SymbolicValueError: STFT does not currently support complex types

[kabyanil]

Describe the bug

I am trying to export the preprocessor class of Quartznet model to onnx. My requirement is to run preprocessing in a non-python environment. However, the export is failing.

Steps/Code to reproduce bug

The bug can be reproduced using the following code -

quartznet = nemo_asr.models.EncDecCTCModel.from_pretrained(model_name="QuartzNet15x5Base-En")

preprocessor = quartznet.preprocessor.to(device)

preprocessor.eval()

preprocessor.export("prep.onnx", onnx_opset_version=17)

The following error is produced:

{
	"name": "SymbolicValueError",
	"message": "STFT does not currently support complex types  [Caused by the value '97 defined in (%97 : Float(*, *, strides=[2946, 1], requires_grad=0, device=cpu) = onnx::Reshape[allowzero=0](%86, %96), scope: nemo.collections.asr.modules.audio_preprocessing.AudioToMelSpectrogramPreprocessor::/nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures::featurizer # /home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/functional.py:703:0
)' (type 'Tensor') in the TorchScript graph. The containing node has kind 'onnx::Reshape'.] 
    (node defined in /home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/functional.py(703): stft
/home/ik/Documents/aakhor/asr/NeMo/nemo/collections/asr/parts/preprocessing/features.py(308): <lambda>
/home/ik/Documents/aakhor/asr/NeMo/nemo/collections/asr/parts/preprocessing/features.py(423): forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/utils/_contextlib.py(116): decorate_context
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1726): _slow_forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1747): _call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1736): _wrapped_call_impl
/home/ik/Documents/aakhor/asr/NeMo/nemo/collections/asr/modules/audio_preprocessing.py(292): get_features
/home/ik/Documents/aakhor/asr/NeMo/nemo/collections/asr/modules/audio_preprocessing.py(91): forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/utils/_contextlib.py(116): decorate_context
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1726): _slow_forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1747): _call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1736): _wrapped_call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/jit/_trace.py(130): wrapper
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/jit/_trace.py(139): forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1747): _call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1736): _wrapped_call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/jit/_trace.py(1500): _get_trace_graph
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(904): _trace_and_get_graph_from_model
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(997): _create_jit_graph
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(1113): _model_to_graph
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(1564): _export
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(502): export
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/__init__.py(375): export
/home/ik/Documents/aakhor/asr/NeMo/nemo/core/classes/exportable.py(221): _export
/home/ik/Documents/aakhor/asr/NeMo/nemo/core/classes/exportable.py(114): export
/tmp/ipykernel_10033/3317768717.py(8): <module>
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3577): run_code
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3517): run_ast_nodes
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3334): run_cell_async
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/async_helpers.py(128): _pseudo_sync_runner
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3130): _run_cell
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3075): run_cell
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/zmqshell.py(549): run_cell
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/ipkernel.py(449): do_execute
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelbase.py(778): execute_request
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/ipkernel.py(362): execute_request
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelbase.py(437): dispatch_shell
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelbase.py(534): process_one
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelbase.py(545): dispatch_queue
/home/ik/miniconda3/envs/nemo/lib/python3.11/asyncio/events.py(84): _run
/home/ik/miniconda3/envs/nemo/lib/python3.11/asyncio/base_events.py(1936): _run_once
/home/ik/miniconda3/envs/nemo/lib/python3.11/asyncio/base_events.py(608): run_forever
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/tornado/platform/asyncio.py(205): start
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelapp.py(739): start
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/traitlets/config/application.py(1075): launch_instance
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel_launcher.py(18): <module>
<frozen runpy>(88): _run_code
<frozen runpy>(198): _run_module_as_main
)

    Inputs:
        #0: 86 defined in (%86 : Float(*, *, *, strides=[2946, 2946, 1], requires_grad=0, device=cpu) = onnx::Pad[mode=\"reflect\"](%59, %85), scope: nemo.collections.asr.modules.audio_preprocessing.AudioToMelSpectrogramPreprocessor::/nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures::featurizer # /home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/functional.py:5096:0
    )  (type 'Tensor')
        #1: 96 defined in (%96 : int[] = prim::ListConstruct(%90, %95), scope: nemo.collections.asr.modules.audio_preprocessing.AudioToMelSpectrogramPreprocessor::/nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures::featurizer
    )  (type 'List[int]')
    Outputs:
        #0: 97 defined in (%97 : Float(*, *, strides=[2946, 1], requires_grad=0, device=cpu) = onnx::Reshape[allowzero=0](%86, %96), scope: nemo.collections.asr.modules.audio_preprocessing.AudioToMelSpectrogramPreprocessor::/nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures::featurizer # /home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/functional.py:703:0
    )  (type 'Tensor')",
	"stack": "---------------------------------------------------------------------------
SymbolicValueError                        Traceback (most recent call last)
Cell In[10], line 8
      4 preprocessor = quartznet.preprocessor.to(device)
      6 preprocessor.eval()
----> 8 preprocessor.export(\"prep.onnx\", onnx_opset_version=18)

