[NHWC] InferShape Layout conversion fix. #372
Conversation
|
just want to confirm the NHWC weight convention and what is the intended behavior. @yuruofeifei @yidawang @yzhliu can you confirm? |
|
cc @piiswrong This seems to be a common weight convention for NHWC used in packages. |
|
@tqchen I don't know if it is common in other frameworks, but this was the format I figured out while I was trying to vectorize NHWC convolution. Suppose the filter layout is [num_filter, height, width, in_channel]. Then I would vector-load from input channel dimension. But since input channel axis is one of the reduce axes, I cannot do data parallel multiply-add-accumulate. Instead, I need to reduce over vector elements. |
|
@tqchen I'm implementing a more general layout system, in new design, weight has its own layout, e.g., OIHW (oc, ic, height, width). We can talk about it later. |
* [DOC] Add Build TVM Runtime on Device section Add Build TVM Runtime on Device section * Add echo USE_RPC=1>> config.mk in code-block Add echo USE_RPC=1>> config.mk in Build TVM Runtime on Device section * [DOC] Fix small problem
No description provided.