Lower adaptive pooling to output size 1 as a global reduce in 1D

[LeSingh1]

Problem

nn.AdaptiveAvgPool1d(1) over a flexible length — global pooling across a variable-length sequence, the standard head of a 1-D convnet — aborts conversion:

import torch, torch.nn as nn, coremltools as ct

model = nn.AdaptiveAvgPool1d(1).eval()
x = torch.rand(1, 64, 8)
ct.convert(
    torch.jit.trace(model, x),
    inputs=[ct.TensorType(shape=ct.Shape([1, 64, ct.RangeDim(upper_bound=20)]))],
    convert_to="mlprogram",
)

File ".../sympy/core/expr.py", line 375, in __float__
    raise TypeError("Cannot convert expression to float")
TypeError: Cannot convert expression to float

_adaptive_pool1d always slices the input into one window per output element, computing the window boundaries in python from the pooled dimension:

start = _math.floor(i * in_dimension / out_dimension)
end = _math.ceil((i + 1) * in_dimension / out_dimension)

For a flexible input in_dimension is a sympy symbol, so floor / ceil raise out of sympy, from the middle of the converter with no indication of what the user did wrong.

The 2-D equivalent has no such problem: _adaptive_pool2d special-cases output size (1, 1) into a global reduce_mean / reduce_max, which needs no window boundaries, so nn.AdaptiveAvgPool2d((1, 1)) over a flexible input converts today. The 1-D path just never got the same treatment.

Fix

Give _adaptive_pool1d the two cases _adaptive_pool2d already has:

• output size == 1 → a single global reduce over the last axis. This is exactly what the window loop already computes for that case (its one window spans the whole dimension), so the result is unchanged for static shapes — it just no longer needs to know the size.
• any other output size with a symbolic input shape → the same clear ValueError _adaptive_pool2d raises, instead of a sympy TypeError.

Tests

• TestAdaptiveAvgPool::test_adaptive_avg_pool1d_symbolic_input — nn.AdaptiveAvgPool1d(1) with a RangeDim length. Fails with the sympy TypeError without the fix.
• TestAdaptiveAvgPool::test_adaptive_avg_pool1d_output_size_1 — output size 1 with static shapes over rank 3 and rank 2 input, pinning the new global-reduce path against torch. The existing test_adaptive_max_pool1d parametrizations only use output sizes 3 and 5, so output size 1 was uncovered.

The symbolic test skips the torch.export frontends, which decompose AdaptiveAvgPool1d(1) into a mean op before the converter sees it.

Verification

pytest coremltools/converters/mil/frontend/torch/test/test_torch_ops.py -k "TestAdaptiveAvgPool or TestAdaptiveMaxPool"

618 passed, 2 skipped on macOS / Apple silicon, torch 2.12, across mlprogram + neuralnetwork and the TorchScript / TorchExport / ExecuTorch frontends.