File ~/Documents/aakhor/asr/NeMo/nemo/core/classes/exportable.py:114, in Exportable.export(self, output, input_example, verbose, do_constant_folding, onnx_opset_version, check_trace, dynamic_axes, check_tolerance, export_modules_as_functions, keep_initializers_as_inputs)
    112 model = self.get_export_subnet(subnet_name)
    113 out_name = augment_filename(output, subnet_name)
--> 114 out, descr, out_example = model._export(
    115     out_name,
    116     input_example=input_example,
    117     verbose=verbose,
    118     do_constant_folding=do_constant_folding,
    119     onnx_opset_version=onnx_opset_version,
    120     check_trace=check_trace,
    121     dynamic_axes=dynamic_axes,
    122     check_tolerance=check_tolerance,
    123     export_modules_as_functions=export_modules_as_functions,
    124     keep_initializers_as_inputs=keep_initializers_as_inputs,
    125 )
    126 # Propagate input example (default scenario, may need to be overriden)
    127 if input_example is not None:

File ~/Documents/aakhor/asr/NeMo/nemo/core/classes/exportable.py:221, in Exportable._export(self, output, input_example, verbose, do_constant_folding, onnx_opset_version, check_trace, dynamic_axes, check_tolerance, export_modules_as_functions, keep_initializers_as_inputs)
    219     dynamic_axes = get_dynamic_axes(self.input_module.input_types_for_export, input_names)
    220     dynamic_axes.update(get_dynamic_axes(self.output_module.output_types_for_export, output_names))
--> 221 torch.onnx.export(
    222     jitted_model,
    223     input_example,
    224     output,
    225     input_names=input_names,
    226     output_names=output_names,
    227     verbose=verbose,
    228     do_constant_folding=do_constant_folding,
    229     dynamic_axes=dynamic_axes,
    230     opset_version=onnx_opset_version,
    231     keep_initializers_as_inputs=keep_initializers_as_inputs,
    232     export_modules_as_functions=export_modules_as_functions,
    233 )
    235 if check_trace:
    236     verify_runtime(self, output, check_trace_input, input_names, check_tolerance=check_tolerance)

File ~/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/__init__.py:375, in export(model, args, f, kwargs, export_params, verbose, input_names, output_names, opset_version, dynamic_axes, keep_initializers_as_inputs, dynamo, external_data, dynamic_shapes, report, verify, profile, dump_exported_program, artifacts_dir, fallback, training, operator_export_type, do_constant_folding, custom_opsets, export_modules_as_functions, autograd_inlining, **_)
    369 if dynamic_shapes:
    370     raise ValueError(
    371         \"The exporter only supports dynamic shapes \"
    372         \"through parameter dynamic_axes when dynamo=False.\"
    373     )
--> 375 export(
    376     model,
    377     args,
    378     f,  # type: ignore[arg-type]
    379     kwargs=kwargs,
    380     export_params=export_params,
    381     verbose=verbose is True,
    382     input_names=input_names,
    383     output_names=output_names,
    384     opset_version=opset_version,
    385     dynamic_axes=dynamic_axes,
    386     keep_initializers_as_inputs=keep_initializers_as_inputs,
    387     training=training,
    388     operator_export_type=operator_export_type,
    389     do_constant_folding=do_constant_folding,
    390     custom_opsets=custom_opsets,
    391     export_modules_as_functions=export_modules_as_functions,
    392     autograd_inlining=autograd_inlining,
    393 )
    394 return None

File ~/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py:502, in export(model, args, f, kwargs, export_params, verbose, training, input_names, output_names, operator_export_type, opset_version, do_constant_folding, dynamic_axes, keep_initializers_as_inputs, custom_opsets, export_modules_as_functions, autograd_inlining)
    499 if kwargs is not None:
    500     args = args + (kwargs,)
--> 502 _export(
    503     model,
    504     args,
    505     f,
    506     export_params,
    507     verbose,
    508     training,
    509     input_names,
    510     output_names,
    511     operator_export_type=operator_export_type,
    512     opset_version=opset_version,
    513     do_constant_folding=do_constant_folding,
    514     dynamic_axes=dynamic_axes,
    515     keep_initializers_as_inputs=keep_initializers_as_inputs,
    516     custom_opsets=custom_opsets,
    517     export_modules_as_functions=export_modules_as_functions,
    518     autograd_inlining=autograd_inlining,
    519 )
    521 return None

