Problems to use a converted Py Torch Vanilla NN in a Pipeline when it has hidden layers with different shapes from input layer

[sjordine]

🐞Describing the bug

• I have create a PyTorch very simple vanilla neural network:

class SimpleModel(nn.Module):
  def __init__(self):
    super().__init__()
    self.linear1 = nn.Linear(3, 4)
    #self.activation1 = nn.ReLU()
    #self.linear2 = nn.Linear(4, 3)
    self.linear3 = nn.Linear(4, 1)

  def forward(self, x):
     x = self.linear1(x)
     #x = self.activation1(x)
     #x = self.linear2(x)
     x = self.linear3(x)
     return x

And converted it to a Core ML package using Core ML Tools

input = torch.rand(1,3)
model.eval()
traced_model = torch.jit.trace(model, input)
mlmodel = ct.convert(
    traced_model,
    inputs=[ct.TensorType(name="input", shape=input.shape)],
)
mlmodel.save("Test.mlpackage")

This model can be open in Xcode with no problem

But, if I try to use it in a Pipeline:

pipeline_network = pipeline.Pipeline (
   input_features = [("input",datatypes.Array(1,3))],
   output_features=[("linear_1",datatypes.Array(1,1))]
)
pipeline_network.add_model(mlmodel)
pipeline_spec = pipeline_network.spec
ct.utils.convert_double_to_float_multiarray_type(pipeline_spec)
ct.utils.save_spec(pipeline_spec, "Test-Pipeline.mlpackage")

It does not open on XCode and give me the following error:

This problem does not occur if self.linear1 = nn.Linear(3, 3) and self.linear3 = nn.Linear(3, 1).

To Reproduce

• Here is my full example:

!pip install torch==2.1.2
!pip install --upgrade coremltools

!rm -Rf /content/Test.mlpackage
!rm -Rf /content/Test-Pipeline.mlpackage

!rm -Rf /content/simple-export.zip
!rm -Rf /content/pipeline-export.zip

import torch
import torch.nn as nn

class SimpleModel(nn.Module):
  def __init__(self):
    super().__init__()
    self.linear1 = nn.Linear(3, 4) #changing 4 to 3 here and in linear3 make it work
    #self.activation1 = nn.ReLU()
    #self.linear2 = nn.Linear(4, 3)
    self.linear3 = nn.Linear(4, 1) #changing 4 to 3 here and in linear1 make it work

  def forward(self, x):
     x = self.linear1(x)
     #x = self.activation1(x)
     #x = self.linear2(x)
     x = self.linear3(x)
     return x

input = torch.rand(1,3)

model = SimpleModel()

model.eval()

traced_model = torch.jit.trace(model, input)

import coremltools as ct
import numpy as np

mlmodel = ct.convert(
    traced_model,
    inputs=[ct.TensorType(name="input", shape=input.shape)],
    outputs=[ct.TensorType(name="linear_1")],
    compute_precision = ct.precision.FLOAT32,
    compute_units=ct.ComputeUnit.CPU_AND_GPU
    #minimum_deployment_target = ct.target.iOS17,
    #convert_to="mlprogram"
)

mlmodel.save("Test.mlpackage")

!zip -r /content/simple-export.zip /content/Test.mlpackage

from coremltools.models import pipeline
from coremltools.models import datatypes

import coremltools.models as models

pipeline_network = pipeline.Pipeline (
   input_features = [("input",datatypes.Array(1,3))],
   output_features=[("linear_1",datatypes.Array(1,1))]
)

pipeline_network.add_model(mlmodel)

pipeline_spec = pipeline_network.spec

ct.utils.convert_double_to_float_multiarray_type(pipeline_spec)

ct.utils.save_spec(pipeline_spec, "Test-Pipeline.mlpackage")

!zip -r /content/pipeline-export.zip /content/Test-Pipeline.mlpackage

Test.mlpackage will open without problem in Xcode, but Test-Pipeline.mlpackage will present the outputSchema error...

System environment

• coremltools version: 8.1
• Colab environment
• Py torch version: 2.1.2

[JT11-11]

I'm interested to take this issue .

[sjordine]

I'm interested to take this issue .

If you take this and need any further info, I will be glad to help! Thanks so much!