Skip to content

testconvnet-usage.org

Latest commit

 

History

History
73 lines (67 loc) · 2.89 KB

testconvnet-usage.org

File metadata and controls

73 lines (67 loc) · 2.89 KB

command

–show-estimation=fc\_j2 [–estimation-type=AHEBuffy]

–save-images

–save-estimation

–show-cost=, –cost-idx=

–analyze-feature-name –feature-channel

–save-res-patch=all/average/allpatchdata/allpatchdata-feature –save-feature-path=

–calc-hist=actjoint [–save-feature-path]=

example

--analyze-feature-name=conv3,2 --feature-channel=64 --calc-hist=actjoint,occ@any_number
--save-feature-path=somewhere

2 is the idx of indicatormap in the data list

–show_response=random

This is average response of all the filers for some random images

–save-response=separate

Show response for each images for all the filters

–save-indmap-show=all –save-feature-path [–matched-path –ind-type=]

note

matched is used for removing unmatched ind-type indicate whether the type of indicator map joint/part

example

python testconvnet.py -f /media/SYSTEM/Storage/saved/backup/ConvNet__2013-10-22_20.42.46 --data-path=/media/DATA/pure_shuffled_batch_train_modified128 --analyze-feature-name=fc_i2 --feature-channel=7 --save-indmap-show=all --save-feature-path=/media/SYSTEM/Storage/imagedata/results/indmap_20000 --mini=32 --test-range=20000

–ubd-image-folder= –save-feature-path [–ubd-fix-input-var==]

–save-cost-path

example

python testconvnet.py -f /media/SYSTEM/Storage/saved/backup/Clusters/c8k19/ConvNet__2013-12-30_10.13.54  --save-cost-path=/media/SYSTEM/Storage/imagedata/results/track/2013-12-28/costs/cost_macid_19_type_8

–save-feature-name –forward-pass-feature=1 –save-feature-path

–forward-pass-feature=1, then just forward pass the feature (joints8) otherwise, it will save the estimated pose in images

–do-evaluation=fc\_j2 [–evaluation-type=mpjpe] [–save-evaluation=]

python testconvnet.py -f /opt/visal/tmp/for_sijin/Data/saved/c8k16/ConvNet__2014-06-13_12.55.33 --do-evaluation=fc_j2,humaneva_body,1 --test-one=0

Note

reorder\_channel

Only meaningful for net10-22(net10-24)

# for conv2
     clist = [15, 2, 0, 5, 8,11, 1, 7,\
              10,13, 4,12, 6,14, 9, 3] 
# for conv3
     clist = [ 9, 2, 8,12,4, 7, 0, 3,\
              15, 6,14, 5,1,11,10,13]

Useful

Current Best Network

/media/SYSTEM/Storage/saved/backup/ConvNet__2013-10-22_20.42.46

–data-path=/media/DATA/valid_re_shuffled_batch_train_modified128 –data-path=/media/DATA/pure_shuffled_batch_train_modified128

layer_list = [(9,1),(3,2),(5,1),(3,2), (5,1),(3,2)]
d = {'conv1':0, 'pool1':1, 'conv2':2,'pool2':3,'conv3':4,\
                 'pool3':5}
            extra_d = {'fc_i2':-1}

type4=c8k15/ConvNet__2013-12-13_20.08.29

iteration = 407.89 –data-path=/media/DATA/pos_neg_shuffled_train128

type7=c8k16/ConvNet__2013-12-28_22.54.22

iteration = 357.49 –data-path=/media/DATA/pos_neg_shuffled_train128