0.50s

Processors | 1 |
---|

Show Command Line

Display graphs:

Testing CNN Backward Pass: Test1, backward pass with linear activation network - compare with finite difference added Conv layer 2 x 5 x 5 added Conv layer 2 x 3 x 3 added MaxPool layer 2 x 2 x 2 Do Forward Pass Do Backward Pass Testing weight gradients: layer: 0 / 6 Weight gradient from back-propagation - vector size is 1 Activation gradient from back-propagation - vector size is 2 Layer 0 : output D x H x W 2 5 5 input D x H x W 2 4 4 layer output size 2 Evaluate the Derivatives with Finite difference and compare with BP for Layer 0 0 - 0 , 0 : -7.18855 from BP -7.18855 2.00335e-11 0 - 0 , 1 : -18.557 from BP -18.557 2.1896e-12 0 - 0 , 2 : 31.0615 from BP 31.0615 1.80472e-11 Testing weight gradients: layer: 1 / 6 Weight gradient from back-propagation - vector size is 1 Activation gradient from back-propagation - vector size is 2 Layer 1 : output D x H x W 2 3 3 input D x H x W 2 5 5 layer output size 2 Evaluate the Derivatives with Finite difference and compare with BP for Layer 1 0 - 0 , 0 : 5.04341 from BP 5.04341 1.72622e-10 0 - 0 , 1 : -54.9362 from BP -54.9362 1.29486e-11 0 - 0 , 2 : 22.3178 from BP 22.3178 4.77348e-11 Testing weight gradients: layer: 2 / 6 Weight gradient from back-propagation - vector size is 1 Activation gradient from back-propagation - vector size is 2 Layer 2 : output D x H x W 2 2 2 input D x H x W 2 3 3 layer output size 2 Evaluate the Derivatives with Finite difference and compare with BP for Layer 2 0 - 0 , 0 : 0 from BP 0 0 0 - 0 , 1 : 0 from BP 0 0 0 - 0 , 2 : 0 from BP 0 0 Testing weight gradients: layer: 3 / 6 Layer 3 has no weights Activation gradient from back-propagation - vector size is 1 Layer 3 : output D x H x W 1 1 8 input D x H x W 2 2 2 layer output size 1 Evaluate the Derivatives with Finite difference and compare with BP for Layer 3 Testing weight gradients: layer: 4 / 6 Weight gradient from back-propagation - vector size is 1 Activation gradient from back-propagation - vector size is 1 Layer 4 : output D x H x W 1 1 3 input D x H x W 1 1 8 layer output size 1 Evaluate the Derivatives with Finite difference and compare with BP for Layer 4 0 - 0 , 0 : -14.3381 from BP -14.3381 2.59206e-11 0 - 0 , 1 : -15.986 from BP -15.986 9.62831e-12 0 - 0 , 2 : -14.3381 from BP -14.3381 2.59206e-11 Testing weight gradients: layer: 5 / 6 Weight gradient from back-propagation - vector size is 1 Activation gradient from back-propagation - vector size is 1 Layer 5 : output D x H x W 1 1 1 input D x H x W 1 1 3 layer output size 1 Evaluate the Derivatives with Finite difference and compare with BP for Layer 5 0 - 0 , 0 : 52.6036 from BP 52.6036 1.33035e-12 0 - 0 , 1 : -14.99 from BP -14.99 6.89569e-12 0 - 0 , 2 : -2.0514 from BP -2.0514 7.69305e-11 Testing weight gradients: maximum relative error: [NON-XML-CHAR-0x1B][32m7.48414e-10[NON-XML-CHAR-0x1B][39m Test2, more complex network architecture no dropout added Conv layer 12 x 7 x 7 added Conv layer 6 x 5 x 5 added MaxPool layer 6 x 3 x 3 Do Forward Pass Do Backward Pass Testing weight gradients: layer: 0 / 6 Weight gradient from back-propagation - vector size is 1 BP Weight Gradient ( 12 x 16 ) , ...... skip