loss 在降低，但是Test Accuracy of the model =0.00000%

[charliex2]

pytorch-captcha-recognition git:(master) ✗ python captcha_train.py
init net
('epoch:', 0, 'step:', 0, 'loss:', 0.7110539674758911)
('epoch:', 1, 'step:', 0, 'loss:', 0.20382478833198547)
('epoch:', 2, 'step:', 0, 'loss:', 0.2277430146932602)
('epoch:', 3, 'step:', 0, 'loss:', 0.2373160570859909)
('epoch:', 4, 'step:', 0, 'loss:', 0.22053484618663788)
('epoch:', 5, 'step:', 0, 'loss:', 0.21534274518489838)
('epoch:', 6, 'step:', 0, 'loss:', 0.19580186903476715)
('epoch:', 7, 'step:', 0, 'loss:', 0.1724429875612259)
('epoch:', 8, 'step:', 0, 'loss:', 0.15848731994628906)
('epoch:', 9, 'step:', 0, 'loss:', 0.1449590027332306)
('epoch:', 10, 'step:', 0, 'loss:', 0.14423207938671112)
('epoch:', 11, 'step:', 0, 'loss:', 0.1371624767780304)
('epoch:', 12, 'step:', 0, 'loss:', 0.12939313054084778)
('epoch:', 13, 'step:', 0, 'loss:', 0.12587422132492065)
('epoch:', 14, 'step:', 0, 'loss:', 0.12168151885271072)
('epoch:', 15, 'step:', 0, 'loss:', 0.12061356008052826)
('epoch:', 16, 'step:', 0, 'loss:', 0.11823482811450958)
('epoch:', 17, 'step:', 0, 'loss:', 0.11660336703062057)
('epoch:', 18, 'step:', 0, 'loss:', 0.11383380740880966)
('epoch:', 19, 'step:', 0, 'loss:', 0.11137279868125916)
('epoch:', 20, 'step:', 0, 'loss:', 0.10716967284679413)
('epoch:', 21, 'step:', 0, 'loss:', 0.10241743177175522)
('epoch:', 22, 'step:', 0, 'loss:', 0.10174597799777985)
('epoch:', 23, 'step:', 0, 'loss:', 0.09972652792930603)
('epoch:', 24, 'step:', 0, 'loss:', 0.09792148321866989)
('epoch:', 25, 'step:', 0, 'loss:', 0.09143608063459396)
('epoch:', 26, 'step:', 0, 'loss:', 0.09200140088796616)
('epoch:', 27, 'step:', 0, 'loss:', 0.09041691571474075)
('epoch:', 28, 'step:', 0, 'loss:', 0.08658407628536224)
('epoch:', 29, 'step:', 0, 'loss:', 0.08212457597255707)
save last model
➜ pytorch-captcha-recognition git:(master) ✗ python captcha_test.py
load cnn net.
Test Accuracy of the model on the 3 test images: 0.000000 %

[Wang-Xiaodong1899]

你要训练epoch至少15次看看

[MortonWang]

-ShadowShadowWong

你好，遇到了与@charliex2 相同的问题

在batch=512的情况下，学习率从0.001到0.1，训练epoch<=100，结果是，best_val_acc=18%, best_test_acc=1%
请教模型的搭建以及超参数的设置

• 图片：200 x 70
• 训练集：10w
• 验证集：1000
• 测试集：2000

-Morton

[jian1299]

我也遇到了相同的问题
在batch=512的情况下，学习率从0.001，训练epoch 65，损失函数0.0102
结果best_val_acc=8%, best_test_acc=7%
请教模型的搭建以及超参数的设置

图片：160 x 80
训练集：10w
验证集：1000
测试集：2000