search

warmspringwinds / tf-image-segmentation

Image Segmentation framework based on Tensorflow and TF-Slim library
MIT License
549 stars 188 forks source link

Can we fine tune the FCN_8s network on our own dataset? #14

Open robot010 opened 7 years ago

robot010 commented 7 years ago

Hi, I was wondering if we can utilize the fcn_8s_train.py file to fine-tuned the FCN8s on our own dataset. It seems that when I do this, there is an incompatible shape problem because my dataset only has 7 classes. Besides, the fcn_8s_train.py file is trained on FCN_16s, am I right? So it might be improper for me to use that script. Do I have to use the tf_recored script to convert my dataset first? Any suggestion is welcome. Thanks.

MrChristo59 commented 7 years ago

Did you manage to tune the network with your dataset ? I also want to do that but can't find the right way.

robot010 commented 7 years ago

@MrChristo59 I would like to suggest you check on this site: https://github.com/MarvinTeichmann/tensorflow-fcn, which I found out how to fine-tune on my own dataset. I noticed that for FCN, there is no need to have the deconv layer pre-trained, as long as we wanted to change the number of output neurons, we will need to train the deconv layer from stratch.