onnx export error #107922
Comments
|
Please share the onnx version you use. ONNX Script requires a onnx version >= 1.14. Thanks! |
windows: |
linux: |
|
Could you install the latest onnx (1.15) and onnxscript (uninstall onnxscript-preview first)? Thanks! |
|
Traceback (most recent call last): |
|
pip install --upgrade onnxscript |
|
pip install packaging. We will fix this in the next release |
|
PS F:\ai_export> pip install --upgrade onnxscript The above exception was the direct cause of the following exception: Traceback (most recent call last): |
after install, it still not work |
after install, it still not work |
|
dynamo doesn’t support Windows yet. I suggest using WSL or if you like hack the check_if_dynamo_supported function above to always return True |
|
Looks like duplicate of #112844 |
|
pip install --upgrade onnxscript packaging The above exception was the direct cause of the following exception: Traceback (most recent call last): |
use wsl still not work |
|
Please check the function call as the signature has changed: There are also known issues reported in #112844 |
python ./main.py from user code: Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information The above exception was the direct cause of the following exception: Traceback (most recent call last): |
|
import torch class DataCov(nn.Module): def export(): if name == 'main': |
|
{ |
|
@WangHHY19931001 Hi, Did you sloved this error? |
it still error /home/yafeng_wang/miniconda3/envs/onnx_export/lib/python3.12/site-packages/torch/onnx/_internal/exporter.py:137: UserWarning: torch.onnx.dynamo_export only implements opset version 18 for now. If you need to use a different opset version, please register them with register_custom_op. The above exception was the direct cause of the following exception: Traceback (most recent call last): |
|
{ |
|
Set following environment variable. https://github.com/pytorch/pytorch/blob/main/torch/onnx/__init__.py#L473 |
still not work |
As my checking, dynamo export works, but next code in PyTorch tutorials do not work, because all code there supposed to use legacy exporter. |
|
We are in the process of updating the tutorials. For now, please test with |
Yes, as I said export works. I could export my model. |
PyTorch ONNX Conversion ReportError messagesNo errorsExported programExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, b_transform_0_spectrogram_window: "f32[1536]", b_transform_0_mel_scale_fb: "f32[769, 128]", x1: "f32[1, 1, 576000]"):
# File: /home/dingli/miniconda3/envs/ai_mos_export_onnx/lib/python3.12/site-packages/torchaudio/transforms/_transforms.py:110 in forward, code: return F.spectrogram(
view: "f32[1, 576000]" = torch.ops.aten.view.default(x1, [-1, 576000]); x1 = None
view_1: "f32[1, 1, 576000]" = torch.ops.aten.view.default(view, [1, 1, 576000]); view = None
reflection_pad1d: "f32[1, 1, 577536]" = torch.ops.aten.reflection_pad1d.default(view_1, [768, 768]); view_1 = None
view_2: "f32[1, 577536]" = torch.ops.aten.view.default(reflection_pad1d, [1, 577536]); reflection_pad1d = None
unfold: "f32[1, 751, 1536]" = torch.ops.aten.unfold.default(view_2, -1, 1536, 768); view_2 = None
mul: "f32[1, 751, 1536]" = torch.ops.aten.mul.Tensor(unfold, b_transform_0_spectrogram_window); unfold = b_transform_0_spectrogram_window = None
_fft_r2c: "c64[1, 751, 769]" = torch.ops.aten._fft_r2c.default(mul, [2], 0, True); mul = None
