I am using an older version of TVM@a64feed.
FasterRCNN was downloaded from here: http://download.tensorflow.org/models/object_detection/faster_rcnn_inception_v2_coco_2018_01_28.tar.gz
VM Compilation reaches ManifestAlloc stage within 5 minute. Since the graph becomes huge after this point, FoldConstants takes about 1 hour. Since the FoldConstants is invoked 3 times after this point, it takes about 3 hour to finish the OptimizeModule().
Lowering takes about 5 min.
Codegen then takes over 2 hour.
First I would like to focus on the FoldConstant issue. Is there way to accelerate the interpreter or have an alternative way to determine the values for constants?
Any information related to slowness that you guys can share will be helpful for me.
Also, if there is an already working variant of Faster RCNN available, could you please share the link?
Thank you for the support. Regards, Ashutosh