InternImage
InternImage copied to clipboard
OpenGVLab
Cannot export classification model to ONNX
I tried to export a classification model to ONNX model according to README in the classification directory. But it was failed with following error messages.
(internimage) ubuntu@ip-10-0-0-189:~/InternImage/classification$ python export.py --model_name internimage_t_1k_224 --ckpt_dir ckpt_dir/ --onnx
=> merge config from ./configs/internimage_t_1k_224.yaml
using core type: DCNv3
using activation layer: GELU
using main norm layer: LN
using dpr: linear, 0.2
level2_post_norm: False
level2_post_norm_block_ids: None
res_post_norm: False
remove_center: False
Traceback (most recent call last):
File "export.py", line 121, in <module>
main()
File "export.py", line 112, in main
torch2onnx(args, cfg)
File "export.py", line 65, in torch2onnx
output_names=['output'])
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/__init__.py", line 309, in export
export_modules_as_functions)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 122, in export
custom_opsets=custom_opsets, export_modules_as_functions=export_modules_as_functions)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 724, in _export
dynamic_axes=dynamic_axes)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 507, in _model_to_graph
module=module)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 232, in _optimize_graph
graph = torch._C._jit_pass_onnx(graph, operator_export_type)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/__init__.py", line 359, in _run_symbolic_method
return utils._run_symbolic_method(*args, **kwargs)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 846, in _run_symbolic_method
return symbolic_fn(g, *args)
File "/home/ubuntu/InternImage/classification/ops_dcnv3/functions/dcnv3_func.py", line 105, in symbolic
remove_center=int(remove_center),
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 952, in _graph_op
n = g.insertNode(_newNode(g, opname, outputs, *args, **kwargs))
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 905, in _newNode
_add_attribute(n, k, v, aten=aten)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 869, in _add_attribute
" must be suffixed with type, e.g. 'dim_i' or 'dims_i'").format(key))
IndexError: Invalid attribute specifier 'remove_center' names must be suffixed with type, e.g. 'dim_i' or 'dims_i'
The environment is
- AWS EC2
- using AWS Deep Learning AMI GPU CUDA 11.3 (Ubuntu 20.04)
- Anaconda 2023.09-0 (September 29, 2023)
- others are same as written in README
I found that changes of commit 8f2d158399b95b2188b86dd46da93c25c2dbb7cc have probably caused this problem.
https://github.com/OpenGVLab/InternImage/commit/8f2d158399b95b2188b86dd46da93c25c2dbb7cc
$ python export.py --model_name internimage_t_1k_224 --ckpt_dir ckpt_dir/ --onnx
=> merge config from ./configs/internimage_t_1k_224.yaml
