训练时间
在mbp的i5的cpu上训练了3轮,花的时间如下
Epoch 1/3
- 737s - loss: 0.1415 - val_loss: 0.0874
Epoch 2/3
- 608s - loss: 0.0807 - val_loss: 0.0577
Epoch 3/3
- 518s - loss: 0.0636 - val_loss: 0.0499
kaggle gpu telsa
Epoch 1/3
- 40s - loss: 0.1544 - val_loss: 0.0956
Epoch 2/3
- 38s - loss: 0.0871 - val_loss: 0.0665
Epoch 3/3
- 38s - loss: 0.0690 - val_loss: 0.0478
对比gpu和cpu,时间相差了1,2个数量级
GeForce GTX 1080
Epoch 1/3
- 47s - loss: 0.1349 - val_loss: 0.0890
Epoch 2/3
- 45s - loss: 0.0787 - val_loss: 0.0670
Epoch 3/3
- 43s - loss: 0.0625 - val_loss: 0.0466
在本地开发环境上的入门级显卡1080上,训练时间后和kaggle的环境相差不多。
Epoch=50
输出前后几轮的训练时间
Epoch 1/50
- 52s - loss: 0.1253 - val_loss: 0.0795
Epoch 2/50
- 48s - loss: 0.0738 - val_loss: 0.0565
Epoch 3/50
- 48s - loss: 0.0616 - val_loss: 0.0477
Epoch 4/50
- 49s - loss: 0.0534 - val_loss: 0.0378
Epoch 5/50
- 49s - loss: 0.0484 - val_loss: 0.0375
####################
Epoch 19/50
- 50s - loss: 0.0270 - val_loss: 0.0249
Epoch 20/50
- 50s - loss: 0.0257 - val_loss: 0.0241
Epoch 21/50
- 48s - loss: 0.0256 - val_loss: 0.0255
Epoch 22/50
- 50s - loss: 0.0247 - val_loss: 0.0255
Epoch 23/50
- 48s - loss: 0.0246 - val_loss: 0.0219
最终结果
50轮次,大概花了一个多小时,kaggle上的准确率从0.66提升到0.74,后续再考虑优化其他超参数,继续提升准确率