train own dataset but cbf find hdf5

lucasjinreal / keras_frcnn

Keras Implementation of faster-rcnn

521 stars 278 forks source link

train own dataset but cbf find hdf5 #48

Open ycAlex11 opened 5 years ago

ycAlex11 commented 5 years ago

@jinfagang could you explain what is the kitti_frcnn_last.hdf5? As my understanding is a file to save weight(or the net) if you dont want train the whole data from beginning. so what I am going to do for this if i am going to train my own data? tbh this is not a really useful readme!!

TianwenZhang0825 commented 5 years ago

start train , the .hdf5 file is not a must. !!!!!!!add ./mole to your project, otherwise you will not get the weights files.

ycAlex11 commented 5 years ago

thanks for your reply, i dont really get your point.... where I am going to add ./mode(i guess it should be a folder calls mode ??) And another thing, how do I increase utilization ratio of GPU? Btw, i added your wechat cheers

dek8v5 commented 5 years ago

As I understand, you need to download the pre trained model if you dont wanna train your model. https://github.com/fchollet/deep-learning-models/releases/tag/v0.1 download the first model resnet50... there. Then put it inside model dir. Change cfg.model_path in line 173-175 with cfg.base_net_weights, this cfg.base_net_weights is pointed to your model that you just downloaded in model dir (by Keras-frcnn config.py). Correct me if I am wrong. Thanks