Closed twongjirad closed 6 years ago
confirmed that the values in the data blobs are completely different. showing plot. seems there is an inversion in the wire and time axes somewhere. maybe when i made the image2d file.
despite image2d and lmdb difference I was able to "remove the neutrino" with the zero tool and the network probability changed
before
after
commit is on the rgb_caffemod branch btw https://github.com/LArbys/LArCV/commit/e4d5342df5b200909b45636cbd73b35d1aac9dcd
MATCHES!!!
super !
committed changes @ https://github.com/LArbys/LArCV/commit/c9a268b7b148b0520c32c059cac565df6d4091c5
NOTE: I turned off the thresholding functionality for the 12 channel data
Also disabled the additional mean subtraction.
The order of thresholding and mean subtraction has to be discussed.
the scores from the gui interface was matched to a python script used to run the BNB. This turns out to be different from the scores when I run the BNB data through the network using 'caffe test'.
New targets:
I0506 01:00:48.000690 18747 caffe.cpp:275] Batch 0, probt = 0.230329 I0506 01:00:48.000699 18747 caffe.cpp:275] Batch 0, probt = 0.769671 I0506 01:00:48.046295 18747 caffe.cpp:275] Batch 1, acc = 1 I0506 01:00:48.046342 18747 caffe.cpp:275] Batch 1, probt = 0.966876 I0506 01:00:48.046350 18747 caffe.cpp:275] Batch 1, probt = 0.033124 I0506 01:00:48.087106 18747 caffe.cpp:275] Batch 2, acc = 1 I0506 01:00:48.087162 18747 caffe.cpp:275] Batch 2, probt = 0.600837 I0506 01:00:48.087169 18747 caffe.cpp:275] Batch 2, probt = 0.399163 I0506 01:00:48.131290 18747 caffe.cpp:275] Batch 3, acc = 1 I0506 01:00:48.131337 18747 caffe.cpp:275] Batch 3, probt = 0.61427 I0506 01:00:48.131343 18747 caffe.cpp:275] Batch 3, probt = 0.38573 I0506 01:00:48.172060 18747 caffe.cpp:275] Batch 4, acc = 1 I0506 01:00:48.172116 18747 caffe.cpp:275] Batch 4, probt = 0.618634 I0506 01:00:48.172122 18747 caffe.cpp:275] Batch 4, probt = 0.381366
So I verified that the network data for both the GUI and 'caffe test' mode are the same. The sum of diffs for all pixels is zero. Still, the network output is different by a few percent. WTF.
I believe this is resolved, or please reopen if it persists!
But at least the forward pass is working.
First thing to check is if the image in the datum is in the same order as the data in the Image2D object.
Check the mean blob, too.