在huggingface上制作小demo

在huggingface上制作小demo

今天好兄弟让我帮他搞一个模型，他有小样本的化学数据，想让我根据这些数据训练一个小模型，他想用这个模型预测一些值

最终我简单训练了一个小模型，起初想把这个模型和GUI界面打包成exe发给他，但是发现打包后3.9GB,太大了吧！！！后来我又找了别的方案，即将训练好的模型以及相关代码、环境配置文件上传到huggingface上，通过hf的界面端直接使用这个模型，接下来我回顾一下整个流程

1.训练模型并写一个简单的GUI

训练数据

模型输入值：substance、N、C、C/N、K、Cellulose、Hemicellulose、Lignin

模型输出值：Alkaliniy400、Alkalinity600

训练代码train.py

由于样本较小，为了减小误差，这里采用5折交叉验证

import pandas as pd
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import TensorDataset, DataLoader
from sklearn.model_selection import KFold
from sklearn.preprocessing import StandardScaler, OneHotEncoder
import joblib


# 定义神经网络模型
class AlkalinityNet(nn.Module):
    def __init__(self, input_dim):
        super(AlkalinityNet, self).__init__()
        self.model = nn.Sequential(
            nn.Linear(input_dim, 64),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(64, 32),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(32, 2)  # 输出两个值：400℃ 和 600℃ 碱度
        )

    def forward(self, x):
        return self.model(x)


def train_model(model, train_loader, criterion, optimizer, device):
    model.train()
    running_loss = 0.0
    for X_batch, y_batch in train_loader:
        X_batch = X_batch.to(device)
        y_batch = y_batch.to(device)
        optimizer.zero_grad()
        outputs = model(X_batch)
        loss = criterion(outputs, y_batch)
        loss.backward()
        optimizer.step()
        running_loss += loss.item() * X_batch.size(0)
    return running_loss


def evaluate_model(model, val_loader, criterion, device):
    model.eval()
    running_loss = 0.0
    with torch.no_grad():
        for X_batch, y_batch in val_loader:
            X_batch = X_batch.to(device)
            y_batch = y_batch.to(device)
            outputs = model(X_batch)
            loss = criterion(outputs, y_batch)
            running_loss += loss.item() * X_batch.size(0)
    return running_loss


def main():
    # 读取 Excel 数据
    data = pd.read_excel('~/PycharmProjects/Alkalinity/datasets/data.xlsx')

    # 假设第一列为物质名称，列名为 "Substance"
    # 数值特征
    num_features = ["N", "C", "C/N", "K", "Cellulose", "Hemicellulose", "Lignin"]
    targets = ["Alkalinity400", "Alkalinity600"]

    # 提取物质类别和数值特征
    substances = data["Substance"].values.reshape(-1, 1)
    X_num = data[num_features].values
    y = data[targets].values.astype(np.float32)

    # 对物质类别使用 OneHotEncoder 编码
    encoder = OneHotEncoder(sparse_output=False, handle_unknown='ignore')
    X_cat = encoder.fit_transform(substances)
    joblib.dump(encoder, 'encoder.pkl')

    # 对数值特征进行标准化
    scaler_X = StandardScaler()
    X_num_scaled = scaler_X.fit_transform(X_num)
    joblib.dump(scaler_X, 'scaler_X.pkl')

    # 拼接类别特征和数值特征
    X_combined = np.hstack([X_cat, X_num_scaled]).astype(np.float32)

    # 对目标值进行标准化
    scaler_y = StandardScaler()
    y_scaled = scaler_y.fit_transform(y)
    joblib.dump(scaler_y, 'scaler_y.pkl')

    # 转换为 PyTorch tensor
    X_tensor = torch.from_numpy(X_combined)
    y_tensor = torch.from_numpy(y_scaled)

    # 设置设备
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    # 交叉验证设置
    kf = KFold(n_splits=5, shuffle=True, random_state=42)
    num_epochs = 100
    batch_size = 8
    criterion = nn.MSELoss()
    fold_losses = []

    print("开始 5 折交叉验证...")
    for fold, (train_index, val_index) in enumerate(kf.split(X_tensor), 1):
        X_train, X_val = X_tensor[train_index], X_tensor[val_index]
        y_train, y_val = y_tensor[train_index], y_tensor[val_index]

