Detectron: Accessing net's output

[raviv]

Hi,

How can I access the network's outputs after each iteration?
I assume this would be somehow available after each call to
workspace.RunNet(model.net.Proto().name)

My end goal is to visualize inference once every few iterations to "debug" the net.

Thanks!

[rbgirshick]

You can fetch blobs from the Caffe2 workspace, which will copy the GPU tensor into a numpy ndarray. For example, this line fetches the gpu_0/rois blob which contains the region of interest predictions from the RPN part of the network.

[raviv]

Thanks @rbgirshick.
Can I access the inputs (images) the same way? I'd like to see the effect of augmentation on my images.
If so, what would be the blobs's name for that?
Also, what's the best place in the code to inject my custom augmentations before net forward?

Appreciate the help!

[manyaafonso]

Hi @raviv , would you be willing to share your modified code for this visualization? Thanks.

[raviv]

@manyaafonso, I eventually ditched Detectron so can't help you with that.

[manyaafonso]

@raviv thanks for getting back quickly. A coworker of mine also ditched detectron for matterport's maskrcnn, so I'll give that a try too.