printing (too many elements ) Activation gradient from back-propagation - vector size is 4 Activation Gradient ( 12 x 49 ) , ...... skip printing (too many elements ) Layer 0 : output D x H x W 12 7 7 input D x H x W 1 8 8 layer output size 4 Layer Output ( 12 x 49 ) , ...... skip printing (too many elements ) Evaluate the Derivatives with Finite difference and compare with BP for Layer 0 0 - 0 , 0 : -4.57912 from BP -4.57912 1.6113e-10 0 - 0 , 1 : 2.29655 from BP 2.29655 1.05075e-10 0 - 0 , 2 : 9.06929 from BP 9.06929 9.43657e-11 Testing weight gradients: layer: 1 / 6 Weight gradient from back-propagation - vector size is 1 BP Weight Gradient ( 6 x 108 ) , ...... skip printing (too many elements ) Activation gradient from back-propagation - vector size is 4 Activation Gradient ( 6 x 25 ) , ...... skip printing (too many elements ) Layer 1 : output D x H x W 6 5 5 input D x H x W 12 7 7 layer output size 4 Layer Output ( 6 x 25 ) , ...... skip printing (too many elements ) Evaluate the Derivatives with Finite difference and compare with BP for Layer 1 0 - 0 , 0 : 0.314135 from BP 0.314135 1.37732e-09 0 - 0 , 1 : -2.54656 from BP -2.54656 2.03892e-10 0 - 0 , 2 : -4.51668 from BP -4.51668 1.09001e-10 Testing weight gradients: layer: 2 / 6 Weight gradient from back-propagation - vector size is 1 BP Weight Gradient ( 6 x 54 ) , ...... skip printing (too many elements ) Activation gradient from back-propagation - vector size is 4 Layer 2 : output D x H x W 6 3 3 input D x H x W 6 5 5 layer output size 4 Evaluate the Derivatives with Finite difference and compare with BP for Layer 2 0 - 0 , 0 : 0 from BP 0 0 0 - 0 , 1 : 0 from BP 0 0 0 - 0 , 2 : 0 from BP 0 0 Testing weight gradients: layer: 3 / 6 Layer 3 has no weights Activation gradient from back-propagation - vector size is 1 Activation Gradient ( 4 x 54 ) , ...... skip printing (too many elements ) Layer 3 : output D x H x W 1 1 54 input D x H x W 6 3 3 layer output size 1 Layer Output ( 4 x 54 ) , ...... skip printing (too many elements ) Evaluate the Derivatives with Finite difference and compare with BP for Layer 3 Testing weight gradients: layer: 4 / 6 Weight gradient from back-propagation - vector size is 1 BP Weight Gradient ( 20 x 54 ) , ...... skip printing (too many elements ) Activation gradient from back-propagation - vector size is 1 Layer 4 : output D x H x W 1 1 20 input D x H x W 1 1 54 layer output size 1 Evaluate the Derivatives with Finite difference and compare with BP for Layer 4 0 - 0 , 0 : 5.07252 from BP 5.07252 5.19509e-11 0 - 0 , 1 : 4.70495 from BP 4.70495 7.70506e-12 0 - 0 , 2 : 5.58944 from BP 5.58944 4.78862e-11 Testing weight gradients: layer: 5 / 6 Weight gradient from back-propagation - vector size is 1 Activation gradient from back-propagation - vector size is 1 Layer 5 : output D x H x W 1 1 2 input D x H x W 1 1 20 layer output size 1 Evaluate the Derivatives with Finite difference and compare with BP for Layer 5 0 - 0 , 0 : 20.763 from BP 20.763 2.07679e-11 0 - 0 , 1 : 1.82507 from BP 1.82507 5.62671e-12 0 - 0 , 2 : 1.7897 from BP 1.7897 1.75692e-10 Testing weight gradients: maximum relative error: [NON-XML-CHAR-0x1B][33m1.74945e-07[NON-XML-CHAR-0x1B][39m