Has anyone been able to train a different network architecture?

[louisquinn]

I am considering how I can modify this framework to train a different network, say Inception V3.

I know I have to modify the VGGnet_test.py and VGGnet_train.py files to match Inception's architecture (can be found here: https://github.com/tensorflow/models/blob/master/inception/inception/slim/inception_model.py)

Also the tensor to take the Inception's features from (I think is) mixed_10/join (see pic below). Has anyone had any success with this? Or could provide some insight to what else could be required?

[lk251]

Not yet, but this error comes to mind, there might be others https://github.com/smallcorgi/Faster-RCNN_TF/pull/83/commits/84c52bac6d1d16c727ce7120acf9d1773b291d51

Happy to join in this endeavor after I manage to use a different dataset (I expect a bumpy road, eg. n_classes is fixed to 21 in VGGnet_train.py

[cemysf]

I copied the original _VGGnettrain.py and _VGGnettest.py in lib/networks with different names After that, I modified these files (only difference that I've made was the number of output classes) Then I edit _getnetwork function for these in the factory.py in lib/networks

Also, you may need to construct a new dataset handler in lib/datasets and alter train.py and test.py in lib/fast_rcnn