模型性能对齐的一些疑问

您好：我在使用如下命令进行训练时

python train.py --exp_name bam_xlsr --train_root ./data/raw/train --dev_root ./data/raw/dev

模型运行到第50个Epoch训练结束,此时得到的模型结果如下所示：

2024-11-11 13:04:20 - INFO - Start training.
2024-11-11 13:57:31 - INFO - Epoch [0]: test eer 0.08273609685967305
2024-11-11 13:57:31 - INFO - Epoch [0]: test acc 0.9156937292458732
2024-11-11 13:57:31 - INFO - Epoch [0]: test precision 0.8736008092477612
2024-11-11 13:57:31 - INFO - Epoch [0]: test recall 0.973322301026438
2024-11-11 13:57:31 - INFO - Epoch [0]: test F1 0.9207694084344986
2024-11-11 13:57:31 - INFO - ---------------------------------------------------------
2024-11-11 13:57:32 - INFO - Epoch [0]: binary test eer 0.04723030337679158
2024-11-11 13:57:32 - INFO - Epoch [0]: binary test acc 0.9660445642234948
2024-11-11 13:57:32 - INFO - Epoch [0]: binary test precision 0.8601362946750566
2024-11-11 13:57:32 - INFO - Epoch [0]: binary test recall 0.9154038565888908
2024-11-11 13:57:32 - INFO - Epoch [0]: binary test F1 0.8869099138953536
2024-11-11 13:57:32 - INFO - ---------------------------------------------------------

我又尝试使用如下命令在上一次训练的基础上继续训练：

 python train.py --train_root ./data/raw/train --dev_root ./data/raw/dev --checkpoint exp/bam_wavlm/train/lightning_logs/version_1/checkpoints/47-0.05829.ckpt --continue_train True

继续训练50个Epoch后的模型在测试集上的结果如下：

2024-11-12 16:19:58 - INFO - Start training.
2024-11-12 17:12:42 - INFO - Epoch [0]: test eer 0.0941809846777977
2024-11-12 17:12:42 - INFO - Epoch [0]: test acc 0.9011662415579832
2024-11-12 17:12:42 - INFO - Epoch [0]: test precision 0.8511679264792399
2024-11-12 17:12:42 - INFO - Epoch [0]: test recall 0.9739278879804376
2024-11-12 17:12:42 - INFO - Epoch [0]: test F1 0.9084193546277286
2024-11-12 17:12:42 - INFO - ---------------------------------------------------------
2024-11-12 17:12:44 - INFO - Epoch [0]: binary test eer 0.052092465838956326
2024-11-12 17:12:44 - INFO - Epoch [0]: binary test acc 0.9618287694306752
2024-11-12 17:12:44 - INFO - Epoch [0]: binary test precision 0.8396237913382628
2024-11-12 17:12:44 - INFO - Epoch [0]: binary test recall 0.9117157894272467
2024-11-12 17:12:44 - INFO - Epoch [0]: binary test F1 0.8741859958504029
2024-11-12 17:12:44 - INFO - ---------------------------------------------------------

为什么结果反而下降了呢?我在加载官方发布的模型得到的结果如下：

2024-11-07 10:49:46 - INFO - Start training.
2024-11-07 11:40:07 - INFO - Epoch [0]: test eer 0.037801628315819805
2024-11-07 11:40:07 - INFO - Epoch [0]: test acc 0.9591388741838716
2024-11-07 11:40:07 - INFO - Epoch [0]: test precision 0.9390206460636512
2024-11-07 11:40:07 - INFO - Epoch [0]: test recall 0.9826253915597752
2024-11-07 11:40:07 - INFO - Epoch [0]: test F1 0.9603282935104467
2024-11-07 11:40:07 - INFO - ---------------------------------------------------------
2024-11-07 11:40:08 - INFO - Epoch [0]: binary test eer 0.03609208325332974
2024-11-07 11:40:08 - INFO - Epoch [0]: binary test acc 0.9753616632176717
2024-11-07 11:40:08 - INFO - Epoch [0]: binary test precision 0.9017411342977851
2024-11-07 11:40:08 - INFO - Epoch [0]: binary test recall 0.9321854444870895
2024-11-07 11:40:08 - INFO - Epoch [0]: binary test F1 0.9167105922469723
2024-11-07 11:40:08 - INFO - --------------------------------------------------

请问造成这种差异的原因有可能是啥呢？

media-sec-lab / BAM

模型性能对齐的一些疑问 #5