different Accuracy between during training and real output via ipynb. (file attached)

[전호진]

Hi, I'm trying to use caffe for object detection and have several failure for weeks,

I have made training set and test(validation) set by create_lmdb_isBall.sh (which is modified version of caffe imagenet example).

and mean file by make_isBall_mean.sh(which is modified version of caffe imagenet example).

training set has 4 category and 1 false category (0~4:, noBall, yesBall, NULL, fire, pillar) 

and test(validation) set has 2 category( 0~1: noBall, yesBall). 

during training, test net output (accuracy) shows 0.93(93 percent) from 40k.

for me, this percentage is pretty good. so after seeing this, I ran training until 140k and stopped it.

(accuracy was also 0.93)

after this, for classification of this network, 

I used 3636Net_onClick.ipynb for classification .(which is modified version of caffe imagenet example, 00-classification.ipynb).

(data for classification is same as training set.)

but result was really, really bad.

overall classification rate is 0.2~0.43 now.

why this is happening?

for more information, I attached my prototxt, log, ipynb, sh files and 40k and 140k result.