Faster-rcnn.pytorch: train on own dataset

[wjx2]

Hi,
If I want to use your code to train on our dataset, where should I modify?

[Worulz]

@wjx2

fastest option is to use the pascal voc pattern. I recommend downloading the dataset and checking the paths. Then place your data in the same format.

Another option, if you want to create a more elegant solution. you can find out how it finds the file paths if you execute and print code from lib/datasets/pascal_voc.py. if you'd like to make a new imdb dataset add a line in train_net.py similar to how dataset name is collected and similarly add it to lib/datasets/factory.py then follow the pascal_voc.py format and create a new script with your dataset name. the python print and input function will be your best friend.

[wjx2]

@Worulz I couldn't find the network, I mean that in caffe faste rcnn, I should modify num_class etc. But I couldn't find this project to modify. And I use the pascal voc pattern as you recommend.

[Worulz]

@wjx2 it's hard coded in this line. This will feed the classes into the train and test script. but not the demo.

When you use the demo.py to see predictions. Make sure to change the classes here.

[wjx2]

@Worulz Thank you .
I have modify the file you mentioned above. And should I modify the network to adapt my datasets?(such as my dataset is not 20 classes).Like the py-faster-rcnn, we also should change the num_out and num_classes in the network. And I can't find where the network shoul modify in this project.

[Worulz]

I believe it should be this line. It looks like it's calling the new number of classes, so you should be good. I've been able to run it with my own dataset with the changes I'm mentioned above.

just as a sanity check I would recommend playing with this line.. if you have 20 classes (included background). change it to say 50?. And see if any error (relating to mis-matched dimensions) occurs when you run the train script. If it does, then it does automatically extracts your new number of classes from your dataset script.

[wjx2]

@Worulz Thank you!