The version of TVM used in fast-depth and model transformation pipline

[MRinel]

The version of TVM used in fast-depth and model transformation pipline

[dwofk]

Hi @MRinel,

We used a version of TVM that was between v0.4 and v0.5. To get the specific version of TVM used in our pipeline, please do:

git clone --recursive https://github.com/dmlc/tvm
cd tvm
git reset --hard ab4946c8b80da510a5a518dca066d8159473345f
git submodule update --init

You could then proceed with building TVM as specified here.

Operators are defined in tvm/nnvm/python/nnvm/frontend/onnx.py. You can search that file for class Upsample(OnnxOpConverter). In older versions of TVM, the Upsample operator had not yet been implemented, which may explain why you are seeing a NotImplementedError. In the version specified above, this operator had been added.

For our work, we modified this operator slightly to:

class Upsample(OnnxOpConverter):
    """ Operator converter for Upsample (nearest mode).
    """

    @classmethod
    def _impl_v7(cls, inputs, attr, params):
        scales = attr.get('scales')
        #assert len(scales) == 4 and scales[0] == 1.0 and scales[1] == 1.0 and scales[2] == scales[3]
        mode = attr.get('mode')
        if mode == b'nearest':
            method = "NEAREST_NEIGHBOR"
        elif mode == b'bilinear':
            method = "BILINEAR"
        else:
            raise ValueError("Invalid ONNX upsample mode: {}".format(mode))
        return _sym.upsampling(inputs[0], scale=2, method=method, layout='NCHW')

After modifying this file, you would need to rerun the Python package installation.

Yes, the pipeline we used was mobilenet-nnconv5dw-skipadd-pruned.pth.tar --> onnx model --> compile/deploy with TVM. We performed auto-tuning for CPU and GPU backends using AutoTVM.