feat(train_page): 添加模型训练超参数配置功能
- 新增学习率、批次大小、最大训练步数等超参数输入组件 - 实现超参数在训练过程中的动态应用 - 调整训练参数以适应不同硬件环境 - 优化训练过程，支持按步数保存模型
This commit is contained in:
carry
parent
5094febcb4
commit
2d39b91764
@ -13,16 +13,16 @@ from unsloth.chat_templates import get_chat_template
|
||||
from tools import formatting_prompts_func
|
||||
|
||||
sys.path.append(str(Path(__file__).resolve().parent.parent))
|
||||
from global_var import get_model, get_tokenizer, get_datasets
|
||||
from global_var import get_model, get_tokenizer, get_datasets, get_workdir
|
||||
from tools import formatting_prompts_func
|
||||
|
||||
def train_page():
|
||||
with gr.Blocks() as demo:
|
||||
|
||||
gr.Markdown("## 微调")
|
||||
# 获取数据集列表并设置初始值
|
||||
datasets_list = [str(ds["name"]) for ds in get_datasets().all()]
|
||||
initial_dataset = datasets_list[0] if datasets_list else None
|
||||
|
||||
dataset_dropdown = gr.Dropdown(
|
||||
choices=datasets_list,
|
||||
value=initial_dataset, # 设置初始选中项
|
||||
@ -30,14 +30,27 @@ def train_page():
|
||||
allow_custom_value=True,
|
||||
interactive=True
|
||||
)
|
||||
|
||||
# 新增超参数输入组件
|
||||
learning_rate_input = gr.Number(value=2e-4, label="学习率")
|
||||
per_device_train_batch_size_input = gr.Number(value=1, label="每设备训练批次大小", precision=0)
|
||||
max_steps_input = gr.Number(value=60, label="最大训练步数", precision=0)
|
||||
epoch_input = gr.Number(value=1, label="训练轮数", precision=0)
|
||||
save_steps_input = gr.Number(value=20, label="保存步数", precision=0) # 新增保存步数输入框
|
||||
|
||||
train_button = gr.Button("开始微调")
|
||||
|
||||
# 训练状态输出
|
||||
output = gr.Textbox(label="训练日志", interactive=False)
|
||||
|
||||
def train_model(dataset_name):
|
||||
def train_model(dataset_name, max_seq_length, learning_rate, per_device_train_batch_size, epoch, save_steps):
|
||||
# 使用动态传入的超参数
|
||||
learning_rate = float(learning_rate)
|
||||
per_device_train_batch_size = int(per_device_train_batch_size)
|
||||
epoch = int(epoch)
|
||||
save_steps = int(save_steps) # 新增保存步数参数
|
||||
|
||||
# 模型配置参数
|
||||
max_seq_length = 4096 # 最大序列长度
|
||||
dtype = None # 数据类型，None表示自动选择
|
||||
load_in_4bit = False # 使用4bit量化加载模型以节省显存
|
||||
|
||||
@ -78,7 +91,6 @@ def train_page():
|
||||
loftq_config=None,
|
||||
)
|
||||
|
||||
# 配置分词器
|
||||
tokenizer = get_chat_template(
|
||||
tokenizer,
|
||||
chat_template="qwen-2.5",
|
||||
@ -91,48 +103,50 @@ def train_page():
|
||||
dataset = dataset.map(formatting_prompts_func,
|
||||
fn_kwargs={"tokenizer": tokenizer},
|
||||
batched=True)
|
||||
print(dataset[5])
|
||||
|
||||
|
||||
# 初始化SFT训练器
|
||||
trainer = SFTTrainer(
|
||||
model=model, # 待训练的模型
|
||||
model=model, # 待训练的模型
|
||||
tokenizer=tokenizer, # 分词器
|
||||
train_dataset=dataset, # 训练数据集
|
||||
dataset_text_field="text", # 数据集字段的名称
|
||||
max_seq_length=max_seq_length, # 最大序列长度
|
||||
max_seq_length=model.max_seq_length, # 最大序列长度
|
||||
data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer),
|
||||
dataset_num_proc=1, # 数据集处理的并行进程数，提高CPU利用率
|
||||
dataset_num_proc=1, # 数据集处理的并行进程数
|
||||
packing=False,
|
||||
args=TrainingArguments(
|
||||
per_device_train_batch_size=2, # 每个GPU的训练批次大小
|
||||
per_device_train_batch_size=per_device_train_batch_size, # 每个GPU的训练批次大小
|
||||
gradient_accumulation_steps=4, # 梯度累积步数,用于模拟更大的batch size
|
||||
warmup_steps=5, # 预热步数,逐步增加学习率
|
||||
learning_rate=2e-4, # 学习率
|
||||
lr_scheduler_type="linear", # 线性学习率调度器
|
||||
max_steps=60, # 最大训练步数（一步 = 处理一个batch的数据）
|
||||
# 根据硬件支持选择训练精度
|
||||
warmup_steps=int(epoch * 0.1), # 预热步数,逐步增加学习率
|
||||
learning_rate=learning_rate, # 学习率
|
||||
lr_scheduler_type="linear", # 线性学习率调度器
|
||||
max_steps=int(epoch * len(dataset)/per_device_train_batch_size), # 最大训练步数（一步 = 处理一个batch的数据）
|
||||
fp16=not is_bfloat16_supported(), # 如果不支持bf16则使用fp16
|
||||
bf16=is_bfloat16_supported(), # 如果支持则使用bf16
|
||||
logging_steps=1, # 每1步记录一次日志
|
||||
optim="adamw_8bit", # 使用8位AdamW优化器节省显存，几乎不影响训练效果
|
||||
weight_decay=0.01, # 权重衰减系数,用于正则化，防止过拟合
|
||||
seed=3407, # 随机数种子
|
||||
output_dir="outputs", # 保存模型检查点和训练日志
|
||||
seed=114514, # 随机数种子
|
||||
output_dir=get_workdir() + "/checkpoint/", # 保存模型检查点和训练日志
|
||||
save_strategy="steps", # 按步保存中间权重
|
||||
save_steps=20, # 每20步保存一次中间权重
|
||||
save_steps=save_steps, # 使用动态传入的保存步数
|
||||
# report_to="tensorboard", # 将信息输出到tensorboard
|
||||
),
|
||||
)
|
||||
|
||||
# 开始训练，resume_from_checkpoint为True表示从最新的模型开始训练
|
||||
trainer_stats = trainer.train(resume_from_checkpoint = True)
|
||||
|
||||
|
||||
# 开始训练
|
||||
trainer_stats = trainer.train(resume_from_checkpoint=True)
|
||||
|
||||
train_button.click(
|
||||
fn=train_model,
|
||||
inputs=dataset_dropdown,
|
||||
inputs=[
|
||||
dataset_dropdown,
|
||||
learning_rate_input,
|
||||
per_device_train_batch_size_input,
|
||||
max_steps_input,
|
||||
epoch_input,
|
||||
save_steps_input
|
||||
],
|
||||
outputs=output
|
||||
)
|
||||
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user