Failure to export torch.nn.functional.pad with int32 tensor

[gsigms]

Convert does not seem to support torch pad with int32 input type

ValueError: In op, of type pad, named val, the named input constant_valmust have the same data type as the named inputx`. However, constant_val has dtype fp32 whereas x has dtype int32.

venv/lib/python3.9/site-packages/coremltools/converters/mil/mil/input_type.py:137: ValueError`

import coremltools as ct
import torch

class Example(torch.nn.Module):

    def forward(self, x):
        x = torch.nn.functional.pad(x, (0, 2), value=0)
        return x

module = Example().eval()
traced = torch.jit.trace(module, (torch.IntTensor([[1, 2], [3, 4]]),))

coremlmodel = ct.convert(traced,
    source='pytorch',
    convert_to='mlprogram',
    inputs=[ct.TensorType(name="x", dtype=ct.converters.mil.mil.types.int32, shape=(2, 2)),],
    outputs=[ct.TensorType(name="y", dtype=ct.converters.mil.mil.types.int32)],
    compute_units=ct.ComputeUnit.ALL,
    compute_precision=ct.precision.FLOAT16,
)

System environment (please complete the following information):

coremltools version: 6.2 mac os x Ventura torch: 1.13.1 python 3.9

[jakesabathia2]

@gsigms Thanks for filing this issue. However, I strongly won't recommend using the int32 input data type, since the most of the coreml op only accept float inputs. Please change the input type to float32 and I think that could solve the issue. Thanks

[bunnyhero]

My desired use case is using pad to populate a tensor that will be used as the reps argument for tile, later. The reps parameter requires int32. I guess I'll cast back and forth instead.