data2/text/DLP/1032151.txt

Ticket Name: TDA2PXEVM: How can I extract two CNN output tensors on TIDL (TDA2)

Query Text:
Part Number: TDA2PXEVM Other Parts Discussed in Thread: TDA2 Hi sir! My network have two (maybe more) output layers (CxHxW) : Cx56x56 & Cx28x28. How can I get two outputs from TDIL?

Responses:
Hi IIuo, Sorry for the delay. This is not supported in TIDL on TDA2. Thanks, Praveen

Hi Praveen I proposed a method to solve this problem ： using Transpose Conv and Concat. Step 1. I add a Transposed Conv layer with 4x4 kernel, stride 2 to the small size layer (for upsampling : Cx28x28 -> Cx56x56) Step 2. I fill filter manually to make Transposed Conv layer working as insert 0 between original output elements [e11 e12 -> [ e11 0 e12 0 e21 e22] 0 0 0 0 e21 0 e22 0 0 0 0 0 ] Step 3. I use Concat layer to get my output tensor : 2Cx56x56 So I can simply add stride =1 / 2 to decode different output tensor. Those works have been tested on Caffe. Hope that can help others.

Thanks for sharing.