Whats the range of first layer filters in my .caffemodel

[jim D.]

Hi everyone,

First, I would like to ask a simple question. Caffe handles images in range [0,255] and not in [0,1] right? I think thats why scaling factor exists

Now the important part. I am using caffe on linux only from console. I dont use python and I had some trouble to find how to analyze the .caffemodel file. But right now I really need some information from it. I have a .caffemodel file which is actually the outcome of Alexnet for my dataset. There are only 500 repetitions. I just need to know: Whats the range of first layer filters, after that repetitions.

I wanna use predefined kernels in an other net of mine (in the first layer), those kernels wont be trained. So I really want to know in what range I should initialize my kernel values, given my current dataset. I also post below, a link of my megasync where you can find a .zip file with .caffemodel and .prototxt train_val.

I really need this information asap guys. If someone of you could open that .caffemodel and tell in what range are first layer's values of kernels it would help me a lot. Of course all information you can give me about how to install the required python version and dependencies and how I can use caffe headers inside python and maybe a script to do the work above are welcome.

But this information is very crucial for me and anyone who can download that .caffemodel file (.prototxt is also contained) and answer me asap, he would really help me a lot.

.zip file (.caffemodel+.prototxt) on :

https://mega.nz/#!moUghIrA!fJ0pV_H98MXEJd5dnM8ypnt8iJB1sSg4dpEOkwajkqM

[Przemek D]

Per your first question: that's correct, images are loaded in [0;255] range.

Installing python is straightforward, you want to use either the 2.7 version or the current newest - just follow the manual how to compile pycaffe (see the Python and/or MATLAB Caffe section).
Extracting kernel weights is simpler than you think. This example will guide you through the process.