update4 2019-7-14
model.fit( )里 用validata_data 取代 validate_split 224× 224,app.yaml 及processor.py文件修改 224× 224
update5 2019-7-15
validate merge:将train data 和 test data都丢进去训练 batchnormalize earlystopping by loss ,val_loss
update6 2019-7-16
ImageDataGenerator save_to_dir:None或字符串,该参数能让你将提升后的图片保存起来,用以可视化 验证了使用效果:设训练集是300个,batch里提取的前300个是原生图片,第2次生成的300个是原生300个的变化,第3次重新生成原生300个的变化。
ReduceLROnPlateau 实验记录
- ResNet50 relu args.BATCH = 32 steps_per_epoch=15, Epoch 32/100
- 5s - loss: 6.6469 - acc: 0.5647 - val_loss: 12.5721 - val_acc: 0.2200
2.ResNet50 relu args.BATCH = 32 steps_per_epoch=15, Epoch 89/100
- 5s - loss: 0.2045 - acc: 0.9510 - val_loss: 13.2168 - val_acc: 0.1800
3.ResNet50 relu args.BATCH = 16 steps_per_epoch=60, Epoch 63/100
- 12s - loss: 0.0900 - acc: 0.9792 - val_loss: 13.2168 - val_acc: 0.1800
4.ResNet50 relu args.BATCH = 16 steps_per_epoch=60, 删除了 rescale=1./255, 之前的归一化重复了? Epoch 97/100
- 11s - loss: 0.0698 - acc: 0.9896 - val_loss: 0.0140 - val_acc: 0.9900
5.ResNet50 relu args.BATCH = 32 steps_per_epoch=150, Epoch = 300 save_best_only = True flyai : 86%