# Reduce rank and set values according to 3rd dimension **URL:** \ **Category:** SameDiff\ **Created:** [September 27, 2022, 12:32pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029 "2022-09-27T12:32:14Z")\ **Posts on this page:** 13\ **Page:** 1
**Author:** ![SidneyLann](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/sidneylann/32/367_2.png) [@SidneyLann](https://community.konduit.ai/u/SidneyLann)\ **Post date:** [September 27, 2022, 12:32pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/1 "2022-09-27T12:32:14Z")
[ [[1,0,0,0],[0,1,0,0],[0,0,1,0]], [[1,0,0,0],[0,1,0,0],[0,0,1,0]] ] How to get below result from above SDVariable? Reduce 3rd dimension, and the value = 0.1 when 1 exists in the first or 2rd position, otherwise value = 1. [ [0.1,0.1,1], [0.1,0.1,1] ] ---
**Author:** ![agibsonccc](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/agibsonccc/32/697_2.png) [@agibsonccc](https://community.konduit.ai/u/agibsonccc)\ **Post date:** [September 27, 2022, 12:34pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/2 "2022-09-27T12:34:23Z")
@SidneyLann mind posting some code to start from? I’ll take that and modify it. I can say for sure that SDVariable does have get(…) just like on the ndarrays. You might be able to use that as is. ---
**Author:** ![SidneyLann](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/sidneylann/32/367_2.png) [@SidneyLann](https://community.konduit.ai/u/SidneyLann)\ **Post date:** [September 27, 2022, 12:45pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/3 "2022-09-27T12:45:11Z")
SDVariable input = sameDiff.placeHolder(“input”, DataType.FLOAT, -1, timeSteps, numFeatrues); SDVariable label = sameDiff.placeHolder(“label”, DataType.FLOAT, -1, timeSteps, numLabelClasses); SDVariable weight = label.argmax(“weight”, 2); … sameDiff.loss.softmaxCrossEntropy(“loss”, label, out, weight); I want to generate weights dynamic according to label SDVariable . ---
**Author:** ![agibsonccc](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/agibsonccc/32/697_2.png) [@agibsonccc](https://community.konduit.ai/u/agibsonccc)\ **Post date:** [September 27, 2022, 12:49pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/4 "2022-09-27T12:49:03Z")
Do you need just a subset of the dynamic weights then? If so, I think var.get(SDIndex.(…)) methods will do pretty well for you. You can even do it with variables. ---
**Author:** ![SidneyLann](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/sidneylann/32/367_2.png) [@SidneyLann](https://community.konduit.ai/u/SidneyLann)\ **Post date:** [September 27, 2022, 12:52pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/5 "2022-09-27T12:52:15Z")
I am not just get a subset, also to assign new values to it. and the new values according to the 3rd dimension. It may need to use nested for loops. ---
**Author:** ![agibsonccc](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/agibsonccc/32/697_2.png) [@agibsonccc](https://community.konduit.ai/u/agibsonccc)\ **Post date:** [September 27, 2022, 12:53pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/6 "2022-09-27T12:53:13Z")
@SidneyLann then you can use put(…) - thttps://github.com/deeplearning4j/deeplearning4j/blob/2668876811fbebf1fafc09f954a34dfcd8c95780/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java#L1807 This actually does dynamic looping for you internally. ---
**Author:** ![SidneyLann](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/sidneylann/32/367_2.png) [@SidneyLann](https://community.konduit.ai/u/SidneyLann)\ **Post date:** [September 27, 2022, 1:06pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/7 "2022-09-27T13:06:24Z")
It seems I can’t use it, because I actually need to generate a toPut, but it need pass a toPut. and my case need to assign by condition. ---
**Author:** ![agibsonccc](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/agibsonccc/32/697_2.png) [@agibsonccc](https://community.konduit.ai/u/agibsonccc)\ **Post date:** [September 27, 2022, 1:08pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/8 "2022-09-27T13:08:53Z")
@SidneyLann you can probably reuse the control flow. We have the boolean eval conditions etc all there. We also have where all in there. Am I correct in understanding you need to update your weights where they meet a certain condition? Is that the idea? Current samediff should be able to do all that as is. ---
**Author:** ![SidneyLann](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/sidneylann/32/367_2.png) [@SidneyLann](https://community.konduit.ai/u/SidneyLann)\ **Post date:** [September 27, 2022, 1:14pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/9 "2022-09-27T13:14:26Z")
SDVariable label = sameDiff.placeHolder(“label”, DataType.FLOAT, -1, timeSteps, numLabelClasses); SDVariable weight =…( label ) I need generate weight dynamic from label, the label has rank 3, and the weight has rank 2. This is label [ [[1,0,0,0],[0,1,0,0],[0,0,1,0]], [[1,0,0,0],[0,1,0,0],[0,0,1,0]] ] And this is weight [ [0.1,0.1,1], [0.1,0.1,1] ] ---
**Author:** ![agibsonccc](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/agibsonccc/32/697_2.png) [@agibsonccc](https://community.konduit.ai/u/agibsonccc)\ **Post date:** [September 27, 2022, 1:33pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/10 "2022-09-27T13:33:34Z")
@SidneyLann is this for weighted loss? If not why not make that a placeholder? ---
**Author:** ![SidneyLann](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/sidneylann/32/367_2.png) [@SidneyLann](https://community.konduit.ai/u/SidneyLann)\ **Post date:** [September 27, 2022, 1:45pm UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/11 "2022-09-27T13:45:55Z")
sameDiff.loss.softmaxCrossEntropy(“loss”, label, lstm1, weight); This should be weighted loss? Yes, can use placeholder, but it need more maintenance work. ---
**Author:** ![agibsonccc](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/agibsonccc/32/697_2.png) [@agibsonccc](https://community.konduit.ai/u/agibsonccc)\ **Post date:** [September 28, 2022, 5:44am UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/12 "2022-09-28T05:44:25Z")
@SidneyLann could you clarify what you mean by maintenance work? Is there something we could do on our side to make that easier for you? ---
**Author:** ![SidneyLann](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/sidneylann/32/367_2.png) [@SidneyLann](https://community.konduit.ai/u/SidneyLann)\ **Post date:** [September 28, 2022, 8:49am UTC](https://community.konduit.ai/t/reduce-rank-and-set-values-according-to-3rd-dimension/2029/13 "2022-09-28T08:49:40Z")
The logic to generate the weights INDArray is almost the same as label array, so if the label logic is amended, the same thing need to do for weights. Ok. I generate the weights and use placeholder and it work now.