using core type: DCNv3
using activation layer: GELU
using main norm layer: LN
using dpr: linear, 0.2
level2_post_norm: False
level2_post_norm_block_ids: None
res_post_norm: False
remove_center: False
Traceback (most recent call last):
File "export.py", line 121, in <module>
main()
File "export.py", line 112, in main
torch2onnx(args, cfg)
File "export.py", line 65, in torch2onnx
output_names=['output'])
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/__init__.py", line 309, in export
export_modules_as_functions)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 122, in export
custom_opsets=custom_opsets, export_modules_as_functions=export_modules_as_functions)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 724, in _export
dynamic_axes=dynamic_axes)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 499, in _model_to_graph
graph, params, torch_out, module = _create_jit_graph(model, args)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 440, in _create_jit_graph
graph, torch_out = _trace_and_get_graph_from_model(model, args)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/onnx/utils.py", line 391, in _trace_and_get_graph_from_model
torch.jit._get_trace_graph(model, args, strict=False, _force_outplace=False, _return_inputs_states=True)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/jit/_trace.py", line 1166, in _get_trace_graph
outs = ONNXTracedModule(f, strict, _force_outplace, return_inputs, _return_inputs_states)(*args, **kwargs)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/jit/_trace.py", line 132, in forward
self._force_outplace,
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/jit/_trace.py", line 118, in wrapper
outs.append(self.inner(*trace_inputs))
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1098, in _slow_forward
result = self.forward(*input, **kwargs)
File "/home/ubuntu/InternImage/classification/models/intern_image.py", line 766, in forward
x = self.forward_features(x)
File "/home/ubuntu/InternImage/classification/models/intern_image.py", line 725, in forward_features
x = level(x)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1098, in _slow_forward
result = self.forward(*input, **kwargs)
File "/home/ubuntu/InternImage/classification/models/intern_image.py", line 512, in forward
x = blk(x)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1098, in _slow_forward
result = self.forward(*input, **kwargs)
File "/home/ubuntu/InternImage/classification/models/intern_image.py", line 430, in forward
x = _inner_forward(x)
File "/home/ubuntu/InternImage/classification/models/intern_image.py", line 416, in _inner_forward
x = x + self.drop_path(self.dcn(self.norm1(x)))
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1098, in _slow_forward
result = self.forward(*input, **kwargs)
File "/home/ubuntu/InternImage/classification/ops_dcnv3/modules/dcnv3.py", line 345, in forward
