out of memory issue when generating perturbations with VGG and ResNet?

[auzyze]

Hi,
Thanks for your code, I learn a lot from it.

When I run the python code with inception5h model, it works very well and I can reproduce perturbation maps. However, when I try the code with VGG or ResNet models, I always get GPU memory full errors. I used the tensorflow SLIM pretrained models, and freeze them into .pb file, just like the format of inception5h model in your example. Do you have any idea to work around this issue? I use a K40C 11G Mem GPU.

Thanks!

[husseinfawzi]

Hi,

Did you modify the lines persisted_input = persisted_sess.graph.get_tensor_by_name("input:0") and persisted_output = persisted_sess.graph.get_tensor_by_name("softmax2_pre_activation:0") to the layer names of your new network?

Besides, if you are using the older version of the code, num_classes is set to 1001. The other networks might be using another convention (1000 labels only), so you might have to modify this too.

[auzyze]

Thanks for the comments, yes, I've modified tensor name accordingly:

VGG19:

persisted_input = persisted_sess.graph.get_tensor_by_name("input:0")
persisted_output = persisted_sess.graph.get_tensor_by_name("vgg_19/fc8/squeezed:0")

ResNet-50:

persisted_input = persisted_sess.graph.get_tensor_by_name("input:0")
persisted_output = persisted_sess.graph.get_tensor_by_name("resnet_v2_50/SpatialSqueeze:0")

num_classes of VGG model in slim library is 1000, so I set:

scalar_out = [tf.slice(persisted_output, [0, i], [1, 1]) for i in range(0, 1000)]
dydx = [tf.gradients(scalar_out[i], [persisted_input])[0] for i in range(0, 1000)]

[husseinfawzi]

Apologies for the (very) late reply.
Could you try to compute perturbations for the following VGG model https://s3.amazonaws.com/cadl/models/vgg16.tfmodel
Input is images:0 and output is fc8:0. I could not reproduce your error with this model (GPU Titan XP).
Please use the latest version of the code.