Access to intermediate layer TridentNet

[simaiden]

I want to access to an intermediate layer of TridentNet to get a feature vector, in particular feature vector after doing the roi pool operation, how can i perform this?

Thanks

[RogerChern]

Our wrapper library mxnext provides a mxnext.foward_debug operator that accepts a callback with signature callback(gpu_id, num_iter, *data).

You can add the following to your Head

def save_roi_feat(gpu_id, num_iter, roi_feat):
    if gpu_id == 0 and num_iter + 1 % 100 == 0:
        np.save("trident_roi_feature_%d" % int(num_iter), roi_feat)

roi_feat = X.forward_debug(roi_feat, callback=save_roi_feat)

[simaiden]

Thanks for reply! Sorry about the dumb question but is RpnHead the Head
intridentnet_r50v2c4_c5_1x.py ? or may I have to save roi_extractor = RoiExtractor(RoiParam)?

Thanks in advance

[RogerChern]

I think that inserting the hook at the input of BboxC5Head may be a good idea.

https://github.com/TuSimple/simpledet/blob/f67fca08b33d957b43eb1db5e3745a3a9e3803b5/symbol/builder.py#L558-L579

You can put the following snippet in just before L564

def save_roi_feat(gpu_id, num_iter, roi_feat):
    if gpu_id == 0 and num_iter + 1 % 100 == 0:
        import numpy as np
        np.save("trident_roi_feature_%d" % int(num_iter), roi_feat)

conv_feat = X.forward_debug(conv_feat, callback=save_roi_feat)

And you can find trident_roi_feat_XXX.npy in your directory.

[simaiden]

I have an error:

RuntimeError: num_iter_b5f728d4-dac5-4563-be3a-0a3f8a5c690a is not presented

Maybe the code you posted ir just for training? How can i get the roi feats in inference?

Thanks!

[RogerChern]

A known issue of the debug utility. I will look into it. In the meantime, you may set your network for training and lr and wd to 0 as a workaround.

…

On Thu, Jan 2, 2020 at 1:26 AM Simón Sepúlveda Osses < ***@***.***> wrote: I have an error: RuntimeError: num_iter_b5f728d4-dac5-4563-be3a-0a3f8a5c690a is not presented Maybe the code you posted ir just for training? How can i get the roi feats in inference? Thanks! — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#285>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABGODH7LLEGZ5KJZEIINKJLQ3TG4TANCNFSM4KASMTUQ> .

[simaiden]

Thanks @RogerChern , I'll give a try