self.remove_center)
File "/home/ubuntu/anaconda3/envs/internimage/lib/python3.7/site-packages/torch/cuda/amp/autocast_mode.py", line 110, in decorate_fwd
return fwd(*args, **kwargs)
File "/home/ubuntu/InternImage/classification/ops_dcnv3/functions/dcnv3_func.py", line 49, in forward
group_channels, offset_scale, ctx.im2col_step, **kwargs)
TypeError: dcnv3_forward(): incompatible function arguments. The following argument types are supported:
1. (arg0: at::Tensor, arg1: at::Tensor, arg2: at::Tensor, arg3: int, arg4: int, arg5: int, arg6: int, arg7: int, arg8: int, arg9: int, arg10: int, arg11: int, arg12: int, arg13: float, arg14: int, arg15: int) -> at::Tensor
Invoked with: tensor([[[[-4.6765e-01, -4.2197e-01, 6.0807e-02, ..., 1.0714e-01,
5.4229e-01, 1.9236e-02],
[-1.0968e-01, 2.8766e-01, 2.4005e-01, ..., -6.1618e-02,
7.7031e-01, 5.1768e-02],
[-3.8835e-01, 3.7148e-01, -1.2781e-01, ..., -2.2630e-01,
4.8191e-01, 3.8509e-02],
...,
[-2.7780e-01, 5.5153e-01, -1.3241e-01, ..., -2.0382e-01,
4.1559e-01, 3.0929e-01],
[-4.4095e-01, 5.9974e-01, 6.8127e-03, ..., 6.2164e-03,
2.8223e-01, 7.7597e-02],
[-4.4174e-02, 3.5850e-01, 2.4612e-01, ..., -3.9771e-01,
8.4857e-01, -3.9927e-02]],
[[ 5.4233e-02, -1.1951e-01, -1.0995e-01, ..., 2.2794e-01,
3.4253e-01, 1.1734e-01],
[-1.9957e-01, -4.4482e-01, -5.7037e-01, ..., -2.7812e-01,
-3.8517e-01, 1.2238e-01],
[ 3.0182e-01, -1.4350e-01, -5.1033e-01, ..., -5.5779e-02,
1.1078e-01, 3.9052e-02],
...,
[ 2.5119e-01, -1.2241e-01, -4.9132e-02, ..., -2.7280e-01,
-2.3096e-02, -1.9215e-01],
[-5.2790e-01, 3.7251e-02, 2.1808e-01, ..., -3.8262e-01,
-3.1527e-02, 1.1366e-02],
[ 4.3142e-01, -4.2621e-02, -1.5395e-01, ..., -2.0733e-01,
-1.1442e-01, -1.8908e-02]],
[[-3.5645e-01, 6.6657e-02, 5.2657e-01, ..., 2.2563e-01,
5.6672e-01, -8.8984e-02],
[-3.9088e-01, 6.6069e-01, 9.8569e-02, ..., -2.7920e-01,
-2.3524e-01, 1.2398e-01],
[-4.8302e-01, -1.1691e-01, 3.0573e-01, ..., -3.4909e-01,
-1.9151e-01, -2.5360e-01],
...,
[-3.0950e-01, 2.9786e-01, 7.0305e-02, ..., -2.2046e-01,
-7.6774e-02, 1.5473e-01],
[ 5.3203e-01, -3.0058e-01, 1.5836e-03, ..., -7.5064e-02,
4.0212e-01, 2.4620e-01],
[-1.9144e-01, 9.2998e-05, -8.9839e-02, ..., -3.4417e-01,
2.4298e-01, -1.2038e-01]],
...,
[[ 1.2952e-01, -4.0685e-01, 8.5675e-01, ..., 6.6157e-01,
5.7936e-01, -4.3543e-01],
[-1.7371e-01, 1.0144e-01, 4.1771e-01, ..., 2.0275e-01,
-9.2321e-02, 6.3147e-04],
[-2.9344e-02, -4.4817e-01, 1.4441e-01, ..., -1.2976e-01,
-3.7015e-01, -5.9597e-02],
...,
[ 6.8723e-03, 4.2173e-01, -1.6490e-01, ..., -1.2973e-01,
8.1808e-02, -1.6639e-01],
[-3.8542e-01, 3.0564e-01, -5.3050e-03, ..., -3.9722e-01,
6.1607e-02, -2.4981e-02],
[ 1.0085e-01, -2.0305e-01, -1.6138e-01, ..., -7.7343e-02,
-1.3430e-01, 1.9485e-01]],
[[-7.9828e-02, -1.8743e-01, 3.2408e-01, ..., 1.7548e-01,
4.7418e-01, -1.9687e-01],
