Execution Time0.12s

Test: TMVA-DNN-BatchNormalization (Passed)
Build: master-x86_64-fedora29-gcc8-dbg (root-fedora29-3.cern.ch) on 2019-11-14 11:43:28

Test Timing: Passed
Processors1

Show Command Line
Display graphs:

Test output
Testing Backpropagation:
DEEP NEURAL NETWORK:   Depth = 3  Input = ( 1, 10, 4 )  Batch size = 10  Loss function = R
	Layer 0	 DENSE Layer: 	 ( Input =     4 , Width =     2 ) 	Output = (  1 ,    10 ,     2 ) 	 Activation Function = Identity
	Layer 1	 BATCH NORM Layer: 	 ( Input =     2 ) 
	Layer 2	 DENSE Layer: 	 ( Input =     2 , Width =     1 ) 	Output = (  1 ,    10 ,     1 ) 	 Activation Function = Identity
input 

10x4 matrix is as follows

     |      0    |      1    |      2    |      3    |
---------------------------------------------------------
   0 |     0.9989     -0.4348      0.7818    -0.03005 
   1 |     0.8243    -0.05672     -0.9009     -0.0747 
   2 |   0.007912     -0.4108       1.391     -0.9851 
   3 |   -0.04894      -1.443      -1.061      -1.388 
   4 |     0.7674      -0.736      0.5797     -0.3821 
   5 |      2.061      -1.235       1.165     -0.4542 
   6 |    -0.1348     -0.4996     -0.1824       1.844 
   7 |    -0.2428       1.997    0.004806     -0.4222 
   8 |      1.541     0.09474       1.525       1.217 
   9 |    -0.1363     -0.1992     -0.2938     -0.1184 

 training batch 1 mu var00.0968956
output DL 

10x2 matrix is as follows

     |      0    |      1    |
-------------------------------
   0 |     0.3485      -2.286 
   1 |    -0.8521       1.286 
   2 |      1.777      -3.419 
   3 |     -0.641       0.642 
   4 |     0.3061      -2.109 
   5 |     0.4473      -4.214 
   6 |     -1.927       1.008 
   7 |      1.533       1.033 
   8 |     0.2405       -3.13 
   9 |    -0.2628       0.476 

output BN 
output DL feature 0 mean 0.0968956	output DL std 1.09574
output DL feature 1 mean -1.07133	output DL std 2.15562
output of BN 

10x2 matrix is as follows

     |      0    |      1    |
-------------------------------
   0 |      0.242     -0.5941 
   1 |    -0.9129       1.153 
   2 |      1.616      -1.148 
   3 |    -0.7098      0.8378 
   4 |     0.2012     -0.5074 
   5 |     0.3371      -1.537 
   6 |     -1.947       1.017 
   7 |      1.381       1.029 
   8 |     0.1382      -1.007 
   9 |    -0.3461      0.7567 

output BN feature 0 mean 1.11022e-17	output BN std 1.05404
output BN feature 1 mean 1.33227e-16	output BN std 1.05408
Testing weight gradients   for    layer 0
weight gradient for layer 0

2x4 matrix is as follows

     |      0    |      1    |      2    |      3    |
---------------------------------------------------------
   0 |    0.01216     -0.0119      0.0129    0.003798 
   1 |   0.009467    -0.01023    0.008976    0.004105 

weights for layer 0

2x4 matrix is as follows

     |      0    |      1    |      2    |      3    |
---------------------------------------------------------
   0 |    -0.1087      0.5813      0.8768     -0.8087 
   1 |    -0.4985      0.5603      -1.958      0.4685 

 training batch 2 mu var00.0968984
compute loss for weight  -0.10869  -0.1087 result 0.0407086
 training batch 3 mu var00.0968956
compute loss for weight  -0.10871  -0.1087 result 0.0407083
 training batch 4 mu var00.0968963
compute loss for weight  -0.108695  -0.1087 result 0.0407085
 training batch 5 mu var00.0968956
compute loss for weight  -0.108705  -0.1087 result 0.0407084
   --dy = 0.0121605 dy_ref = 0.0121605
 training batch 6 mu var00.0968951
compute loss for weight  0.581276  0.581266 result 0.0407083
 training batch 7 mu var00.0968956
compute loss for weight  0.581256  0.581266 result 0.0407086
 training batch 8 mu var00.0968954
compute loss for weight  0.581271  0.581266 result 0.0407084
 training batch 9 mu var00.0968956
compute loss for weight  0.581261  0.581266 result 0.0407085
   --dy = -0.0119007 dy_ref = -0.0119007
 training batch 10 mu var00.0968959
compute loss for weight  0.876843  0.876833 result 0.0407086
 training batch 11 mu var00.0968956
compute loss for weight  0.876823  0.876833 result 0.0407083
 training batch 12 mu var00.0968957
compute loss for weight  0.876838  0.876833 result 0.0407085
 training batch 13 mu var00.0968956
compute loss for weight  0.876828  0.876833 result 0.0407084
   --dy = 0.0129046 dy_ref = 0.0129046
 training batch 14 mu var00.0968956
compute loss for weight  -0.808685  -0.808695 result 0.0407085
 training batch 15 mu var00.0968956
compute loss for weight  -0.808705  -0.808695 result 0.0407084
 training batch 16 mu var00.0968956
compute loss for weight  -0.80869  -0.808695 result 0.0407085
 training batch 17 mu var00.0968956
compute loss for weight  -0.8087  -0.808695 result 0.0407084
   --dy = 0.0037976 dy_ref = 0.0037976
Testing weight gradients   for    layer 1
weight gradient for layer 1

