search

GeorgeSeif / Semantic-Segmentation-Suite

Semantic Segmentation Suite in TensorFlow. Implement, train, and test new Semantic Segmentation models easily!
2.51k stars 880 forks source link

How to train it well? #219

Open mengxia1994 opened 5 years ago

mengxia1994 commented 5 years ago

I use train.py to run about 300 epoch, other settings are also default(CamVid, FC-DenseNet56). But when i use the model to predict a image, its performance is very bad. Then i use one of the train image. It is much better but still far away from satisfied. The val performance stop improving from almost 250 epoch, precision and recall won't surpass 0.9. What should i do to obtain a model just like the main page show?

mojoee commented 4 years ago

i am also having issues training the Network with my own dataset at the moment. I have tried the usual things, as e.g. adjusting learning rate, optimizer, crop size, batch-size but it doesn't help