I have a model that have Squeeze & Excitation blocks in it. When I try to run optimization on it, I get errors on the avgpool Shift delta. Are models with S&E blocks not supported?
Since SE blocks blocks the inference pipeline, we have the option to apply Tiled SE and excite - this is a method to transform the SE blocks into an equivalent structure that serves the same purpose, but doesn’t block the inference pipeline.
It can be don be using the following alls commands when running optimization:
pre_quantization_optimization(se_optimization, method=tse, mode=sequential, count=X, tile_height=Y)
where X is the number of SE blocks in the model and Y is the number in which the avgpool kernel would be divided by (if the kernel is 40 and Y=5, then the kernel applied would be 8).
You can read about the mechanism in the article: