PyTorch转换为Paddle; Paddle不支持 convert failed node:_features_conv0_Conv_output_0, op_type is Conv

[zuishenke123]

感谢您参与 X2Paddle 社区！ 问题模版为了 X2Paddle 能更好的迭代，例如新功能发布、 RoadMaps 和错误跟踪. 😸

问题描述

将PyTorch官方库中的DenseNet模型转换为Paddle版本，并用于测试基于ONNX的模型转换性能。PyTorch模型转换ONNX成功，但是ONNX转换为PaddlePaddle失败

• 错误信息
  -Converting node 434 ... Traceback (most recent call last):
  File "D:\anaconda3\envs\LID-CMC\lib\site-packages\x2paddle\op_mapper\onnx2paddle\opset_legacy.py", line 110, in run_mapping
  res = func(*args, **kwargs)
  File "D:\anaconda3\envs\LID-CMC\lib\site-packages\x2paddle\op_mapper\onnx2paddle\opset_legacy.py", line 2304, in Conv
  _rename_or_remove_weight(
  File "D:\anaconda3\envs\LID-CMC\lib\site-packages\x2paddle\op_mapper\onnx2paddle\opset_legacy.py", line 68, in _rename_or_remove_weight
  raise KeyError('{} not a key in {}'.format(origin_name, weights.keys()))
  KeyError: "x2paddle_onnx__Conv_1167 not a key in dict_keys(['x2paddle_features_denseblock1_denselayer1_norm1_weight', 'x2paddle_features_denseblock1_denselayer1_norm1_bias', 'x2paddle_features_denseblock1_denselayer1_conv2_weight', 'x2paddle_features_denseblock1_denselayer2_norm1_weight', 'x2paddle_features_denseblock1_denselayer2_norm1_bias', 'x2paddle_features_denseblock1_denselayer2_conv2_weight', 'x2paddle_features_denseblock1_denselayer3_norm1_weight', 'x2paddle_features_denseblock1_denselayer3_norm1_bias', 'x2paddle_features_denseblock1_denselayer3_conv2_weight', 'x2paddle_features_denseblock1_denselayer4_norm1_weight', 'x2paddle_features_denseblock1_denselayer4_norm1_bias', 'x2paddle_features_denseblock1_denselayer4_conv2_weight', 'x2paddle_features_denseblock1_denselayer5_norm1_weight', 'x2paddle_features_denseblock1_denselayer5_norm1_bias', 'x2paddle_features_denseblock1_denselayer5_conv2_weight', 'x2paddle_features_denseblock1_denselayer6_norm1_weight', 'x2paddle_features_denseblock1_denselayer6_norm1_bias', 'x2paddle_features_denseblock1_denselayer6_conv2_weight', 'x2paddle_features_transition1_norm_weight', 'x2paddle_features_transition1_norm_bias', 'x2paddle_features_transition1_conv_weight', 'x2paddle_features_denseblock2_denselayer1_conv2_weight', 'x2paddle_features_denseblock2_denselayer2_conv2_weight', 'x2paddle_features_denseblock2_denselayer3_conv2_weight', 'x2paddle_features_denseblock2_denselayer4_conv2_weight', 'x2paddle_features_denseblock2_denselayer5_conv2_weight', 'x2paddle_features_denseblock2_denselayer6_norm1_weight', 'x2paddle_features_denseblock2_denselayer6_norm1_bias', 'x2paddle_features_denseblock2_denselayer6_conv2_weight', 'x2paddle_features_denseblock2_denselayer7_norm1_weight', 'x2paddle_features_denseblock2_denselayer7_norm1_bias', 'x2paddle_features_denseblock2_denselayer7_conv2_weight', 'x2paddle_features_denseblock2_denselayer8_norm1_weight', 'x2paddle_features_denseblock2_denselayer8_norm1_bias', 'x2paddle_features_denseblock2_denselayer8_conv2_weight', 'x2paddle_features_denseblock2_denselayer9_norm1_weight', 'x2paddle_features_denseblock2_denselayer9_norm1_bias', 'x2paddle_features_denseblock2_denselayer9_conv2_weight', 'x2paddle_features_denseblock2_denselayer10_norm1_weight', 'x2paddle_features_denseblock2_denselayer10_norm1_bias', 'x2paddle_features_denseblock2_denselayer10_conv2_weight', 'x2paddle_features_denseblock2_denselayer11_norm1_weight', 'x2paddle_features_denseblock2_denselayer11_norm1_bias', 'x2paddle_features_denseblock2_denselayer11_conv2_weight', 'x2paddle_features_denseblock2_denselayer12_norm1_weight', 'x2paddle_features_denseblock2_denselayer12_norm1_bias', 'x2paddle_features_denseblock2_denselayer12_conv2_weight', 'x2paddle_features_transition2_norm_weight', 'x2paddle_features_transition2_norm_bias', 'x2paddle_features_transition2_conv_weight', 'x2paddle_features_denseblock3_denselayer1_conv2_weight', 'x2paddle_features_denseblock3_denselayer2_conv2_weight', 'x2paddle_features_denseblock3_denselayer3_conv2_weight', 'x2paddle_features_denseblock3_denselayer4_conv2_weight', 'x2paddle_features_denseblock3_denselayer5_conv2_weight', 'x2paddle_features_denseblock3_denselayer6_conv2_weight', 'x2paddle_features_denseblock3_denselayer7_conv2_weight', 'x2paddle_features_denseblock3_denselayer8_conv2_weight', 'x2paddle_features_denseblock3_denselayer9_conv2_weight', 'x2paddle_features_denseblock3_denselayer10_norm1_weight', 'x2paddle_features_denseblock3_denselayer10_norm1_bias', 'x2paddle_features_denseblock3_denselayer10_conv2_weight', 'x2paddle_features_denseblock3_denselayer11_norm1_weight', 'x2paddle_features_denseblock3_denselayer11_norm1_bias', 'x2paddle_features_denseblock3_denselayer11_conv2_weight', 'x2paddle_features_denseblock3_denselayer12_norm1_weight', 'x2paddle_features_denseblock3_denselayer12_norm1_bias', 'x2paddle_features_denseblock3_denselayer12_conv2_weight', 'x2paddle_features_denseblock3_denselayer13_norm1_weight', 'x2paddle_features_denseblock3_denselayer13_norm1_bias', 'x2paddle_features_denseblock3_denselayer13_conv2_weight', 'x2paddle_features_denseblock3_denselayer14_norm1_weight', 'x2paddle_features_denseblock3_denselayer14_norm1_bias', 'x2paddle_features_denseblock3_denselayer14_conv2_weight', 'x2paddle_features_denseblock3_denselayer15_norm1_weight', 'x2paddle_features_denseblock3_denselayer15_norm1_bias', 'x2paddle_features_denseblock3_denselayer15_conv2_weight', 'x2paddle_features_denseblock3_denselayer16_norm1_weight', 'x2paddle_features_denseblock3_denselayer16_norm1_bias', 'x2paddle_features_denseblock3_denselayer16_conv2_weight', 'x2paddle_features_denseblock3_denselayer17_norm1_weight', 'x2paddle_features_denseblock3_denselayer17_norm1_bias', 'x2paddle_features_denseblock3_denselayer17_conv2_weight', 'x2paddle_features_denseblock3_denselayer18_norm1_weight', 'x2paddle_features_denseblock3_denselayer18_norm1_bias', 'x2paddle_features_denseblock3_denselayer18_conv2_weight', 'x2paddle_features_denseblock3_denselayer19_norm1_weight', 'x2paddle_features_denseblock3_denselayer19_norm1_bias', 'x2paddle_features_denseblock3_denselayer19_conv2_weight', 'x2paddle_features_denseblock3_denselayer20_norm1_weight', 'x2paddle_features_denseblock3_denselayer20_norm1_bias', 'x2paddle_features_denseblock3_denselayer20_conv2_weight', 'x2paddle_features_denseblock3_denselayer21_norm1_weight', 'x2paddle_features_denseblock3_denselayer21_norm1_bias', 'x2paddle_features_denseblock3_denselayer21_conv2_weight', 'x2paddle_features_denseblock3_denselayer22_norm1_weight', 'x2paddle_features_denseblock3_denselayer22_norm1_bias', 'x2paddle_features_denseblock3_denselayer22_conv2_weight', 'x2paddle_features_denseblock3_denselayer23_norm1_weight', 'x2paddle_features_denseblock3_denselayer23_norm1_bias', 'x2paddle_features_denseblock3_denselayer23_conv2_weight', 'x2paddle_features_denseblock3_denselayer24_norm1_weight', 'x2paddle_features_denseblock3_denselayer24_norm1_bias', 'x2paddle_features_denseblock3_denselayer24_conv2_weight', 'x2paddle_features_transition3_norm_weight', 'x2paddle_features_transition3_norm_bias', 'x2paddle_features_transition3_conv_weight', 'x2paddle_features_denseblock4_denselayer1_conv2_weight', 'x2paddle_features_denseblock4_denselayer2_conv2_weight', 'x2paddle_features_denseblock4_denselayer3_conv2_weight', 'x2paddle_features_denseblock4_denselayer4_conv2_weight', 'x2paddle_features_denseblock4_denselayer5_conv2_weight', 'x2paddle_features_denseblock4_denselayer6_conv2_weight', 'x2paddle_features_denseblock4_denselayer7_conv2_weight', 'x2paddle_features_denseblock4_denselayer8_conv2_weight', 'x2paddle_features_denseblock4_denselayer9_conv2_weight', 'x2paddle_features_denseblock4_denselayer10_conv2_weight', 'x2paddle_features_denseblock4_denselayer11_conv2_weight', 'x2paddle_features_denseblock4_denselayer12_conv2_weight', 'x2paddle_features_denseblock4_denselayer13_conv2_weight', 'x2paddle_features_denseblock4_denselayer14_conv2_weight', 'x2paddle_features_denseblock4_denselayer15_conv2_weight', 'x2paddle_features_denseblock4_denselayer16_conv2_weight', 'x2paddle_classifier_weight', 'x2paddle_classifier_bias', 'x2paddle_onnx__Conv_1166', 'x2paddle_onnx__Conv_1169', 'x2paddle_onnx__Conv_1172', 'x2paddle_onnx__Conv_1175', 'x2paddle_onnx__Conv_1178', 'x2paddle_onnx__Conv_1181', 'x2paddle_onnx__Conv_1184', 'x2paddle_onnx__Conv_1187', 'x2paddle_onnx__Conv_1190', 'x2paddle_onnx__Conv_1193', 'x2paddle_onnx__Conv_1196', 'x2paddle_onnx__Conv_1199', 'x2paddle_onnx__Conv_1202', 'x2paddle_onnx__Conv_1205', 'x2paddle_onnx__Conv_1208', 'x2paddle_onnx__Conv_1211', 'x2paddle_onnx__Conv_1214', 'x2paddle_onnx__Conv_1217', 'x2paddle_onnx__Conv_1220', 'x2paddle_onnx__Conv_1223', 'x2paddle_onnx__Conv_1226', 'x2paddle_onnx__Conv_1229', 'x2paddle_onnx__Conv_1232', 'x2paddle_onnx__Conv_1235', 'x2paddle_onnx__Conv_1238', 'x2paddle_onnx__Conv_1241', 'x2paddle_onnx__Conv_1244', 'x2paddle_onnx__Conv_1247', 'x2paddle_onnx__Conv_1250', 'x2paddle_onnx__Conv_1253', 'x2paddle_onnx__Conv_1256', 'x2paddle_onnx__Conv_1259', 'x2paddle_onnx__Conv_1262', 'x2paddle_onnx__Conv_1265', 'x2paddle_onnx__Conv_1268', 'x2paddle_onnx__Conv_1271', 'x2paddle_onnx__Conv_1274', 'x2paddle_onnx__Conv_1277', 'x2paddle_onnx__Conv_1280', 'x2paddle_onnx__Conv_1283', 'x2paddle_onnx__Conv_1286', 'x2paddle_onnx__Conv_1289', 'x2paddle_onnx__Conv_1292', 'x2paddle_onnx__Conv_1295', 'x2paddle_onnx__Conv_1298', 'x2paddle_onnx__Conv_1301', 'x2paddle_onnx__Conv_1304', 'x2paddle_onnx__Conv_1307', 'x2paddle_onnx__Conv_1310', 'x2paddle_onnx__Conv_1313', 'x2paddle_onnx__Conv_1316', 'x2paddle_onnx__Conv_1319', 'x2paddle_onnx__Conv_1322', 'x2paddle_onnx__Conv_1325', 'x2paddle_onnx__Conv_1328', 'x2paddle_onnx__Conv_1331', 'x2paddle_onnx__Conv_1334', 'x2paddle_onnx__Conv_1337', 'x2paddle_onnx__Conv_1340', 'x2paddle__features_transition1_pool_Constant_output_0', 'x2paddle__features_transition2_pool_Constant_output_0', 'x2paddle__features_transition3_pool_Constant_output_0', 'conv0.weight'])"

