day39模型的可视化和推理@浙大疏锦行
主要针对隐藏层神经元的个数进行了修改
# 实验 1: 原始配置 (隐藏层神经元 = 10)print("=== 实验 1: 原始配置 (Hidden Size = 10) ===")model_base=MLP(input_size=4,hidden_size=10,output_size=3).to(device)time_base,acc_base,losses_base=train_and_evaluate(model_base,learning_rate=0.01,num_epochs=10000,desc="Base Model")print(f"Base Model - Time:{time_base:.2f}s, Accuracy:{acc_base*100:.2f}%")# 实验 2: 增加隐藏层神经元 (隐藏层神经元 = 50)print("\n=== 实验 2: 增加隐藏层神经元 (Hidden Size = 50) ===")model_large=MLP(input_size=4,hidden_size=50,output_size=3).to(device)time_large,acc_large,losses_large=train_and_evaluate(model_large,learning_rate=0.01,num_epochs=10000,desc="Large Model")print(f"Large Model - Time:{time_large:.2f}s, Accuracy:{acc_large*100:.2f}%")# 实验 3: 减少隐藏层神经元 (隐藏层神经元 = 4)print("\n=== 实验 3: 减少隐藏层神经元 (Hidden Size = 4) ===")model_small=MLP(input_size=4,hidden_size=4,output_size=3).to(device)time_small,acc_small,losses_small=train_and_evaluate(model_small,learning_rate=0.01,num_epochs=10000,desc="Small Model")print(f"Small Model - Time:{time_small:.2f}s, Accuracy:{acc_small*100:.2f}%")=== 实验 1: 原始配置 (Hidden Size = 10) === Base Model: 10000/10000 [00:12<00:00, 780.84epoch/s, Loss=0.0943] Base Model: 10000/10000 [00:12<00:00, 780.84epoch/s, Loss=0.0943] Base Model - Time: 12.81s, Accuracy: 96.67% === 实验 2: 增加隐藏层神经元 (Hidden Size = 50) === Large Model: 10000/10000 [00:12<00:00, 793.83epoch/s, Loss=0.0857] Large Model: 10000/10000 [00:12<00:00, 793.83epoch/s, Loss=0.0857] Large Model - Time: 12.60s, Accuracy: 96.67% === 实验 3: 减少隐藏层神经元 (Hidden Size = 4) === Small Model: 10000/10000 [00:13<00:00, 761.09epoch/s, Loss=0.0849] Small Model - Time: 13.14s, Accuracy: 96.67%可视化
# 可视化对比plt.figure(figsize=(15,6))# Loss Curveplt.subplot(1,2,1)plt.plot(losses_base,label='Hidden=10')plt.plot(losses_large,label='Hidden=50')plt.plot(losses_small,label='Hidden=4')plt.xlabel('Steps (x100 epochs)')plt.ylabel('Loss')plt.title('Training Loss Comparison')plt.legend()plt.grid(True)# Accuracy and Time Bar Chartplt.subplot(1,2,2)models=['Hidden=10','Hidden=50','Hidden=4']accs=[acc_base*100,acc_large*100,acc_small*100]# Convert to percentagetimes=[time_base,time_large,time_small]x=np.arange(len(models))width=0.35ax1=plt.gca()ax2=ax1.twinx()bars1=ax1.bar(x-width/2,accs,width,label='Accuracy (%)',color='skyblue')bars2=ax2.bar(x+width/2,times,width,label='Time (s)',color='salmon')ax1.set_ylabel('Accuracy (%)')ax2.set_ylabel('Time (s)')ax1.set_ylim(0,110)# Accuracy 0-100+ax1.set_xticks(x)ax1.set_xticklabels(models)plt.title('Performance Comparison')# Add legendslines1,labels1=ax1.get_legend_handles_labels()lines2,labels2=ax2.get_legend_handles_labels()ax1.legend(lines1+lines2,labels1+labels2,loc='upper left')plt.tight_layout()plt.show()@浙大疏锦行
