how to use the crop_max_patches.py?

[visonpon]

how to use the crop_max_patches.py? can you give examples of the script usage like use the optimize_image.py script? @yosinski

[visonpon]

assert args.do_maxes or args.do_deconv……,'specify at least one do_*option to output.' what's that mean to specify? @yosinski

[tsungjenh]

@visonpon Im stuck at find_max_acts.py. Did you successfully run it? If I use cifar10 dataset do I need to transform the binary to jpg then put it all together in to a folder before I run this script or? And also what does it mean parser.add_argument('outfile', type = str, help = 'output filename for pkl')?? Please help!!!

[heitorrapela]

"@visonpon can you give examples of the script usage like use the optimize_image.py script?" Answer this question, you can see this link: ./optimize_image.py --decay 0.0001 --blur-radius 1.0 --blur-every 4 --max-iter 1000 --lr-policy constant --lr-params "{'lr': 100.0}"

You can change the parameters too. There are many parameters like place to save (defult: optimize_results/opt, always change this parameter, or you will not generate the desired output, by an error of trying to overwrite previous files - I get this error after 1k ite, something like one hour running the script) and neuron to pick up (default: 130)... I'm using data-size (224,224) to run the googLeNet and if you have some errors you can try stackoverflow like this one link, where you need to modify this file: caffe/python/caffe/io.py

For your first question, i'm trying to run it... xD

[heitorrapela]

@visonpon

1. how to use the crop_max_patches.py? can you give examples of the script usage like use the optimize_image.py script?

• python crop_max_patches.py output.pkl '../models/bvlc-googlenet/deploy.prototxt' '../models/bvlc-googlenet/bvlc_googlenet.caffemodel' '../input_images/' 'out.txt' 'outputFolder' 'pool5/7x7_s1' --do-deconv --do-maxes --N 9
• "Must be" : positional arguments: (List generated with python crop_max_patches --h) nmt_pkl Which pickled NetMaxTracker to load. (output of the find_max_acts.py) net_prototxt Network prototxt to load (see the script) net_weights Network weights to load (see the script) datadir Directory to look for files in (Folder to input images) filelist List of image files to consider, one per line. (**To understantd that I saw that in the code you need something like "filename ouf input image" + "label", so I created a .txt file with each image and label per line **)

2. assert args.do_maxes or args.do_deconv……,'specify at least one do_*option to output.' what's that mean to specify? --do-maxes Output max patches. --do-deconv Output deconv patches. --do-deconv-norm Output deconv-norm patches. --do-backprop Output backprop patches. --do-backprop-norm Output backprop-norm patches. --do-info Output info file containing max filenames and labels. My choice was --do-deconv --do-maxes with --N 9

@pawanhsu
"Im stuck at find_max_acts.py. Did you successfully run it? If I use cifar10 dataset do I need to transform the binary to jpg then put it all together in to a folder before I run this script or? And also what does it mean parser.add_argument('outfile', type = str, help = 'output filename for pkl')?? Please help!!!"

1. You need to use the .jpg image that you feed your net.
2. parser.add_argument('outfile', type = str, help = 'output filename for pkl') - (output of the find_max_acts.py)

Some results that i'm working with googLeNet - Output of pool5/7x7_s1 (I stop the script in the begin because I need to go out, it generated something like 500+- images from 9.5k+-, .

- Unit0/deconv_002.jpg - Unit0/maxim_002.jpg - Unit0051/deconv_004.jpg - Unit0051/maxim_004.jpg