File ~/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py:1564, in _export(model, args, f, export_params, verbose, training, input_names, output_names, operator_export_type, export_type, opset_version, do_constant_folding, dynamic_axes, keep_initializers_as_inputs, fixed_batch_size, custom_opsets, add_node_names, onnx_shape_inference, export_modules_as_functions, autograd_inlining)
   1561     dynamic_axes = {}
   1562 _validate_dynamic_axes(dynamic_axes, model, input_names, output_names)
-> 1564 graph, params_dict, torch_out = _model_to_graph(
   1565     model,
   1566     args,
   1567     verbose,
   1568     input_names,
   1569     output_names,
   1570     operator_export_type,
   1571     val_do_constant_folding,
   1572     fixed_batch_size=fixed_batch_size,
   1573     training=training,
   1574     dynamic_axes=dynamic_axes,
   1575 )
   1577 # TODO: Don't allocate a in-memory string for the protobuf
   1578 defer_weight_export = (
   1579     export_type is not _exporter_states.ExportTypes.PROTOBUF_FILE
   1580 )

File ~/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py:1117, in _model_to_graph(model, args, verbose, input_names, output_names, operator_export_type, do_constant_folding, _disable_torch_constant_prop, fixed_batch_size, training, dynamic_axes)
   1114 params_dict = _get_named_param_dict(graph, params)
   1116 try:
-> 1117     graph = _optimize_graph(
   1118         graph,
   1119         operator_export_type,
   1120         _disable_torch_constant_prop=_disable_torch_constant_prop,
   1121         fixed_batch_size=fixed_batch_size,
   1122         params_dict=params_dict,
   1123         dynamic_axes=dynamic_axes,
   1124         input_names=input_names,
   1125         module=module,
   1126     )
   1127 except Exception as e:
   1128     torch.onnx.log(\"Torch IR graph at exception: \", graph)

File ~/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py:639, in _optimize_graph(graph, operator_export_type, _disable_torch_constant_prop, fixed_batch_size, params_dict, dynamic_axes, input_names, module)
    636     _C._jit_pass_onnx_set_dynamic_input_shape(graph, dynamic_axes, input_names)
    637 _C._jit_pass_onnx_lint(graph)
--> 639 graph = _C._jit_pass_onnx(graph, operator_export_type)
    640 _C._jit_pass_onnx_lint(graph)
    641 _C._jit_pass_lint(graph)

File ~/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py:1836, in _run_symbolic_function(graph, block, node, inputs, env, values_in_env, new_nodes, operator_export_type)
   1831     if symbolic_fn is not None:
   1832         # TODO Wrap almost identical attrs assignment or comment the difference.
   1833         attrs = {
   1834             k: symbolic_helper._node_get(node, k) for k in node.attributeNames()
   1835         }
-> 1836         return symbolic_fn(graph_context, *inputs, **attrs)
   1838 attrs = {
   1839     k + \"_\" + node.kindOf(k)[0]: symbolic_helper._node_get(node, k)
   1840     for k in node.attributeNames()
   1841 }
   1842 if namespace == \"onnx\":
   1843     # Clone node to trigger ONNX shape inference

File ~/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/symbolic_helper.py:281, in parse_args.<locals>.decorator.<locals>.wrapper(g, *args, **kwargs)
    275 if len(kwargs) == 1:
    276     assert \"_outputs\" in kwargs, (
    277         f\"Symbolic function {fn.__name__}'s '**kwargs' can only contain \"
    278         f\"'_outputs' key at '**kwargs'. \"
    279         f\"{FILE_BUG_MSG}\"
    280     )
--> 281 return fn(g, *args, **kwargs)

File ~/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/symbolic_opset17.py:135, in stft(g, input, n_fft, hop_length, win_length, window, normalized, onesided, return_complex)
    133 # Checks
    134 if return_complex:
--> 135     raise errors.SymbolicValueError(
    136         msg=\"STFT does not currently support complex types\", value=input
    137     )
    139 # Get STFT sizes
    140 frame_step_value = hop_length if hop_length is not None else n_fft // 4

