nearest exact interpolation backward is not supported by PyT

Error message

nearest exact interpolation backward is not supported by PyTorch/tch backend

What it means

The burn-tch (LibTorch) backend does not implement the backward pass for NearestExact interpolation mode; calling it panics. PyTorch has no direct upsample_nearest_exact2d_backward binding exposed via tch, so the backend intentionally aborts instead of silently producing wrong gradients. This is an unimplemented-feature panic, not a data or environment problem.

Source

Thrown at crates/burn-tch/src/ops/module.rs:460

        grad: TchTensor,
        output_size: [usize; 2],
        options: InterpolateOptions,
    ) -> TchTensor {
        let output_size = output_size.map(|e| e as i64);
        let [n, c, h_in, w_in] = x.shape().dims();
        let input_size = [n as i64, c as i64, h_in as i64, w_in as i64];
        let align_corners = options.align_corners;

        let tensor = match options.mode {
            InterpolateMode::Nearest => tch::Tensor::upsample_nearest2d_backward(
                &grad.tensor,
                output_size,
                input_size,
                None,
                None,
            ),
            InterpolateMode::NearestExact => {
                panic!(
                    "nearest exact interpolation backward is not supported by PyTorch/tch backend"
                )
            }
            InterpolateMode::Bilinear => tch::Tensor::upsample_bilinear2d_backward(
                &grad.tensor,
                output_size,
                input_size,
                align_corners,
                None,
                None,
            ),
            InterpolateMode::Bicubic => tch::Tensor::upsample_bicubic2d_backward(
                &grad.tensor,
                output_size,
                input_size,
                align_corners,
                None,
                None,

View on GitHub (pinned to d16f7ba2ed)

Solutions

Switch the interpolate layer to InterpolateMode::Nearest for tch (backward is supported and usually numerically close)
Use InterpolateMode::Bilinear or Bicubic, both of which have supported backward passes in tch
Run this model/graph on a backend that implements NearestExact backward (e.g. burn-cube/wgpu)
Guard the mode at startup: validate InterpolateOptions before building the training loop and reject NearestExact when the tch backend is selected

Example fix

// before
let options = InterpolateOptions::new(InterpolateMode::NearestExact);
// after
let options = InterpolateOptions::new(InterpolateMode::Nearest);

Defensive patterns

Strategy: validation

Validate before calling

fn ensure_interpolate_backward_supported(options: &InterpolateOptions) {
    match options.mode {
        InterpolateMode::NearestExact | InterpolateMode::Lanczos3 => {
            panic!("mode {:?} has no backward in burn-tch; use Nearest/Bilinear/Bicubic", options.mode)
        }
        _ => {}
    }
}

Try / catch

let result = std::panic::catch_unwind(|| model.forward(x).backward());
match result {
    Ok(grads) => grads,
    Err(_) => fallback_to_supported_backend_or_mode(),
}

Prevention

Never use InterpolateMode::NearestExact or Lanczos3 in models that will train on burn-tch
Centralize interpolate options in one config module and assert backend compatibility at startup
Add a smoke test that backprops through every layer/mode combination you ship
Prefer Bilinear for resize layers on tch

When it happens

Trigger: Backward pass of a model whose interpolate (upsample) layer uses InterpolateOptions { mode: InterpolateMode::NearestExact } on the tch backend — i.e. during autodiff training or gradient computation of a resize/upsample op configured with mode=nearest-exact.

Common situations: Porting a model from the burn-cube/wgpu backend (where NearestExact works) to burn-tch for training; configuring interpolation mode to match an ONNX/PyTorch 'nearest-exact' resize; shared training code that lets a config file pick the mode.

Related errors

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/7cc5e1acb714e794. Report an issue: GitHub.