I believe that we could avoid compute_shape in ExtractImagePatches' create_kernel_tensor function by using a Range op that takes a TensorDim parameter along with Scatter. However, such a Range op has not yet been implemented. Moreover, I'm not quite sure if switching to TensorDim there would get in the way of optimizations noticing that everything involved was a constant.
Investigate whether the above approach would work as desired, and if so implement it.
I believe that we could avoid
compute_shapein ExtractImagePatches'create_kernel_tensorfunction by using a Range op that takes aTensorDimparameter along with Scatter. However, such a Range op has not yet been implemented. Moreover, I'm not quite sure if switching to TensorDim there would get in the way of optimizations noticing that everything involved was a constant.Investigate whether the above approach would work as desired, and if so implement it.
See also https://github.com/plaidml/openvino/pull/96