Is error rates top-1 or top-5 in experiment result?

[haikuoyao]

Hi there,

I try to reproduce experiment result.

Sorry, this might be a silly question. Is error rates top-1 or top-5 in this table?

Thanks.

[taineleau-zz]

TOP 1.

[haikuoyao]

@taineleau Thank you. May I ask another question? In paper, we compute the average (absolute) weight to conduct an experiment on Feature Reuse. What is exactly the weight? And how do we compute it? Thanks a lot.

[taineleau-zz]

@haikuoyao Let me try to answer your question. @liuzhuang13 Please correct me if I am wrong. Basically, the weight of a convolution layer is of shape (n_input_planes, n_output_plane, filter_width, filter_height). If we are looking at a specific layer l, say l = 3, this layer accepts features from layer 0, 1 and 2, which are concatenated as the input to layer 3.

Suppose we store the weight in w (a python variable). (Note that, the number of channel of layer 0 is 24, and the growth rate k is 12). Hence, the weight for the grid (s=0, l=3) is w[:24, :, :, :]; For (s=1, l=3), it's w[24:36, :, :, :]; For (s=2, l=3), it's w[36:48, :, :, :].

[haikuoyao]

Thanks a lot @taineleau . Sorry to ask another question. How long did it take to classify one single image with densenet-161.t7 model by classify.lua in you guys' experiment? It takes around 2s here on GPU. Seems it's right. It shouldn't take so long, right?

[haikuoyao]

oh... When I classified more than one image, it just took a little bit long on the first image and quite fast on rest images. Seems it's okay. Sorry to bother you.

[taineleau-zz]

@haikuoyao You're right. For the first batch, the chips need some more time to get prepared. I believe if you take a look at the data time, it takes up most of the time when you forward the first batch.

[haikuoyao]

@taineleau Thanks a lot. I'm gonna close this issue. :)