vanbasten23/maskrcnn

This repo is used for research into MaskRCNN for the purpose of implementing dynamic shapes.

It includes the following:

• maskrcnn.json -- a chrome trace of one full train iteration
• maskrcnn.py -- a script to run maskrcnn with real inputs from the COCO dataset.
• maskrcnn_inputs.pt -- the tensor inputs used by maskrcnn.py and maskrcnn_profiler.py
• maskrcnn_profiler.py -- an example on running MaskRCNN with PyTorch Profiler
• missing_jit_shape_funcs.py -- a script to generate missing JIT SSA shape functions
• missing_ub_funcs.py -- a script to generate missing ops in LTC for MaskRCNN
• torch_mlir_shape_funcs.py a script to generate the list of JIT shape functions we already implement

Installation

Run python install.py. The script will do the following:

• Recursively checks out torchvision if it hasn't been checked out already
• Builds torchvision