1x2 matrix is as follows

     |      0    |      1    |
-------------------------------
   0 |    0.05135     0.03007 

weights for layer 1

1x2 matrix is as follows

     |      0    |      1    |
-------------------------------
   0 |          1           1 

 training batch 18 mu var00.0968956
compute loss for weight  1.00001  1 result 0.040709
 training batch 19 mu var00.0968956
compute loss for weight  0.99999  1 result 0.0407079
 training batch 20 mu var00.0968956
compute loss for weight  1.00001  1 result 0.0407087
 training batch 21 mu var00.0968956
compute loss for weight  0.999995  1 result 0.0407082
   --dy = 0.0513476 dy_ref = 0.0513476
 training batch 22 mu var00.0968956
compute loss for weight  1.00001  1 result 0.0407087
 training batch 23 mu var00.0968956
compute loss for weight  0.99999  1 result 0.0407081
 training batch 24 mu var00.0968956
compute loss for weight  1.00001  1 result 0.0407086
 training batch 25 mu var00.0968956
compute loss for weight  0.999995  1 result 0.0407083
   --dy = 0.0300693 dy_ref = 0.0300693
Testing weight gradients   for    layer 1
weight gradient for layer 1

1x2 matrix is as follows

     |      0    |      1    |
-------------------------------
   0 |  -4.77e-18   2.602e-18 

weights for layer 1

1x2 matrix is as follows

     |      0    |      1    |
-------------------------------
   0 |          0           0 

 training batch 26 mu var00.0968956
compute loss for weight  1e-05  0 result 0.0407084
 training batch 27 mu var00.0968956
compute loss for weight  -1e-05  0 result 0.0407084
 training batch 28 mu var00.0968956
compute loss for weight  5e-06  0 result 0.0407084
 training batch 29 mu var00.0968956
compute loss for weight  -5e-06  0 result 0.0407084
   --dy = 0 dy_ref = -4.77049e-18
 training batch 30 mu var00.0968956
compute loss for weight  1e-05  0 result 0.0407084
 training batch 31 mu var00.0968956
compute loss for weight  -1e-05  0 result 0.0407084
 training batch 32 mu var00.0968956
compute loss for weight  5e-06  0 result 0.0407084
 training batch 33 mu var00.0968956
compute loss for weight  -5e-06  0 result 0.0407084
   --dy = 1.15648e-13 dy_ref = 2.60209e-18
Testing weight gradients   for    layer 2
weight gradient for layer 2

1x2 matrix is as follows

     |      0    |      1    |
-------------------------------
   0 |     -0.372      0.3278 

weights for layer 2

1x2 matrix is as follows

     |      0    |      1    |
-------------------------------
   0 |     -0.138     0.09174 

 training batch 34 mu var00.0968956
compute loss for weight  -0.138036  -0.138046 result 0.0407047
 training batch 35 mu var00.0968956
compute loss for weight  -0.138056  -0.138046 result 0.0407122
 training batch 36 mu var00.0968956
compute loss for weight  -0.138041  -0.138046 result 0.0407066
 training batch 37 mu var00.0968956
compute loss for weight  -0.138051  -0.138046 result 0.0407103
   --dy = -0.371961 dy_ref = -0.371961
 training batch 38 mu var00.0968956
compute loss for weight  0.0917483  0.0917383 result 0.0407117
 training batch 39 mu var00.0968956
compute loss for weight  0.0917283  0.0917383 result 0.0407052
 training batch 40 mu var00.0968956
compute loss for weight  0.0917433  0.0917383 result 0.0407101
 training batch 41 mu var00.0968956
compute loss for weight  0.0917333  0.0917383 result 0.0407068
   --dy = 0.327772 dy_ref = 0.327772
Testing weight gradients:      maximum relative error: [NON-XML-CHAR-0x1B][32m2.9569e-10[NON-XML-CHAR-0x1B][39m