Shapes of intermediate tensors

Ask a Question

Question

In a onnx graph, I can see the tensor shapes for the inputs and outputs. Is there a way to know what shapes the intermediate tensors are? I consulted the shape inferencing document but it looks like there is only a c++ api.

Further information

• Relevant Area: shape_inference

Notes

Model exported from pytorch

import torch
from torch import nn

class Upsampling(nn.Module):
    def __init__(self):
        super().__init__()
        self.upsample = nn.Upsample(size=[2,2], mode='nearest')

    def forward(self, input, weight):
        weight =  self.upsample(weight)
        return torch.nn.functional.conv2d(input, weight)

torch.onnx.export(Upsampling(), (torch.tensor([[[[1,2],[1,2]]]], dtype=torch.float32), torch.tensor([[[[1]]]], dtype=torch.float32)), "test.onnx")

[jcwchen]

After offline discussion, we found that actually Resize is missing data propagation support so it cannot get shape information from previous Shape op. Then, I made a quick experiment and adding ctx.getSymbolicInput() in Resize op can indeed make it infer the right info. I will submit a PR for it when I have time. Thank you @justinchuby for catching this!

[jcwchen]

~It needs the exact “tensor value” of input “sizes” for Resize, which comes from model.graph.input. The shape for the input “sizes” is not sufficient.~