pytorch->onnx->paddle出错

[smallsung1999]

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. :smile_cat:

问题描述

• 错误信息 AssertionError: Variable dtype not match, Variable [ conv2d_0.w_0 ] need tensor with dtype paddle.float32 but load tensor with dtype paddle.float16
• 错误截图

具体信息

pytorch模型转换至onnx成功，但onnx转换为paddle时，生成x2paddle_code.py后转换失败。

下面是对错误信息的推测： 模型conv前对数据进行了cast操作，将float32数据转换为float16再输入conv，而conv中的w与b均为float16，如何在x2paddle_code.py定义模型结构时指定conv的dtype为float16？ 如有更好的解决办法，欢迎大佬指正。

• 转换模型后用处
  • [ ] 使用 Paddle 框架/ PaddleInference 推理预测
  • [√ ] 使用 Paddle-Lite 做移动端推理
  • [ ] 转换预训练参数，再使用 Paddle 进行模型开发
• 模型来源 AST tiny224：https://github.com/YuanGongND/ast
• 应用场景 用于移动端业务下的音频业务
• 版本信息 PaddlePaddle => :2.4.1: X2Paddle => :1.4.0: 来源框架版本(PyTorch/TF/ONNX/Caffe) => :1.13.1:
• 您的联系方式(邮箱/微信/电话):songzexu@apon.com.cn

[smallsung1999]

model.zip 模型相关文件

[smallsung1999]

torch.onnx.export导出模型时设置do_constant_folding=False，关闭amp，w和b保存为float32

[smallsung1999]

In transformed code:

File "H:\ast\pd_model\x2paddle_code.py", line 690, in forward
    x2paddle__blocks_0_attn_Gather_1_output_0 = paddle.gather(x=x2paddle__blocks_0_attn_Transpose_output_0, index=x2paddle__blocks_0_attn_Constant_5_output_0, axis=0)
    x2paddle__blocks_0_attn_Gather_2_output_0 = paddle.gather(x=x2paddle__blocks_0_attn_Transpose_output_0, index=x2paddle__blocks_0_attn_Constant_6_output_0, axis=0)
    x2paddle__blocks_0_attn_Transpose_1_output_0 = paddle.transpose(x=x2paddle__blocks_0_attn_Gather_1_output_0, perm=[0, 1, 3, 2])   
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
    x2paddle__blocks_0_attn_MatMul_output_0 = paddle.matmul(x=x2paddle__blocks_0_attn_Gather_output_0, y=x2paddle__blocks_0_attn_Transpose_1_output_0)
    x2paddle__blocks_0_attn_Mul_output_0 = paddle.multiply(x=x2paddle__blocks_0_attn_MatMul_output_0, y=x2paddle__blocks_0_attn_Constant_7_output_0)

File "C:\ProgramData\Anaconda3\envs\pytorch\lib\site-packages\paddle\tensor\linalg.py", line 123, in transpose
    raise ValueError(

ValueError: Input(perm) is the permutation of dimensions of Input(x), its length should be equal to dimensions of Input(x), but received dimension of Input(x) is 5, the length of Input(perm) is 4.

model.zip x2paddle_code.zip