Tensorflow deepID CNN模型训练精度在稳定攀升至50-60%acc后下降

时间:2017-09-07 02:33:16

标签: machine-learning tensorflow deep-learning classification convolution

我正在训练CNN模型,根据2000个样本分类100个类别。为了避免过度拟合,我通过噪声和旋转等方法将2k样本论证到20k。

结果让我感到困惑,因为训练acc稳步攀升至40-60%,并迅速下降至低于10%。这种情况发生了好几次,起初我改变学习率来检查这是否有帮助。但即使我将lr设置为低至0.001,这种情况也会退出。

我使用0.2-0.5的dropout和conv和pooling层的规则。我还在resdual块中使用批量标准化。

Epoch 396/1000 - loss: 0.695151 - acc: 0.422    validation acc: 0.36
Epoch 397/1000 - loss: 0.694972 - acc: 0.416    validation acc: 0.34
Epoch 398/1000 - loss: 0.698717 - acc: 0.152    validation acc: 0.12
Epoch 399/1000 - loss: 0.698495 - acc: 0.14 validation acc: 0.14
Epoch 400/1000 - loss: 0.696137 - acc: 0.344    validation acc: 0.26
Epoch 401/1000 - loss: 0.694856 - acc: 0.44 validation acc: 0.28
Epoch 402/1000 - loss: 0.694578 - acc: 0.454    validation acc: 0.32
Epoch 403/1000 - loss: 0.694385 - acc: 0.464    validation acc: 0.34
Epoch 404/1000 - loss: 0.694594 - acc: 0.434    validation acc: 0.36
Epoch 405/1000 - loss: 0.695077 - acc: 0.376    validation acc: 0.36
Epoch 406/1000 - loss: 0.694583 - acc: 0.426    validation acc: 0.36
Epoch 407/1000 - loss: 0.695255 - acc: 0.39 validation acc: 0.32
Epoch 408/1000 - loss: 0.694716 - acc: 0.432    validation acc: 0.34
Epoch 409/1000 - loss: 0.694765 - acc: 0.42 validation acc: 0.36
Epoch 410/1000 - loss: 0.694443 - acc: 0.446    validation acc: 0.36
Epoch 411/1000 - loss: 0.694334 - acc: 0.462    validation acc: 0.38
Epoch 412/1000 - loss: 0.69475 - acc: 0.404 validation acc: 0.36
Epoch 413/1000 - loss: 0.694607 - acc: 0.428    validation acc: 0.36
Epoch 414/1000 - loss: 0.69983 - acc: 0.022 validation acc: 0.02
Epoch 415/1000 - loss: 0.698023 - acc: 0.014    validation acc: 0.02
Epoch 416/1000 - loss: 0.697584 - acc: 0.012    validation acc: 0.0
Epoch 417/1000 - loss: 0.697573 - acc: 0.01 validation acc: 0.0
Epoch 418/1000 - loss: 0.697533 - acc: 0.01 validation acc: 0.0
Epoch 419/1000 - loss: 0.697563 - acc: 0.014    validation acc: 0.0
Epoch 420/1000 - loss: 0.697468 - acc: 0.038    validation acc: 0.0
Epoch 421/1000 - loss: 0.697476 - acc: 0.026    validation acc: 0.12
Epoch 422/1000 - loss: 0.697426 - acc: 0.04 validation acc: 0.12
Epoch 423/1000 - loss: 0.697475 - acc: 0.078    validation acc: 0.12
Epoch 424/1000 - loss: 0.697317 - acc: 0.074    validation acc: 0.08

非常感谢,如果对此问题有任何建议。

0 个答案:

没有答案