        train_dataset = TensorDataset(X_train, y_train)
        val_dataset = TensorDataset(X_val, y_val)

        train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
        val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False)

        model = AlkalinityNet(input_dim=X_tensor.shape[1]).to(device)
        optimizer = optim.Adam(model.parameters(), lr=0.001)

        for epoch in range(num_epochs):
            train_loss = train_model(model, train_loader, criterion, optimizer, device)
            val_loss = evaluate_model(model, val_loader, criterion, device)
            # 这里可以打印每折每轮的损失，也可以选择每隔一定轮数打印一次
            # print(f"Fold {fold}, Epoch {epoch+1}/{num_epochs}, Train Loss: {train_loss/len(train_dataset):.4f}, Val Loss: {val_loss/len(val_dataset):.4f}")

        avg_val_loss = val_loss / len(val_dataset)
        print(f"第 {fold} 折验证 Loss: {avg_val_loss:.4f}")
        fold_losses.append(avg_val_loss)

    print("5 折交叉验证平均 Loss:", np.mean(fold_losses))

    # 在全数据集上训练最终模型
    final_dataset = TensorDataset(X_tensor, y_tensor)
    final_loader = DataLoader(final_dataset, batch_size=batch_size, shuffle=True)

    final_model = AlkalinityNet(input_dim=X_tensor.shape[1]).to(device)
    optimizer = optim.Adam(final_model.parameters(), lr=0.001)

    print("开始在全数据集上训练最终模型...")
    for epoch in range(num_epochs):
        train_loss = train_model(final_model, final_loader, criterion, optimizer, device)
        # 可打印训练进度
        # print(f"Epoch {epoch+1}/{num_epochs}, Loss: {train_loss/len(final_dataset):.4f}")

    # 保存最终模型参数
    torch.save(final_model.state_dict(), 'final_model.pth')
    print("最终模型已保存到 final_model.pth")


if __name__ == '__main__':
    main()

推理代码eval.py

import numpy as np
import torch
import torch.nn as nn
import joblib
from sklearn.preprocessing import StandardScaler, OneHotEncoder


# 定义与训练时相同的网络结构
class AlkalinityNet(nn.Module):
    def __init__(self, input_dim):
        super(AlkalinityNet, self).__init__()
        self.model = nn.Sequential(
            nn.Linear(input_dim, 64),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(64, 32),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(32, 2)
        )

    def forward(self, x):
        return self.model(x)


def predict_alkalinity(input_data):
    """
    参数 input_data 为字典，包含以下键值：
       "Substance": 物质名称，例如 "玉米秸秆"
       "N", "C", "C/N", "K", "Cellulose", "Hemicellulose", "Lignin"
    返回预测的 [400℃ 碱度, 600℃ 碱度]
    """
    # 数值特征顺序与训练时一致
    num_features = ["N", "C", "C/N", "K", "Cellulose", "Hemicellulose", "Lignin"]

    # 提取物质名称与数值特征
    substance = np.array([[input_data["Substance"]]])
    X_num = np.array([input_data[feat] for feat in num_features]).reshape(1, -1).astype(np.float32)

    # 加载保存的 OneHotEncoder 和 StandardScaler
    encoder = joblib.load('encoder.pkl')
    scaler_X = joblib.load('scaler_X.pkl')
    scaler_y = joblib.load('scaler_y.pkl')

    X_cat = encoder.transform(substance)
    X_num_scaled = scaler_X.transform(X_num)

    # 拼接类别特征和数值特征
    X_combined = np.hstack([X_cat, X_num_scaled]).astype(np.float32)

    # 转换为 tensor
    X_tensor = torch.from_numpy(X_combined)

    # 加载模型，注意输入维度需与训练时保持一致
    input_dim = X_combined.shape[1]
    model = AlkalinityNet(input_dim=input_dim)
    # 加载模型参数
    model.load_state_dict(torch.load('final_model.pth', map_location=torch.device('cpu'), weights_only=True))
    model.eval()

    with torch.no_grad():
        y_pred_tensor = model(X_tensor)
    y_pred_scaled = y_pred_tensor.numpy()

