convert_to_analog drops a repeated layer from nn.Sequential #825

Description

Converting nn.Sequential(layer, layer) removes the second application of the shared layer. With FloatingPointRPUConfig, the converted model returns 0.25 instead of 0.0625.

aihwkit 1.2.0 at b5aeea78859a4a028975519e154088d3ae0efeaf, PyTorch 2.5.1+cpu, Python 3.12.3, NumPy 2.5.4, Linux x86_64, GCC 13.3.0, OpenBLAS 0.3.28, native CPU Release build, CUDA off, Conda N/A.

How to reproduce

Using a CPU source build of that commit, save this as sequential_repro.py.

"""Default conversion drops a repeated layer from nn.Sequential."""
import numpy as np
import torch
from torch import nn
from aihwkit.nn.conversion import convert_to_analog
from aihwkit.simulator.configs import FloatingPointRPUConfig

torch.set_num_threads(1)
weight = np.float64(1 / 4)
x = np.float64(1)
expected = np.array([[weight * weight * x]], dtype=np.float64)
layer = nn.Linear(1, 1, bias=False)
with torch.no_grad():
    layer.weight.fill_(float(weight))
digital = nn.Sequential(layer, layer)
input_tensor = torch.tensor([[x]], dtype=torch.float32)
np.testing.assert_array_equal(digital(input_tensor).detach().numpy(), expected)
analog = convert_to_analog(digital, FloatingPointRPUConfig())
actual = analog(input_tensor).detach().numpy()
print(f"Layer count: original={len(digital)}, converted={len(analog)}")
print(f"Output: expected={expected.tolist()}, actual={actual.tolist()}")
np.testing.assert_array_equal(actual, expected)

OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 python sequential_repro.py

Observed output before the final assertion fails.

Layer count: original=2, converted=1
Output: expected=[[0.0625]], actual=[[0.25]]

Expected behavior

The conversion contract promises replacement of all Conv/Linear layers. Both applications should remain and yield (1/4)*(1/4)*1 = 1/16, following the ideal tile equation. The conversion docs are silent about repeated layer registrations.

Other information

The container conversion rebuilds the sequence from named_children(), which omits duplicate module objects. We tested CPU float32 with both inplace settings and passing distinct-layer controls, but did not test CUDA or convolution reuse.

The reproducer is CC0-1.0.

— ChipLab