transpose: "c64[1, 769, 751]" = torch.ops.aten.transpose.int(_fft_r2c, 1, 2); _fft_r2c = None
view_3: "c64[1, 1, 769, 751]" = torch.ops.aten.view.default(transpose, [1, 1, 769, 751]); transpose = None
abs_1: "f32[1, 1, 769, 751]" = torch.ops.aten.abs.default(view_3); view_3 = None
pow_1: "f32[1, 1, 769, 751]" = torch.ops.aten.pow.Tensor_Scalar(abs_1, 2.0); abs_1 = None
# File: /home/dingli/miniconda3/envs/ai_mos_export_onnx/lib/python3.12/site-packages/torchaudio/transforms/_transforms.py:412 in forward, code: mel_specgram = torch.matmul(specgram.transpose(-1, -2), self.fb).transpose(-1, -2)
transpose_1: "f32[1, 1, 751, 769]" = torch.ops.aten.transpose.int(pow_1, -1, -2); pow_1 = None
matmul: "f32[1, 1, 751, 128]" = torch.ops.aten.matmul.default(transpose_1, b_transform_0_mel_scale_fb); transpose_1 = b_transform_0_mel_scale_fb = None
transpose_2: "f32[1, 1, 128, 751]" = torch.ops.aten.transpose.int(matmul, -1, -2); matmul = None
return (transpose_2,)
Graph signature: ExportGraphSignature(input_specs=[InputSpec(kind=<InputKind.BUFFER: 3>, arg=TensorArgument(name='b_transform_0_spectrogram_window'), target='transform.0.spectrogram.window', persistent=True), InputSpec(kind=<InputKind.BUFFER: 3>, arg=TensorArgument(name='b_transform_0_mel_scale_fb'), target='transform.0.mel_scale.fb', persistent=True), InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='x1'), target=None, persistent=None)], output_specs=[OutputSpec(kind=<OutputKind.USER_OUTPUT: 1>, arg=TensorArgument(name='transpose_2'), target=None)])
Range constraints: {}ONNX model<
ir_version=10,
opset_imports={'pkg.onnxscript.torch_lib.common': 1, '': 18, 'pkg.onnxscript.torch_lib': 1},
producer_name='pytorch',
producer_version='2.7.0.dev20250107+cpu',
domain=None,
model_version=None,
>
graph(
name=main_graph,
inputs=(
%"wav_data"<FLOAT,[1,1,576000]>
),
outputs=(
%"ans"<FLOAT,[1,1,128,751]>
),
initializers=(
%"transform.0.spectrogram.window"<FLOAT,[1536]>,
%"transform.0.mel_scale.fb"<FLOAT,[769,128]>
),
) {
0 | # node_Constant_0
%"val_0"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[2]>(array([ -1, 576000]), name=None)}
1 | # node_Cast_1
%"val_1"<?,?> ⬅️ ::Cast(%"val_0") {to=INT64}
2 | # node_Reshape_2
%"view"<FLOAT,[1,576000]> ⬅️ ::Reshape(%"wav_data", %"val_1") {allowzero=0}
3 | # node_Constant_3
%"val_2"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[3]>(array([ 1, 1, 576000]), name=None)}
4 | # node_Cast_4
%"val_3"<?,?> ⬅️ ::Cast(%"val_2") {to=INT64}
5 | # node_Reshape_5
%"view_1"<FLOAT,[1,1,576000]> ⬅️ ::Reshape(%"view", %"val_3") {allowzero=0}
6 | # node_Constant_6
%"val_4"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[2]>(array([768, 768]), name=None)}
7 | # node_aten_reflection_pad1d_7
%"reflection_pad1d"<FLOAT,[1,1,577536]> ⬅️ pkg.onnxscript.torch_lib::aten_reflection_pad1d(%"view_1", %"val_4")
8 | # node_Constant_8
%"val_5"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[2]>(array([ 1, 577536]), name=None)}
9 | # node_Cast_9
%"val_6"<?,?> ⬅️ ::Cast(%"val_5") {to=INT64}
10 | # node_Reshape_10
%"view_2"<FLOAT,[1,577536]> ⬅️ ::Reshape(%"reflection_pad1d", %"val_6") {allowzero=0}
11 | # node__aten_unfold_onnx_11
%"unfold"<FLOAT,[1,751,1536]> ⬅️ pkg.onnxscript.torch_lib::_aten_unfold_onnx(%"view_2") {dim=1, size=1536, step=768, target_end=751, perm=[0, 1, 2]}
12 | # node_Mul_12
%"mul"<FLOAT,[1,751,1536]> ⬅️ ::Mul(%"unfold", %"transform.0.spectrogram.window")
13 | # node_Constant_13