    # 将预测结果反标准化
    y_pred = scaler_y.inverse_transform(y_pred_scaled)
    return y_pred[0]


if __name__ == '__main__':
    # 示例输入，请根据实际物质成分调整数值
    input_data = {
        "Substance": "鸡粪",
        "N": 1.0,
        "C": 40.0,
        "C/N": 40.0,
        "K": 2.5,
        "Cellulose": 29.0,
        "Hemicellulose": 25.0,
        "Lignin": 12.0
    }
    result = predict_alkalinity(input_data)
    print("预测 400℃ 碱度：", result[0])
    print("预测 600℃ 碱度：", result[1])

本地推理看看

GUI界面代码app.py

import tkinter as tk
from tkinter import messagebox
import tkinter.font as tkFont
import joblib
import numpy as np
import torch
import torch.nn as nn

# 定义与训练时一致的模型结构
class AlkalinityNet(nn.Module):
    def __init__(self, input_dim):
        super(AlkalinityNet, self).__init__()
        self.model = nn.Sequential(
            nn.Linear(input_dim, 64),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(64, 32),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(32, 2)  # 输出 400℃ 和 600℃ 的碱度值
        )

    def forward(self, x):
        return self.model(x)

def predict():
    try:
        # 获取用户输入的各项数值
        substance = entry_substance.get()
        N = float(entry_N.get())
        C = float(entry_C.get())
        C_N = float(entry_CN.get())
        K = float(entry_K.get())
        cellulose = float(entry_cellulose.get())
        hemicellulose = float(entry_hemicellulose.get())
        lignin = float(entry_lignin.get())
    except ValueError:
        messagebox.showerror("输入错误", "请确保所有数值项均正确填写")
        return

    # 构造输入字典
    input_data = {
        "Substance": substance,
        "N": N,
        "C": C,
        "C/N": C_N,
        "K": K,
        "Cellulose": cellulose,
        "Hemicellulose": hemicellulose,
        "Lignin": lignin
    }

    try:
        # 加载保存的预处理器
        encoder = joblib.load('encoder.pkl')
        scaler_X = joblib.load('scaler_X.pkl')
        scaler_y = joblib.load('scaler_y.pkl')
    except Exception as e:
        messagebox.showerror("加载错误", f"加载预处理器失败：{e}")
        return

    # 对物质类别特征进行 one-hot 编码
    substance_arr = np.array([[input_data["Substance"]]])
    X_cat = encoder.transform(substance_arr)

    # 数值特征转换与标准化
    X_num = np.array([[input_data["N"], input_data["C"], input_data["C/N"], input_data["K"],
                       input_data["Cellulose"], input_data["Hemicellulose"], input_data["Lignin"]]], dtype=np.float32)
    X_num_scaled = scaler_X.transform(X_num)

    # 拼接特征
    X_combined = np.hstack([X_cat, X_num_scaled]).astype(np.float32)
    X_tensor = torch.from_numpy(X_combined)

    # 加载模型（注意：模型参数文件和输入预处理器需放在同一目录下）
    input_dim = X_combined.shape[1]
    model = AlkalinityNet(input_dim=input_dim)
    try:
        model.load_state_dict(torch.load('final_model.pth', map_location=torch.device('cpu'), weights_only=True))
    except Exception as e:
        messagebox.showerror("加载模型错误", f"加载模型失败：{e}")
        return
    model.eval()

    with torch.no_grad():
        y_pred_tensor = model(X_tensor)
    y_pred_scaled = y_pred_tensor.numpy()

    # 反标准化得到预测值
    y_pred = scaler_y.inverse_transform(y_pred_scaled)
    pred_400 = y_pred[0, 0]
    pred_600 = y_pred[0, 1]

    # 在界面上显示预测结果
    label_pred_400.config(text=str(pred_400))
    label_pred_600.config(text=str(pred_600))

# 创建 GUI 主窗口
root = tk.Tk()
root.title("碱度预测模型")

# 设置参考尺寸与基础字体大小（可根据需要调整）
REF_WIDTH = 800
REF_HEIGHT = 600
BASE_FONT_SIZE = 18

# 用于存放所有需要动态调整字体的控件
widgets_to_update = []

# 创建一个默认字体对象，初始大小 BASE_FONT_SIZE
default_font = tkFont.Font(family="SimSun", size=BASE_FONT_SIZE)

# 辅助函数：创建标签并加入更新列表（居中显示）
def create_label(text, row, col):
    lbl = tk.Label(root, text=text, font=default_font, anchor="center")
    lbl.grid(row=row, column=col, padx=5, pady=5, sticky="nsew")
    widgets_to_update.append(lbl)
    return lbl

