一、你现在在跑的”最终版实验”是什么？

实验名（你以后论文里就这么叫）：

ResNet50 + Temporal Attention Pooling
(Backbone Frozen, Cross-Entropy Loss)

对照对象：

在冻结 backbone 的条件下，引入轻量级时序注意力后，模型在 MARS 数据集上从随机水平稳定收敛，Cross-Entropy loss 从 8.04 下降至 1.48，显著优于简单时序平均的 baseline。

跑了20轮

meanpooling

![[Pasted image 20260122205651.png]]

时序注意力

![[Pasted image 20260122205727.png]]

loss函数变化如下

PS C:\Users\HUASHUO\Desktop\实习_科研\时序建模> & D:/Python/python.exe c:/Users/HUASHUO/Desktop/实习_科研/时序建模/train_attention_frozen.py
===== Start Training (Backbone Frozen) =====
Epoch [1/20]: 100%|██████████████████████████████████| 157/157 [00:36<00:00, 4.36it/s, loss=8.6250]
Epoch [1/20] Average Loss: 8.0385
Epoch [2/20]: 100%|██████████████████████████████████| 157/157 [00:29<00:00, 5.39it/s, loss=7.7270]
Epoch [2/20] Average Loss: 6.5649
Epoch [3/20]: 100%|██████████████████████████████████| 157/157 [00:36<00:00, 4.35it/s, loss=7.4963]
Epoch [3/20] Average Loss: 5.8414
Epoch [4/20]: 100%|██████████████████████████████████| 157/157 [00:39<00:00, 3.97it/s, loss=6.5035]
Epoch [4/20] Average Loss: 5.1334
Epoch [5/20]: 100%|██████████████████████████████████| 157/157 [00:26<00:00, 5.98it/s, loss=7.3543]
Epoch [5/20] Average Loss: 4.5142
Epoch [6/20]: 100%|██████████████████████████████████| 157/157 [00:18<00:00, 8.57it/s, loss=6.2687]
Epoch [6/20] Average Loss: 4.2014
Epoch [7/20]: 100%|██████████████████████████████████| 157/157 [00:19<00:00, 8.14it/s, loss=5.8768]
Epoch [7/20] Average Loss: 3.8086
Epoch [8/20]: 100%|██████████████████████████████████| 157/157 [00:16<00:00, 9.67it/s, loss=5.0060]
Epoch [8/20] Average Loss: 3.5035
Epoch [9/20]: 100%|██████████████████████████████████| 157/157 [00:17<00:00, 9.06it/s, loss=7.2790]
Epoch [9/20] Average Loss: 3.2064
Epoch [10/20]: 100%|█████████████████████████████████| 157/157 [00:19<00:00, 8.07it/s, loss=7.0057]
Epoch [10/20] Average Loss: 2.9417
Epoch [11/20]: 100%|█████████████████████████████████| 157/157 [00:24<00:00, 6.44it/s, loss=6.4427]
Epoch [11/20] Average Loss: 2.7127
Epoch [12/20]: 100%|█████████████████████████████████| 157/157 [00:26<00:00, 5.96it/s, loss=6.1261]
Epoch [12/20] Average Loss: 2.4385
Epoch [13/20]: 100%|█████████████████████████████████| 157/157 [00:22<00:00, 7.03it/s, loss=4.0942]
Epoch [13/20] Average Loss: 2.2597
Epoch [14/20]: 100%|█████████████████████████████████| 157/157 [00:38<00:00, 4.10it/s, loss=4.9109]
Epoch [14/20] Average Loss: 2.1240
Epoch [15/20]: 100%|█████████████████████████████████| 157/157 [00:23<00:00, 6.79it/s, loss=6.6973]
Epoch [15/20] Average Loss: 2.0258
Epoch [16/20]: 100%|█████████████████████████████████| 157/157 [00:20<00:00, 7.66it/s, loss=4.1258]
Epoch [16/20] Average Loss: 1.7999
Epoch [17/20]: 100%|█████████████████████████████████| 157/157 [00:18<00:00, 8.64it/s, loss=5.2614]
Epoch [17/20] Average Loss: 1.7342
Epoch [18/20]: 100%|█████████████████████████████████| 157/157 [00:16<00:00, 9.24it/s, loss=6.4694]
Epoch [18/20] Average Loss: 1.6658
Epoch [19/20]: 100%|█████████████████████████████████| 157/157 [00:17<00:00, 8.93it/s, loss=3.1201]
Epoch [19/20] Average Loss: 1.5710
Epoch [20/20]: 100%|█████████████████████████████████| 157/157 [00:16<00:00, 9.35it/s, loss=4.8086]
Epoch [20/20] Average Loss: 1.4774
===== Training Finished =====

对比