← retour au collage›Modifier / dupliquer

Collage sans titre

#6w5kCatWxGpublic / publicnouvelle version

anonyme créé 2 days ago Expire dans 5 days 4.7 KB syntaxe: text

Vos changements créent un nouveau collage lié à celui-ci — l’original reste intact.

nouvelle version Vos changements créent un nouveau collage lié à celui-ci — l’original reste intact.

Visibilité Fil public Accès public

Note de changement (facultatif)

Ce collage sera affiché dans le fil public. Changez la visibilité si vous voulez seulement le partager par lien.

Annuler

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
import torch.optim as optim
from sklearn.metrics import r2_score

# ---------------------- 全局设置 中文+随机种子 ----------------------
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False
torch.manual_seed(42)
np.random.seed(42)

# 读取数据
df = pd.read_csv("data.csv")
x_raw = df[['x']].values.astype(np.float32)
y_raw = df[['y']].values.astype(np.float32)
x_tensor = torch.from_numpy(x_raw)
y_tensor = torch.from_numpy(y_raw)

# 神经网络定义：单层隐藏层，GELU激活
class MLP(nn.Module):
    def __init__(self, hidden_n):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(1, hidden_n),
            nn.GELU(),
            nn.Linear(hidden_n, 1)
        )
    def forward(self, x):
        return self.net(x)

# ---------------------- 任务2：不同隐藏节点n，cosine decay，lr=0.01，epoch=10000 ----------------------
n_list = [2,4,6,8,16,32]  # 6个n，正好6张子图
epoch_total = 10000
lr0 = 0.01
criterion = nn.MSELoss()

res_fit = []   # 保存预测结果
res_loss = []  # 保存loss曲线
res_r2 = []    # 保存R2
lr_record = [] # 保存学习率变化（取n=8为例）

for n in n_list:
    model = MLP(n)
    opt = optim.Adam(model.parameters(), lr=lr0)
    scheduler = optim.lr_scheduler.CosineAnnealingLR(opt, T_max=epoch_total, eta_min=0)
    loss_hist = []
    for ep in range(epoch_total):
        opt.zero_grad()
        pred = model(x_tensor)
        loss = criterion(pred, y_tensor)
        loss.backward()
        opt.step()
        scheduler.step()
        loss_hist.append(loss.item())
        if n == 8:
            lr_record.append(opt.param_groups[0]['lr'])
    # 预测 & R2
    with torch.no_grad():
        y_pred = model(x_tensor).numpy()
    r2 = r2_score(y_raw, y_pred)
    res_fit.append(y_pred)
    res_loss.append(loss_hist)
    res_r2.append(r2)

# 图1：6个子图，不同n拟合效果
fig1, axes = plt.subplots(2,3,figsize=(14,9))
axes = axes.flatten()
for i, ax in enumerate(axes):
    ax.scatter(x_raw, y_raw, s=10, label="真实数据")
    ax.plot(x_raw, res_fit[i], c='red', lw=2, label=f"n={n_list[i]}, R2={res_r2[i]:.4f}")
    ax.set_title(f"隐藏层节点数 n={n_list[i]}")
    ax.legend()
    ax.grid(alpha=0.3)
fig1.suptitle("不同隐藏层节点数的拟合效果", fontsize=15)
plt.tight_layout()
plt.savefig("fit_compare.jpg", dpi=300)

# 图2：不同n训练损失曲线
plt.figure(figsize=(10,6))
colors = ["#e74c3c","#3498db","#2ecc71","#9b59b6","#f39c12","#1abc9c"]
for i, n in enumerate(n_list):
    plt.plot(res_loss[i], label=f"n={n}", color=colors[i])
plt.xlabel("Epoch")
plt.ylabel("MSE损失")
plt.yscale("log")
plt.title("不同隐藏层节点数的训练损失曲线")
plt.legend()
plt.grid(alpha=0.3)
plt.tight_layout()
plt.savefig("loss_n_curve.jpg", dpi=300)

# 图3：cosine decay学习率变化曲线（n=8）
plt.figure(figsize=(8,5))
plt.plot(lr_record)
plt.xlabel("Epoch")
plt.ylabel("学习率")
plt.title("Cosine Decay学习率变化曲线")
plt.grid(alpha=0.3)
plt.tight_layout()
plt.savefig("lr_cosine_curve.jpg", dpi=300)

# 图4：柱状图 R2对比
plt.figure(figsize=(8,5))
plt.bar([str(i) for i in n_list], res_r2, color=colors)
plt.xlabel("隐藏层节点数 n")
plt.ylabel("决定系数 R²")
plt.title("不同隐藏层节点数的决定系数对比")
plt.ylim(0, 1.05)
for idx,val in enumerate(res_r2):
    plt.text(idx, val+0.01, f"{val:.4f}", ha="center")
plt.tight_layout()
plt.savefig("r2_bar.jpg", dpi=300)

# ---------------------- 任务3：固定n=8，关闭cosine decay，lr=[0.001,0.01,0.1,1] ----------------------
lr_test_list = [0.001, 0.01, 0.1, 1.0]
loss_lr_all = []
for lr in lr_test_list:
    model = MLP(hidden_n=8)
    opt = optim.Adam(model.parameters(), lr=lr)
    # 不设置scheduler，lr固定不变
    loss_hist = []
    for ep in range(epoch_total):
        opt.zero_grad()
        pred = model(x_tensor)
        loss = criterion(pred, y_tensor)
        loss.backward()
        opt.step()
        loss_hist.append(loss.item())
    loss_lr_all.append(loss_hist)

# 图5：不同学习率损失曲线
fig2, axes = plt.subplots(2,2,figsize=(12,9))
axes = axes.flatten()
for i, ax in enumerate(axes):
    ax.plot(loss_lr_all[i])
    ax.set_title(f"学习率 lr={lr_test_list[i]}")
    ax.set_xlabel("Epoch")
    ax.set_ylabel("MSE Loss")
    ax.grid(alpha=0.3)
fig2.suptitle("不同学习率下的训练损失曲线（关闭cosine decay）", fontsize=15)
plt.tight_layout()
plt.savefig("lr_loss_compare.jpg", dpi=300)

plt.show()
print("===== 任务2结果R2 =====")
for n,r2 in zip(n_list, res_r2):
    print(f"n={n}, R2={r2:.4f}")
print("===== 全部绘图完成 =====")