[-5.8841e-01, -1.3589e-01, -3.6579e-01, ..., -3.5333e-01,
2.7716e-02, 1.2175e-01],
[-7.5884e-02, 2.7846e-01, 8.2954e-02, ..., -4.0592e-01,
6.5394e-02, 3.8879e-01],
...,
[-9.1677e-02, 1.4918e-01, -2.3913e-01, ..., -1.6788e-01,
-8.4728e-02, -1.0471e-01],
[ 7.6186e-01, -4.4413e-01, -3.5728e-01, ..., -2.0602e-01,
-4.3217e-02, 2.8945e-01],
[-9.2108e-02, 8.9774e-02, -5.4875e-01, ..., -2.8431e-01,
-1.2403e-01, 7.7167e-02]],
[[-2.3886e-01, -1.3138e-01, 3.9558e-01, ..., 2.6036e-01,
3.7969e-01, -3.6619e-01],
[-1.2701e-01, 5.2801e-01, -4.5345e-02, ..., -2.8627e-01,
2.7628e-02, 1.7306e-01],
[-7.2136e-02, 2.5477e-02, 1.4034e-01, ..., -2.4256e-01,
-2.0102e-01, -1.4860e-01],
...,
[-5.3404e-02, 4.2754e-01, 1.8371e-02, ..., -1.2657e-01,
-2.5988e-01, -1.1474e-01],
[-2.9916e-01, 6.7820e-01, -2.9635e-01, ..., -1.6857e-01,
-1.9597e-02, 1.7084e-01],
[ 4.0832e-01, -3.5684e-02, 9.1102e-01, ..., -2.0259e-01,
3.1585e-01, -3.9386e-01]]]], device='cuda:0',
grad_fn=<AddBackward0>), tensor([[[[-0.5928, 0.6790, 1.3195, ..., 4.5488, -0.9982, 6.1315],
[-0.5782, 0.6907, 1.2045, ..., 4.5504, -1.2954, 6.0554],
[-0.6523, 0.5459, 1.0743, ..., 4.4112, -1.3595, 5.8574],
...,
[-0.5950, 0.5799, 1.2367, ..., 4.6545, -1.0932, 6.2976],
[-0.6645, 0.9539, 1.2341, ..., 4.5624, -1.2763, 6.1479],
[-0.3889, 0.7678, 1.2711, ..., 4.9012, -1.1082, 6.6677]],
[[-0.5805, 0.7389, 1.0743, ..., 4.5676, -0.7191, 6.0390],
[-0.7277, 0.7867, 0.9045, ..., 4.6617, -1.0535, 6.3922],
[-0.9212, 0.8381, 0.7461, ..., 4.5766, -0.9332, 5.9639],
...,
[-0.8907, 0.5791, 1.0640, ..., 4.7195, -0.9491, 6.2605],
[-0.8991, 0.7592, 0.9439, ..., 4.4987, -0.8906, 6.0442],
[-0.9392, 0.9288, 0.9961, ..., 4.7385, -0.8684, 6.5438]],
[[-0.7160, 0.6456, 1.1418, ..., 4.7414, -0.9239, 6.2186],
[-0.8872, 0.6170, 0.9740, ..., 4.7376, -1.0611, 6.2250],
[-0.9317, 0.4504, 1.0573, ..., 4.5252, -1.1890, 5.8526],
...,
[-0.7416, 0.6710, 1.1090, ..., 4.6707, -1.1684, 6.6294],
[-0.9892, 0.4752, 0.9795, ..., 4.4836, -0.9691, 5.8065],
[-0.8462, 0.5647, 1.0336, ..., 4.7973, -0.9707, 6.3381]],
...,
[[-0.4761, 0.7986, 1.0850, ..., 4.8306, -0.8325, 6.5305],
[-0.8435, 0.5465, 1.1957, ..., 4.5269, -1.2263, 6.1522],
[-0.7890, 0.4995, 1.1099, ..., 4.4734, -1.0913, 6.0248],
...,
[-0.8840, 0.6294, 0.9372, ..., 4.4714, -0.9068, 5.8156],
[-0.8075, 0.6659, 1.0219, ..., 4.6203, -1.0859, 6.0948],
[-0.6666, 0.5310, 0.9849, ..., 4.5458, -0.9914, 6.1484]],
[[-0.6323, 0.6200, 1.1372, ..., 4.6640, -0.8417, 6.1914],
[-0.5248, 0.8177, 1.1210, ..., 4.7104, -1.1385, 6.6570],
[-0.9032, 0.6442, 0.9184, ..., 4.8537, -1.0460, 6.2843],
...,
[-0.8778, 0.5821, 0.9644, ..., 4.5857, -0.9153, 6.1122],
[-0.8188, 0.5875, 1.0151, ..., 4.5335, -1.0260, 6.1869],
[-0.7261, 0.6305, 1.0700, ..., 4.7015, -1.0177, 6.2973]],
[[-0.4941, 0.6583, 1.2299, ..., 4.7833, -0.8020, 6.6197],
[-0.6274, 0.6860, 1.1028, ..., 4.7337, -0.9022, 6.5687],
[-0.7527, 0.7736, 0.9567, ..., 4.8420, -0.8560, 6.4241],
