Train on a different dataset

[jay98]

Hey @TimoSaemann ,
I'm confused as of how to use train Enet on a different dataset. I'm trying to train it on KITTI data set.
Thank You

[TimoSaemann]

Have a look into the tutorial: https://github.com/TimoSaemann/ENet/tree/master/Tutorial

[Viswa14]

Hey @TimoSaemann : While training on a different dataset (with different resolution of images) What will be the necessary modifications ? Is there an example for the same ? Thanks a lot

[kli-casia]

Same question

[nk-dev0]

@Viswa14 @kli-nlpr I basically followed the tutorial, but for changing image resolution just change the first layer in the train prototxt to your desired size, and the label divide factor if necessary. Also you'll want to change the num_output param in deconv6_0_0 to your number of classes

[Viswa14]

@nkrall : Thanks for you help!
Did you by any chance encounter a protobuf version mis-match error while compiling the code ? When I compile the caffe with default protobuf 2.x, the compilation is successful but when I execute Training script, I get an error stating Python layer requires protobuf 3.X and if I upgrade my protoc and protobuf version, the caffe compilation fails. Any Suggestion ?

[nk-dev0]

@Viswa14 hmm, no I didn't run into that

[changlinzhang]

@nkrall May I ask how you edit other files? I have edited absolute paths and train_file.txt. Because my the number of new dataset classes is 30, I also changed class numbers. But it didnt seem to converge. (and create_colormap.py needed to be edited? My new dataset doesnt seem to have RGBs of classes)

[jianbozhanghit]

@nkrall : Thanks for you help!
Did you by any chance encounter a protobuf version mis-match error while compiling the code ? When I compile the caffe with default protobuf 2.x, the compilation is successful but when I execute Training script, I get an error stating Python layer requires protobuf 3.X and if I upgrade my protoc and protobuf version, the caffe compilation fails. Any Suggestion ?

@Viswa14 Did you solve the problem? thanks