PKU_NLPDL_Assignment1 / report /task2report.md
Antoine Li
Upload folder using huggingface_hub
0ee6a96 verified
|
Raw
History Blame Contribute Delete
1.39 kB

task2report

姓名:李长烨 学号:2200017853

configurations of CNN model

本CNN模型基本由三重结构组成,第一层为embedding层,第二层包括三个卷积层,卷积核大小分别为3,4,5,用于提取不同长度的gram特征,三个卷积层均通过一个激活函数层和一个最大池化层,其中卷积核大小为3的卷积层激活函数为relu,其他两个卷积层激活函数为tanh,最后一层为全连接层,输入了为三个卷积层结果的拼接,输出为结果logits。

其他参数中,我们通过训练集构建vocabulary, $embed\ size = 100, num\ classes = 4$

训练中采用early-stop,具体参数为$epochs=10, batch\ size=64, learning\ rate=0.001, patience=3$

Classification accuracy on test set

Test Accuracy=0.7750

具体训练中的参数如下

Epoch Loss Dev Accuracy
1/10 84.1876 0.6400
2/10 49.7036 0.7100
3/10 28.8129 0.7310
4/10 14.6686 0.7390
5/10 6.8835 0.7390
6/10 3.5254 0.7630
7/10 1.9510 0.7660
8/10 1.2503 0.7670
9/10 0.8980 0.7710
10/10 0.6942 0.7760
Test Accuracy 0.7750