masahi edited a comment on pull request #6449: URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692441507
@kevinthesun Thanks I'm trying running e2e on my end. I have following questions: 1. How long does it take to compile faster or mask rcnn from torchvision? I remember hearing TF faster rcnn taking 20 min to compile. If it is too slow, it might not be a good idea to run them on CI... 2. Interestingly, mask rcnn seems to have `prim::Loop`, even though it is traced. In general, does tracing make sense for object detection models (that might have data dependent code path)? The loop is coming from https://github.com/pytorch/vision/blob/1a04d3c265679e1a508e7cd627006aaa9ef1ccfb/torchvision/models/detection/roi_heads.py#L457, so this is a partly scripted model. 3. I'm having this error below, is it expected? ``` TVMError: Check failed: allow_alloc: Cannot find the Realization point of tensor Tensor(shape=[1], op.name=box_data_length) During handling of the above exception, another exception occurred: TVMError: Check failed: allow_alloc: Cannot find the Realization point of tensor Tensor(shape=[1], op.name=box_data_length) Error during compile function ----------------------------- #[version = "0.0.5"] fn (%p0: Tensor[(1, ?, ?), float32], Primitive=1) -> (Tensor[(1), int32], Tensor[(1, ?, ?), float32], Tensor[(1, ?), int32]) { vision.get_valid_counts(%p0, meta[relay.attrs.GetValidCountsAttrs][0]) /* ty=(Tensor[(1), int32], Tensor[(1, ?, ?), float32], Tensor[(1, ?), int32]) */ } ``` ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org