...,
[-0.6877, 0.8529, 1.0633, ..., 4.6611, -0.9536, 6.3576],
[-0.7523, 0.7166, 1.0473, ..., 4.7613, -0.9024, 6.3001],
[-0.6006, 0.7225, 1.0519, ..., 4.7283, -0.8990, 6.6008]]]],
device='cuda:0', grad_fn=<AddBackward0>), tensor([[[[0.0582, 0.0902, 0.0730, ..., 0.0925, 0.1473, 0.1316],
[0.0816, 0.0706, 0.0920, ..., 0.0771, 0.1516, 0.1330],
[0.0595, 0.0559, 0.0757, ..., 0.0693, 0.1427, 0.1357],
...,
[0.0497, 0.0875, 0.0602, ..., 0.1015, 0.1339, 0.1121],
[0.0569, 0.0770, 0.0789, ..., 0.0840, 0.1249, 0.1358],
[0.0535, 0.0876, 0.0515, ..., 0.1120, 0.1249, 0.1072]],
[[0.0351, 0.0672, 0.0434, ..., 0.1184, 0.1499, 0.1115],
[0.0285, 0.0568, 0.0371, ..., 0.1190, 0.1599, 0.1127],
[0.0243, 0.0515, 0.0363, ..., 0.1130, 0.1350, 0.1191],
...,
[0.0198, 0.0526, 0.0457, ..., 0.1188, 0.1338, 0.1098],
[0.0239, 0.0560, 0.0456, ..., 0.0991, 0.1318, 0.1188],
[0.0281, 0.0650, 0.0517, ..., 0.0972, 0.1355, 0.1237]],
[[0.0330, 0.0589, 0.0356, ..., 0.1244, 0.1215, 0.1079],
[0.0240, 0.0479, 0.0377, ..., 0.1250, 0.0981, 0.1093],
[0.0191, 0.0615, 0.0468, ..., 0.1433, 0.1210, 0.1016],
...,
[0.0196, 0.0737, 0.0327, ..., 0.1409, 0.0842, 0.1022],
[0.0155, 0.0561, 0.0380, ..., 0.1514, 0.0934, 0.1040],
[0.0168, 0.0532, 0.0353, ..., 0.1395, 0.0935, 0.0934]],
...,
[[0.0593, 0.0598, 0.0369, ..., 0.1055, 0.1213, 0.1170],
[0.0153, 0.0501, 0.0477, ..., 0.1125, 0.1358, 0.1177],
[0.0222, 0.0660, 0.0312, ..., 0.1348, 0.1213, 0.0964],
...,
[0.0329, 0.0486, 0.0364, ..., 0.1475, 0.0951, 0.0975],
[0.0168, 0.0510, 0.0300, ..., 0.1319, 0.0962, 0.1054],
[0.0138, 0.0772, 0.0332, ..., 0.1113, 0.1095, 0.1210]],
[[0.0392, 0.0582, 0.0335, ..., 0.1467, 0.1250, 0.1012],
[0.0174, 0.0611, 0.0312, ..., 0.1188, 0.0983, 0.1089],
[0.0303, 0.0415, 0.0314, ..., 0.1314, 0.0940, 0.1021],
...,
[0.0328, 0.0422, 0.0473, ..., 0.1295, 0.1302, 0.1033],
[0.0177, 0.0501, 0.0304, ..., 0.1322, 0.1225, 0.1058],
[0.0124, 0.0667, 0.0227, ..., 0.1218, 0.1098, 0.1161]],
[[0.0259, 0.0795, 0.0280, ..., 0.1336, 0.1134, 0.1033],
[0.0160, 0.0682, 0.0237, ..., 0.1181, 0.1016, 0.1130],
[0.0171, 0.0654, 0.0226, ..., 0.1153, 0.0839, 0.1107],
...,
[0.0207, 0.0680, 0.0375, ..., 0.1207, 0.1208, 0.1201],
[0.0160, 0.0641, 0.0284, ..., 0.1263, 0.1019, 0.1122],
[0.0157, 0.0729, 0.0223, ..., 0.1279, 0.0907, 0.1017]]]],
device='cuda:0', grad_fn=<ReshapeAliasBackward0>), 3, 3, 1, 1, 1, 1, 1, 1, 4, 16, 1.0, 256
I confirmed that the version of b64d9ca3000f4753e26d3b8c8fec05052114a126, latest commit before 8f2d158399b95b2188b86dd46da93c25c2dbb7cc is possible to export a ckpt model to an ONNX model.
change remove_center=int(remove_center) to remove_center_i=int(remove_center) in dcnv3_func.py can solve it
change remove_center=int(remove_center) to remove_center_i=int(remove_center) in dcnv3_func.py can solve it
The same problem occurs even if I fix it. Do you happen to know the solution?
