AutoAugment is not very efficient on my 5000 item dataset?

google / automl

Google Brain AutoML

Apache License 2.0

6.18k stars 1.45k forks source link

AutoAugment is not very efficient on my 5000 item dataset? #481

Open fitoule opened 4 years ago

fitoule commented 4 years ago

augment

Hi I tried on my dataset training EfficientDet D2 3 classes about 5000 images training (finetuning from D2 Cocodataset). When I try autoaugment policy V2 or V3 it is always not as so good as the raw training without any augment :( How cxan I resolve that ?

Thank you

mingxingtan commented 4 years ago

Yeah, I am not sure how effective of those autoaugmentation: they might depend on the dataset and so on.

For COCO, I found autoaugment needs longer training, so I ended up not using them for COCO.

Koteswara-ML commented 3 years ago

Hey, Did you train on TPU..for training using augmentations, Can you help me with your training parameters .