Timeline for CNN - upsampling backprop gradients across average-pooling layer
Current License: CC BY-SA 4.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Feb 20, 2022 at 9:38 | vote | accept | x.projekt | ||
| Feb 20, 2022 at 9:33 | comment | added | x.projekt | @gunes- I thought of the exact same method, and posted it as an answer. | |
| Feb 20, 2022 at 9:01 | comment | added | gunes | I don't know which way would be best, but it needs reshaping since the second multiplicand is a 4d tensor with indices m,n,i,j. If you flatten dJ/dp (row-major) and construct another matrix with shape 4x9 where each entry holds the derivative of dP/dA (both flattened), then you can form a matrix mult, (dJ/dP) x (dP/dA). | |
| Feb 20, 2022 at 7:08 | comment | added | x.projekt | how to obtain this summation (i.e., $\frac{\partial J}{\partial a_{ij}}=\sum_{m,n}\frac{\partial J}{\partial p_{mn}}\frac{\partial p_{mn}}{\partial a_{ij}}$) using matrix-calculus? | |
| Feb 19, 2022 at 22:08 | history | answered | gunes | CC BY-SA 4.0 |