[Bug] `JukeboxLoss.forward`: `input_amplitude` and `target_amplitude` variable names are swapped

[Zeesejo]

Describe the bug

In monai/losses/spectral_loss.py, JukeboxLoss.forward() has the variable names input_amplitude and target_amplitude swapped:

def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
    input_amplitude = self._get_fft_amplitude(target)   # ← BUG: should be _get_fft_amplitude(input)
    target_amplitude = self._get_fft_amplitude(input)   # ← BUG: should be _get_fft_amplitude(target)

    loss = F.mse_loss(target_amplitude, input_amplitude, reduction="none")

Impact

While this does not affect the numerical output of the loss (since MSE is symmetric: mse(a, b) == mse(b, a)), the naming is semantically inverted which:

1. Makes the code misleading and confusing to anyone reading, extending, or debugging it
2. Creates a silent maintenance hazard — if anyone adds asymmetric processing (e.g. normalization, weighting, gradient masking) based on input vs. target, the swap will cause subtle, hard-to-debug training errors
3. Violates the principle of least surprise for the standard forward(input, target) contract that all PyTorch loss functions follow

Root Cause

In forward(self, input, target):

• input_amplitude is computed from target (wrong)
• target_amplitude is computed from input (wrong)

Expected behavior

def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
    input_amplitude = self._get_fft_amplitude(input)    # correct
    target_amplitude = self._get_fft_amplitude(target)  # correct
    loss = F.mse_loss(target_amplitude, input_amplitude, reduction="none")

Location

monai/losses/spectral_loss.py lines 55-56

Affects all MONAI versions containing JukeboxLoss. Reproducible on dev branch as of 2026-04-11.