How to use your "ResNet152Body" function? about caffe HOT 6 CLOSED

The script is not fully tested for model which is trained with batch normalization (BN), you should make adjust to certain settings if you want to try ResNet.

For your specific problems, you should change mbox_source_layers to layer names you want to use for prediction in ResNet152. And you should use ResNet152Body(net, from_layer='data', use_pool5=False).

from caffe.

yes, i already changed "mbox_source_layers" part.
I try to change layer names(many time) for prediction in ResNet152. but I have still a error .

I0716 11:52:04.729372 1913 net.cpp:752] Ignoring source layer fc7_relu7_0_split
F0716 11:52:04.730145 1913 net.cpp:765] Cannot copy param 0 weights from layer 'conv6_1'; shape mismatch. Source param shape is 256 1024 1 1 (262144); target param shape is 256 2048 1 1 (524288). To learn this layer's parameters from scratch rather than copying from a saved net, rename the layer.

thank you.

from caffe.

It is weird that you have fc7_relu7_0_split, which is not part of ResNet152 (assuming you download it from here). Make sure you are not loading from a previously saved wrong snapshot.

from caffe.

@weiliu89
thank you.
this problem remained previously logs(model...etc). I delete this log.
i try to again.
fortunately, above error not generating!

when resnet execute for training, it generate out of gpu memory.
but, it not generate when I excute using 8 core gpu.
i running(train) and yet not out of memory.

and
I have a questions.
l wonder optimal layer names to use for prediction in ResNet152(using mbox_source_layers variable).
how to select optimal ResNet152 ' layer name?(only test??)

sorry, I am not good at English. and Thank you for the reply.

from caffe.

ResNet has a problem of consuming a lot of memory. I think usually people use ResNet101 for detection. According to my experience, it's performance is similar to VGG (at least on VOC0712) within the SSD framework.

For optimal layers, you can find corresponding layers which has similar feature map size as the ones used in VGG.

from caffe.

tankyou~^^

from caffe.