[Fix] DDRNet: fix tensor size mismatch on non-standard input resolutions

[Selig209]

When using DDRNet with input dimensions that don't divide evenly by 8 (e.g. after keep_ratio=True resizing), the forward pass crashes with:

RuntimeError: The size of tensor a (182) must match the size of tensor b (181) at non-singleton dimension 3

The issue is that out_size is precomputed as (H // 8, W // 8) from the raw input, but the stem's strided convolutions can produce spatial dims that are off by one due to integer division rounding. When comp_c gets resized to out_size and added to x_s, the shapes don't match.

Moved out_size to be computed from the actual tensor shape after the stem, so the resize target always matches x_s.

Fixes #3347

[CLAassistant]

All committers have signed the CLA.

[Copilot]

Pull request overview

Fixes a DDRNet forward-pass crash on inputs whose spatial dimensions don’t divide evenly by 8 by computing the resize target from the post-stem tensor shape (rather than from raw input integer division), preventing off-by-one mismatches during feature fusion.

Changes:

• Compute out_size from x.shape[-2:] after self.stem(x) to ensure all subsequent resize(..., size=out_size) calls match the spatial branch resolution.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This bug fix isn’t covered by a regression test. Please add a unit test for DDRNet that feeds an input whose H/W aren’t divisible by 8 (e.g. 1025x1023 or a keep_ratio-like shape) and asserts the forward pass runs without error and that the feature maps being added/resized have matching spatial shapes.