Training with smaller batch size

jocicmarko / ultrasound-nerve-segmentation

Deep Learning Tutorial for Kaggle Ultrasound Nerve Segmentation competition, using Keras

MIT License

939 stars 328 forks source link

Training with smaller batch size #63

Open mrgloom opened 6 years ago

mrgloom commented 6 years ago

I figure out that with batch size = 8 (original is 32) network doesn't converges, is it possible to train this Unet implementation with smaller batch size? (i.e. with batch size = 1?)

jocicmarko commented 6 years ago

Try adding BatchNormalization layer, it may help you to train with small batch sizes.