%"val_7"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[1]>(array([-1]), name=None)}
14 | # node_Unsqueeze_14
%"val_8"<?,?> ⬅️ ::Unsqueeze(%"mul", %"val_7")
15 | # node_Constant_15
%"val_9"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[1]>(array([0]), name=None)}
16 | # node_Unsqueeze_16
%"val_10"<?,?> ⬅️ ::Unsqueeze(%"val_8", %"val_9")
17 | # node_DFT_17
%"val_11"<?,?> ⬅️ ::DFT(%"val_10") {axis=3, inverse=False, onesided=True}
18 | # node_Squeeze_18
%"_fft_r2c"<FLOAT,[1,751,769,2]> ⬅️ ::Squeeze(%"val_11", %"val_9")
19 | # node_Shape_19
%"val_12"<?,?> ⬅️ ::Shape(%"val_8") {start=0}
20 | # node_Constant_20
%"val_13"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[1]>(array([2]), name=None)}
21 | # node_Gather_21
%"val_14"<?,?> ⬅️ ::Gather(%"val_12", %"val_13") {axis=0}
22 | # node_ReduceProd_22
%"val_15"<?,?> ⬅️ ::ReduceProd(%"val_14") {keepdims=0, noop_with_empty_axes=0}
23 | # node_CastLike_23
%"val_16"<?,?> ⬅️ ::CastLike(%"val_15", %"_fft_r2c")
24 | # node_Transpose_24
%"transpose"<FLOAT,[1,769,751,2]> ⬅️ ::Transpose(%"_fft_r2c") {perm=[0, 2, 1, 3]}
25 | # node_Constant_25
%"val_17"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[4]>(array([ 1, 1, 769, 751]), name=None)}
26 | # node_aten_view_complex_26
%"view_3"<FLOAT,[1,1,769,751,2]> ⬅️ pkg.onnxscript.torch_lib::aten_view_complex(%"transpose", %"val_17")
27 | # node_ReduceL2_27
%"abs_1"<FLOAT,[1,1,769,751]> ⬅️ ::ReduceL2(%"view_3", %"val_7") {keepdims=False, noop_with_empty_axes=0}
28 | # node_Constant_28
%"val_18"<?,?> ⬅️ ::Constant() {value=Tensor<FLOAT,[]>(array(2., dtype=float32), name=None)}
29 | # node_Pow_29
%"pow_1"<FLOAT,[1,1,769,751]> ⬅️ ::Pow(%"abs_1", %"val_18")
30 | # node_Transpose_30
%"transpose_1"<FLOAT,[1,1,751,769]> ⬅️ ::Transpose(%"pow_1") {perm=[0, 1, 3, 2]}
31 | # node_aten_matmul_31
%"matmul"<FLOAT,[1,1,751,128]> ⬅️ pkg.onnxscript.torch_lib::aten_matmul(%"transpose_1", %"transform.0.mel_scale.fb")
32 | # node_Transpose_32
%"ans"<FLOAT,[1,1,128,751]> ⬅️ ::Transpose(%"matmul") {perm=[0, 1, 3, 2]}
return %"ans"<FLOAT,[1,1,128,751]>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib::aten_reflection_pad1d(
inputs=(
%"self"<?,?>,
%"padding"<?,?>
),
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"int64_0_1d"<?,?> ⬅️ ::Constant() {value=TensorProtoTensor<INT64,[1]>(name='int64_0_1d')}
1 | # n1
%"int64_1_1d"<?,?> ⬅️ ::Constant() {value=TensorProtoTensor<INT64,[1]>(name='int64_1_1d')}
2 | # n2
%"int64_0_1d_0"<?,?> ⬅️ ::Constant() {value=TensorProtoTensor<INT64,[1]>(name='int64_0_1d_0')}
3 | # n3
%"start"<?,?> ⬅️ ::Slice(%"padding", %"int64_0_1d", %"int64_1_1d", %"int64_0_1d_0")
4 | # n4
%"int64_1_1d_1"<?,?> ⬅️ ::Constant() {value=TensorProtoTensor<INT64,[1]>(name='int64_1_1d_1')}
5 | # n5
%"int64_2_1d"<?,?> ⬅️ ::Constant() {value=TensorProtoTensor<INT64,[1]>(name='int64_2_1d')}
6 | # n6
%"int64_0_1d_2"<?,?> ⬅️ ::Constant() {value=TensorProtoTensor<INT64,[1]>(name='int64_0_1d_2')}
7 | # n7
%"end"<?,?> ⬅️ ::Slice(%"padding", %"int64_1_1d_1", %"int64_2_1d", %"int64_0_1d_2")
8 | # n8
%"tmp"<?,?> ⬅️ ::Constant() {value_ints=[0]}
9 | # n9
%"tmp_3"<?,?> ⬅️ ::Constant() {value_ints=[0]}
10 | # n10
%"padding_onnx"<?,?> ⬅️ ::Concat(%"tmp", %"start", %"tmp_3", %"end") {axis=0}
11 | # n11
%"return_val"<?,?> ⬅️ ::Pad(%"self", %"padding_onnx") {mode=reflect}
return %"return_val"<?,?>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib::_aten_unfold_onnx(
inputs=(
%"self"<?,?>
),
attributes={
