]["Tweet text"])
print(inputs)
with torch.no_grad():
inputs = {k: v.to(device) for k, v in inputs.items()}
outputs = model.generate(
input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"], max_new_tokens=10, eos_token_id=3
)
print(outputs)
print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True))