SymbolicValueError: STFT does not currently support complex types  [Caused by the value '97 defined in (%97 : Float(*, *, strides=[2946, 1], requires_grad=0, device=cpu) = onnx::Reshape[allowzero=0](%86, %96), scope: nemo.collections.asr.modules.audio_preprocessing.AudioToMelSpectrogramPreprocessor::/nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures::featurizer # /home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/functional.py:703:0
)' (type 'Tensor') in the TorchScript graph. The containing node has kind 'onnx::Reshape'.] 
    (node defined in /home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/functional.py(703): stft
/home/ik/Documents/aakhor/asr/NeMo/nemo/collections/asr/parts/preprocessing/features.py(308): <lambda>
/home/ik/Documents/aakhor/asr/NeMo/nemo/collections/asr/parts/preprocessing/features.py(423): forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/utils/_contextlib.py(116): decorate_context
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1726): _slow_forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1747): _call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1736): _wrapped_call_impl
/home/ik/Documents/aakhor/asr/NeMo/nemo/collections/asr/modules/audio_preprocessing.py(292): get_features
/home/ik/Documents/aakhor/asr/NeMo/nemo/collections/asr/modules/audio_preprocessing.py(91): forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/utils/_contextlib.py(116): decorate_context
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1726): _slow_forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1747): _call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1736): _wrapped_call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/jit/_trace.py(130): wrapper
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/jit/_trace.py(139): forward
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1747): _call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/modules/module.py(1736): _wrapped_call_impl
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/jit/_trace.py(1500): _get_trace_graph
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(904): _trace_and_get_graph_from_model
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(997): _create_jit_graph
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(1113): _model_to_graph
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(1564): _export
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/utils.py(502): export
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/onnx/__init__.py(375): export
/home/ik/Documents/aakhor/asr/NeMo/nemo/core/classes/exportable.py(221): _export
/home/ik/Documents/aakhor/asr/NeMo/nemo/core/classes/exportable.py(114): export
/tmp/ipykernel_10033/3317768717.py(8): <module>
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3577): run_code
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3517): run_ast_nodes
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3334): run_cell_async
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/async_helpers.py(128): _pseudo_sync_runner
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3130): _run_cell
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/IPython/core/interactiveshell.py(3075): run_cell
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/zmqshell.py(549): run_cell
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/ipkernel.py(449): do_execute
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelbase.py(778): execute_request
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/ipkernel.py(362): execute_request
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelbase.py(437): dispatch_shell
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelbase.py(534): process_one
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelbase.py(545): dispatch_queue
/home/ik/miniconda3/envs/nemo/lib/python3.11/asyncio/events.py(84): _run
/home/ik/miniconda3/envs/nemo/lib/python3.11/asyncio/base_events.py(1936): _run_once
/home/ik/miniconda3/envs/nemo/lib/python3.11/asyncio/base_events.py(608): run_forever
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/tornado/platform/asyncio.py(205): start
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel/kernelapp.py(739): start
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/traitlets/config/application.py(1075): launch_instance
/home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/ipykernel_launcher.py(18): <module>
<frozen runpy>(88): _run_code
<frozen runpy>(198): _run_module_as_main
)

    Inputs:
        #0: 86 defined in (%86 : Float(*, *, *, strides=[2946, 2946, 1], requires_grad=0, device=cpu) = onnx::Pad[mode=\"reflect\"](%59, %85), scope: nemo.collections.asr.modules.audio_preprocessing.AudioToMelSpectrogramPreprocessor::/nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures::featurizer # /home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/nn/functional.py:5096:0
    )  (type 'Tensor')
        #1: 96 defined in (%96 : int[] = prim::ListConstruct(%90, %95), scope: nemo.collections.asr.modules.audio_preprocessing.AudioToMelSpectrogramPreprocessor::/nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures::featurizer
    )  (type 'List[int]')
    Outputs:
        #0: 97 defined in (%97 : Float(*, *, strides=[2946, 1], requires_grad=0, device=cpu) = onnx::Reshape[allowzero=0](%86, %96), scope: nemo.collections.asr.modules.audio_preprocessing.AudioToMelSpectrogramPreprocessor::/nemo.collections.asr.parts.preprocessing.features.FilterbankFeatures::featurizer # /home/ik/miniconda3/envs/nemo/lib/python3.11/site-packages/torch/functional.py:703:0
    )  (type 'Tensor')"
}

Expected behavior

The preprocessor should export to onnx successfully.

Environment overview (please complete the following information)

• Environment location: Local Ubuntu 22.04
• Method of NeMo install: from source using pip. Installed in editable mode.

Environment details

• OS version Ubuntu 22.04
• PyTorch version 2.5.1+cu124
• Python version 3.11.10

Additional context

The full model exports fine though. But that does not contain the preprocessor.