dim: UNDEFINED,
size: UNDEFINED,
step: UNDEFINED,
target_end: UNDEFINED,
perm: UNDEFINED
}
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"tmp"<?,?> ⬅️ ::Constant() {value_int=RefAttr('value_int', INT, ref_attr_name='dim')}
1 | # n1
%"tmp_0"<?,?> ⬅️ ::Constant() {value_ints=[-1]}
2 | # n2
%"dims"<?,?> ⬅️ ::Reshape(%"tmp", %"tmp_0")
3 | # n3
%"seq_result"<?,?> ⬅️ ::SequenceEmpty()
4 | # n4
%"i"<?,?> ⬅️ ::Constant() {value_int=0}
5 | # n5
%"target_end"<?,?> ⬅️ ::Constant() {value_int=RefAttr('value_int', INT, ref_attr_name='target_end')}
6 | # n6
%"target_end_cast"<?,?> ⬅️ ::CastLike(%"target_end", %"i")
7 | # n7
%"cond"<?,?> ⬅️ ::Less(%"i", %"target_end_cast")
8 | # n8
%"seq_result_8"<?,?>, %"i_9"<?,?> ⬅️ ::Loop(None, %"cond", %"seq_result", %"i") {body=
graph(
name=loop_body,
inputs=(
%"infinite_loop"<INT64,[]>,
%"cond"<BOOL,[]>,
%"seq_result_1"<?,?>,
%"i_2"<?,?>
),
outputs=(
%"cond_out"<BOOL,[]>,
%"seq_result_4"<?,?>,
%"i_5"<?,?>
),
) {
0 | # n0
%"step"<?,?> ⬅️ ::Constant() {value_int=RefAttr('value_int', INT, ref_attr_name='step')}
1 | # n1
%"step_cast"<?,?> ⬅️ ::CastLike(%"step", %"i_2")
2 | # n2
%"tmp_3"<?,?> ⬅️ ::Mul(%"i_2", %"step_cast")
3 | # n3
%"int64_m1_1d"<?,?> ⬅️ ::Constant() {value=TensorProtoTensor<INT64,[1]>(name='int64_m1_1d')}
4 | # n4
%"starts"<?,?> ⬅️ ::Reshape(%"tmp_3", %"int64_m1_1d")
5 | # n5
%"size"<?,?> ⬅️ ::Constant() {value_int=RefAttr('value_int', INT, ref_attr_name='size')}
6 | # n6
%"size_cast"<?,?> ⬅️ ::CastLike(%"size", %"starts")
7 | # n7
%"ends"<?,?> ⬅️ ::Add(%"starts", %"size_cast")
8 | # n8
%"slice_result"<?,?> ⬅️ ::Slice(%"self", %"starts", %"ends", %"dims")
9 | # n9
%"slice_result_float32"<?,?> ⬅️ ::Cast(%"slice_result") {to=1}
10 | # n10
%"seq_result_4"<?,?> ⬅️ ::SequenceInsert(%"seq_result_1", %"slice_result_float32")
11 | # n11
%"int64_1"<?,?> ⬅️ ::Constant() {value=TensorProtoTensor<INT64,[]>(name='int64_1')}
12 | # n12
%"int64_1_cast"<?,?> ⬅️ ::CastLike(%"int64_1", %"i_2")
13 | # n13
%"i_5"<?,?> ⬅️ ::Add(%"i_2", %"int64_1_cast")
14 | # n14
%"target_end_6"<?,?> ⬅️ ::Constant() {value_int=RefAttr('value_int', INT, ref_attr_name='target_end')}
15 | # n15
%"target_end_6_cast"<?,?> ⬅️ ::CastLike(%"target_end_6", %"i_5")
16 | # n16
%"cond_7"<?,?> ⬅️ ::Less(%"i_5", %"target_end_6_cast")
17 | # n17
%"cond_out"<BOOL,[]> ⬅️ ::Identity(%"cond_7")
return %"cond_out"<BOOL,[]>, %"seq_result_4"<?,?>, %"i_5"<?,?>
}}
9 | # n9
%"concat_result"<?,?> ⬅️ ::ConcatFromSequence(%"seq_result_8") {axis=RefAttr('axis', INT, ref_attr_name='dim'), new_axis=1}
10 | # n10
%"result"<?,?> ⬅️ ::Transpose(%"concat_result") {perm=RefAttr('perm', INTS, ref_attr_name='perm')}
11 | # n11
%"return_val"<?,?> ⬅️ ::CastLike(%"result", %"self")
return %"return_val"<?,?>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib::aten_view_complex(
inputs=(
%"self"<?,?>,
%"size"<?,?>
),
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"size_0"<?,?> ⬅️ ::Cast(%"size") {to=7}
1 | # n1
%"tmp"<?,?> ⬅️ ::Constant() {value_ints=[2]}
2 | # n2
%"complex_size"<?,?> ⬅️ ::Concat(%"size_0", %"tmp") {axis=0}
3 | # n3
%"return_val"<?,?> ⬅️ ::Reshape(%"self", %"complex_size")
return %"return_val"<?,?>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib::aten_matmul(
inputs=(
%"self"<?,?>,
%"other"<?,?>
),
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"return_val"<?,?> ⬅️ ::MatMul(%"self", %"other")
return %"return_val"<?,?>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib.common::Rank(
