Train model from scratch problem

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow

Other

24.65k stars 11.7k forks source link

Train model from scratch problem #1253

Open enoceanwei opened 5 years ago

enoceanwei commented 5 years ago

Dear All,

I revised the backbone architecture, so I need to train a model from scratch (without using the  pretrain .h5 file). I just comment out the 'model.load_weight', however, the result is terrible. the new model can not detect anything almostly.

Could anybody give me some suggestion how to fix the problem? is there anything do I need to changed in the codes? many thanks.

King regards

Wei

jodeiri commented 5 years ago

In the model.train the "layer" argument should changes from "heads" to "all".

ectg commented 5 years ago

Hi @enoceanwei , how did you change the backbone?

enoceanwei commented 5 years ago

Dear @ectg

Sorry to reply a late. I followed the original paper net architecture, you can look thought the 'resnet_graph' function, and changed the blocks numbers which is introduced in the mask paper.

Regards

Wei