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "D:\anaconda3\envs\LID-CMC\lib\runpy.py", line 197, in _run_module_as_main
return _run_code(code, main_globals, None,
File "D:\anaconda3\envs\LID-CMC\lib\runpy.py", line 87, in run_code
exec(code, run_globals)
File "D:\anaconda3\envs\LID-CMC\Scripts\x2paddle.exe_main.py", line 7, in
sys.exit(main())
File "D:\anaconda3\envs\LID-CMC\lib\site-packages\x2paddle\convert.py", line 489, in main
onnx2paddle(
File "D:\anaconda3\envs\LID-CMC\lib\site-packages\x2paddle\convert.py", line 304, in onnx2paddle
mapper = ONNXOpMapper(model)
File "D:\anaconda3\envs\LID-CMC\lib\site-packages\x2paddle\op_mapper\onnx2paddle\onnx_op_mapper.py", line 52, in init
func(node)
File "D:\anaconda3\envs\LID-CMC\lib\site-packages\x2paddle\op_mapper\onnx2paddle\opset_legacy.py", line 112, in run_mapping
raise Exception("convert failed node:{}, op_type is {}".format(
Exception: convert failed node:_features_conv0_Conv_output_0, op_type is Conv

• 错误截图
  

具体信息

详细代码

import torch,os
from torchvision.models.densenet import DenseNet121_Weights, DenseNet

# Initialize Result Path
input_tensor = torch.randn(64, 3, 224, 224)
output_folder = os.path.join(os.getcwd())
if not os.path.exists(output_folder):
    os.makedirs(output_folder)

weights = DenseNet121_Weights.verify(None)
model = DenseNet(32, (6, 12, 24, 16), num_classes=100)
model_name = "densenet121"
if weights is not None:
     model.load_state_dict(weights.get_state_dict(progress=True))
model.eval()

# 指定存储文件夹的名称
if not os.path.exists(os.path.join(output_folder, f"{model_name}_IR_and_pd")):
    os.makedirs(os.path.join(output_folder, f"{model_name}_IR_and_pd"))

try:
    import torch
except Exception as e:
    raise ValueError(f"import torch failed:{e}")

onnx_file = os.path.join(output_folder, f"{model_name}_IR_and_pd", f"testTorch{model_name}.onnx")
try:
   torch.onnx.export(model, input_tensor, onnx_file, verbose=True)
except Exception as e:
   raise ValueError(f"Failed to export ONNX model: {e}")

pd_model_dir = os.path.join(output_folder, f"{model_name}_IR_and_pd")

# 使用 x2paddle 转换 ONNX 模型到 PaddlePaddle 格式
try:
   conversion_command = f"x2paddle --framework=onnx --model={onnx_file} --save_dir={pd_model_dir}"
   conversion_result = os.system(conversion_command)
   print(f"x2paddle conversion command returned: {conversion_result}")
except Exception as e:
   raise ValueError(f"Failed to convert ONNX model with x2paddle: {e}")

print(f"Model conversion completed. ONNX and PaddlePaddle models are saved in '{output_folder}'.")

• 转换模型后用处
  使用 Paddle 框架/ PaddleInference 推理预测

• 模型来源
  densenet121：https://github.com/pytorch/vision/blob/main/torchvision/models/densenet.py

• 应用场景
  用于测试模型转换任务

• 版本信息
  X2Paddle => 1.4.1
  torchvision => 0.15.2
  来源框架版本(PyTorch/TF/ONNX/Caffe):
  onxx => 1.15.0
  Pytorch => 2.0.1
  paddlepaddle-gpu => 2.6.0

• 您的联系方式(邮箱/微信/电话)
  [email protected]