training loss doesn't decrease and bad prediction

rishizek / tensorflow-deeplab-v3

DeepLabv3 built in TensorFlow

MIT License

286 stars 102 forks source link

training loss doesn't decrease and bad prediction #33

Closed hakS07 closed 5 years ago

hakS07 commented 5 years ago

I tried to train DeeplabV3 pre-trained model on a new class(eyes iris) my dtat=RGB iris image+gray-scale iris label:1500 images(300*400) num_iteration=7000 crop_size=513 but the loss is unstable between 0.1 to 0.3 this example for a prediction Capture du 2019-05-27 22-56-27

TillBeemelmanns commented 5 years ago

A Mask-RCNN is usually better for small object detection

hakS07 commented 5 years ago

@TillBeemelmanns thank you