Awesome

FCN-pytorch-easiest

Trying to be the easiest FCN pytorch implementation and just in a get and use fashion

Here I use a handbag semantic segmentation for illustration on how to train FCN on your own dataset and just go to use. To train on your own dataset you just need to see in BagData.py which implements a dataloader in pytorch. What you actually need to do is providing the images file and the correspoding mask images. And for visualization in the training process I use visdom.

requirement

I have tested the code in pytorch 0.3.0.post4 in anaconda python 3.6 in ubuntu 14.04 with GTX1080 in cuda8.0

train

here three images pair is provided in folder last/ and last_msk/ . Here I want to do a handbag semantic segmentation which is stated as belows.

Firstly because visdom is used to visualize the training process, you need open another terminal and run

python -m visdom.server

Then you run in another terminal

python FCN.py

You can open your browser and goto localhost:8097 to see the visulization as following the first row is the prediction.

deploy

and for deploy and inference I also provide a script inference.py. You should be careful about the model path. Bacause I did not provide the trained weights file. :-P

BTW, FCN.py is copy from other repo.