inputs=(
%"input"<?,?>
),
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"tmp"<?,?> ⬅️ ::Shape(%"input")
1 | # n1
%"return_val"<?,?> ⬅️ ::Size(%"tmp")
return %"return_val"<?,?>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib.common::IsScalar(
inputs=(
%"input"<?,?>
),
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"tmp"<?,?> ⬅️ ::Shape(%"input")
1 | # n1
%"tmp_0"<?,?> ⬅️ ::Size(%"tmp")
2 | # n2
%"tmp_1"<?,?> ⬅️ ::Constant() {value_int=0}
3 | # n3
%"return_val"<?,?> ⬅️ ::Equal(%"tmp_0", %"tmp_1")
return %"return_val"<?,?>
}AnalysisPyTorch ONNX Conversion Analysis Model InformationThe model has 0 parameters and 99968 buffers (non-trainable parameters). defaultdict(<class 'int'>, {})Number of buffers per dtype: defaultdict(<class 'int'>, {torch.float32: 99968})Inputs:
Outputs:
The FX graph has 18 nodes in total. Number of FX nodes per op:
Of the call_function nodes, the counts of operators used are:
ONNX Conversion InformationAll operators in the model have registered ONNX decompositions. Decomposition comparisonOps exist only in the ExportedProgram before decomposition: Ops exist only in the ExportedProgram after decomposition: |
|
Collecting environment information... OS: Ubuntu 24.04.1 LTS (x86_64) Python version: 3.12.8 | packaged by Anaconda, Inc. | (main, Dec 11 2024, 16:31:09) [GCC 11.2.0] (64-bit runtime) CPU: Versions of relevant libraries: |
|
still has some error with test: test code : def test_data_cov_onnx(onnx_path):
sess_options = ort.SessionOptions()
sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
providers = [
'CUDAExecutionProvider',
'DmlExecutionProvider',
'CPUExecutionProvider'
]
session = ort.InferenceSession(onnx_path, sess_options,
providers=providers)
src_wav = torch.randn((1, 1, 48000 * 12))
ort_inputs = {session.get_inputs()[0].name: src_wav.numpy(), }
ort_outs = session.run(None, ort_inputs)
ort_outs = ort_outs[0]
ort_outs = torch.from_numpy(ort_outs)
model = DataCov()
model.eval()
deal_1 = model(src_wav)
print(f'Torch Output Shape: {deal_1.shape}, ONNX Output Shape: {ort_outs.shape}')
print(f'Torch Output Min/Max: {torch.min(deal_1)}, {torch.max(deal_1)}')
print(f'ONNX Output Min/Max: {torch.min(ort_outs)}, {torch.max(ort_outs)}')
print(f'Torch Output Mean/Std: {torch.mean(deal_1)}, {torch.std(deal_1)}')
print(f'ONNX Output Mean/Std: {torch.mean(ort_outs)}, {torch.std(ort_outs)}')
np.testing.assert_allclose(deal_1.detach().numpy(), ort_outs.detach().numpy(), rtol=1e-02, atol=1e-04)onnx export code: def export_datacov_onnx(path):
model = DataCov()
model.eval()
src_wav = torch.randn((1, 1, 48000 * 12), requires_grad=True)
input_names = ["wav_data"]
output_names = ["ans"]
args = (src_wav,)
torch.onnx.export(
model,
args,
path,
export_params=True,
opset_version=19,
do_constant_folding=True,
verbose=False,
input_names=input_names,
output_names=output_names,
dynamo=True,
report=True
)
onnx_model = onnx.load(path)
onnx.checker.check_model(onnx_model)model: class DataCov(nn.Module):
def __init__(self):
super(DataCov, self).__init__()
self.transform = nn.Sequential(
torchaudio.transforms.MelSpectrogram(sample_rate=48000, n_fft=1536, hop_length=768, f_min=20, f_max=20000)
)
def forward(self, x1):
return self.transform(x1) |
🐛 Describe the bug
code:
linux error:
windows error:
Versions
linux versions:
windows version:
The text was updated successfully, but these errors were encountered: