TensorFlow frontend Softplus converter still emits native softplus op (fp16 overflow risk on ANE)

[Ashutosh0x]

Problem

The TensorFlow/Keras frontend's Softplus converter at coremltools/converters/mil/frontend/tensorflow/ops.py:2149 still emits the native mb.softplus MIL op:

python
def Softplus(context, node):
...
x = mb.softplus(x=x, name=node.name) # native op → fp16/ANE overflow

This is the same native op that overflows in fp16 on Apple Neural Engine at x ≈ 10.4, causing output collapse to 0.

Context

PR #2725 fixes this for the PyTorch frontend by replacing mb.softplus() with the numerically stable decomposition max(x,0) + log(1+exp(-|x|)). The TF frontend was intentionally left out of scope since both original issues (#2687, #2359) originated from PyTorch models.

Proposed Fix

Apply the same _stable_softplus_mil(x) helper to the TF Softplus converter. Alternatively, implement this at the MIL graph-rewrite pass level so all frontends are covered (the DRY fix, as suggested by @ChinChangYang in the PR #2725 review).

Related

• PR Fix softplus and mish fp16 overflow on ANE via stable decomposition #2725 (PyTorch frontend fix)
• Issue Softplus on Apple Neural Engine has a hard fp16 discontinuity at x ≈ 10.4 (output drops to 0) #2687 (original overflow report)
• Issue High Numerical Errors in Mish Activation with FLOAT16 Precision on Neural Engine #2359 (Mish fp16 errors)