pnnx silently changes circular padding to reflect or constant padding #7061

error log | 日志或报错信息 | ログ

At bbaca40c35ce39f635c947e564710dae42fba142, pnnx converts torch.stft(..., center=True, pad_mode="circular") to a loadable Spectrogram layer with 6=2 (reflect). torchaudio.functional.spectrogram has the same behavior. F.pad(..., mode="circular") becomes Padding without a mode parameter, so it uses constant padding.

The generated ncnn models return numerically different results without an unsupported-padding diagnostic. For the STFT reproduction below, the maximum absolute error is 1.2021422; the result instead matches the reflect-padding reference within 3.95e-6. A (2, 3, 128) torchaudio power spectrogram gives a maximum absolute error of 30.0352. A separate F.pad(x, (3, 4), mode="circular") model with input shape (2, 128) gives 0.76822.

model | 模型 | モデル

This is a TorchScript conversion issue. The following script generates the original model and checks the converted result:

import subprocess
import torch

class Model(torch.nn.Module):
    def __init__(self, mode):
        super().__init__()
        self.mode = mode

    def forward(self, x):
        return torch.view_as_real(torch.stft(
            x, n_fft=32, hop_length=8, win_length=32,
            window=torch.hann_window(32), center=True,
            pad_mode=self.mode, return_complex=True))

torch.manual_seed(0)
x = torch.rand(128)
model = Model("circular").eval()
torch.jit.trace(model, x).save("circular_stft.pt")
subprocess.run(["pnnx", "circular_stft.pt", "inputshape=[128]", "fp16=0"], check=True)

import circular_stft_ncnn
actual = circular_stft_ncnn.test_inference()
print("circular error:", (model(x) - actual).abs().max().item())
print("reflect error:", (Model("reflect")(x) - actual).abs().max().item())

how to reproduce | 复现步骤 | 再現方法

1. Build pnnx and the ncnn Python bindings from the commit above.
2. Run the script with pnnx on PATH. I reproduced this on Windows x64, MSVC 19.44, Python 3.14.3, PyTorch 2.11.0+cpu and torchaudio 2.11.0, using CPU inference and fp16=0.
3. Inspect the generated circular_stft.ncnn.param: it contains Spectrogram ... 5=1 6=2 even though the .pnnx.param retains pad_mode=circular.

The runtime currently supports constant, replicate and reflect padding. Unsupported circular padding should stay unconverted with a diagnostic, rather than become a different supported mode. The normal constant/replicate/reflect cases pass numerical controls. center=False also passes because STFT does not use the padding mode in that case.