# 辅助函数：创建输入框并加入更新列表（内容居中）
def create_entry(row, col):
    ent = tk.Entry(root, font=default_font, justify="center")
    ent.grid(row=row, column=col, padx=5, pady=5, sticky="nsew")
    widgets_to_update.append(ent)
    return ent

# 定义行列权重，使控件居中扩展
for i in range(11):
    root.grid_rowconfigure(i, weight=1)
for j in range(2):
    root.grid_columnconfigure(j, weight=1)

# 创建左侧标签
label_substance = create_label("物质", 0, 0)
label_N = create_label("N", 1, 0)
label_C = create_label("C", 2, 0)
label_CN = create_label("C/N", 3, 0)
label_K = create_label("K", 4, 0)
label_cellulose = create_label("纤维素", 5, 0)
label_hemicellulose = create_label("半纤维素", 6, 0)
label_lignin = create_label("木质素", 7, 0)
label_400 = create_label("400摄氏度碱度", 8, 0)
label_600 = create_label("600摄氏度碱度", 9, 0)

# 创建右侧输入框
entry_substance = create_entry(0, 1)
entry_N = create_entry(1, 1)
entry_C = create_entry(2, 1)
entry_CN = create_entry(3, 1)
entry_K = create_entry(4, 1)
entry_cellulose = create_entry(5, 1)
entry_hemicellulose = create_entry(6, 1)
entry_lignin = create_entry(7, 1)

# 用于显示预测结果的标签（400℃ 和 600℃）
label_pred_400 = create_label("未预测", 8, 1)
label_pred_600 = create_label("未预测", 9, 1)

# 预测按钮（也加入更新列表）
predict_button = tk.Button(root, text="预测", font=default_font, command=predict)
predict_button.grid(row=10, column=0, columnspan=2, pady=10)
widgets_to_update.append(predict_button)

# 定义一个函数，在窗口大小变化时更新所有控件的字体大小
def on_resize(event):
    # 根据窗口当前尺寸与参考尺寸计算缩放比例
    scale_factor = min(event.width / REF_WIDTH, event.height / REF_HEIGHT)
    new_font_size = max(int(BASE_FONT_SIZE * scale_factor), 8)  # 设置最小字体为8
    # 更新所有控件字体
    new_font = (default_font.actual("family"), new_font_size)
    for widget in widgets_to_update:
        widget.config(font=new_font)

# 绑定窗口尺寸变化事件
root.bind("<Configure>", on_resize)

root.mainloop()

本地运行看看效果

2.在huggingface上创建Space

点击new space

填写相关信息后点击Create

3.上传模型、代码、环境配置文件

上传你的模型和相应代码、以及requirements.txt

requirements.txt中直接写需要用到的库

上传文件中需要有个名为 app.py的文件，huggingface会根据这个文件创建网页端应用

为了能够让hf分享别人可以访问的public链接，在app.py中添加参数share=True

# Launch the app with shareable link
if __name__ == "__main__":
    iface.launch(share=True)

为了不让模型自动推理运行，而是让它点击运行时才推理，我们需要将app.py中 live=True设置为False

# Create Gradio interface
iface = gr.Interface(
    fn=predict,
    inputs=[
        gr.Textbox(label="Substance", placeholder="Enter substance"),
        gr.Number(label="N"),
        gr.Number(label="C"),
        gr.Number(label="C/N"),
        gr.Number(label="K"),
        gr.Number(label="Cellulose"),
        gr.Number(label="Hemicellulose"),
        gr.Number(label="Lignin")
    ],
    outputs=[
        gr.Number(label="400℃ Alkalinity"),
        gr.Number(label="600℃ Alkalinity")
    ],
    live=False,  # Disable live prediction to avoid automatic prediction
    title="Alkalinity Prediction Model",
    description="Input relevant data and click the 'Predict' button to get predictions."
)

hf会根据环境配置文件下载相关库并根据app.py创建应用

这样就可以在网页端直接使用模型

将左侧输入值填入后，点击submit后模型输出值显示到右侧

将分享链接分享给好兄弟后，他就可以直接在网页端使用我训练好的简单模型了