**模型部署系列**
# 支持动态输入的超分辨率模型
onnx/2-srcnn-dynamic-onnxRuntime.py
https://github.com/open-mmlab/mmdeploy/blob/master/docs/zh_cn/tutorial/02_challenges.md
## 1, onnx 导出时报错
~~~python
with torch.no_grad():
torch.onnx.export(
model,
(x,3),
onnx_path,
opset_version=11, # ONNX 算子集的版本
input_names=['input', 'factor'], # 是输入、输出 tensor 的名称
output_names=['output'])
~~~
报错如下：
~~~
Traceback (most recent call last):
File "c:/Users/dufei/codes/ai/ai-lab/deploy/onnx/1-srcnn-onnxRuntime copy.py", line 134, in <module>
output_names=['output']) <--------------------------------
File "C:\Users\dufei\.conda\envs\pt10\lib\site-packages\torch\onnx\__init__.py", line 320, in export
..........................
File "C:\Users\dufei\.conda\envs\pt10\lib\site-packages\torch\jit\_trace.py", line 132, in forward
self._force_outplace,
File "C:\Users\dufei\.conda\envs\pt10\lib\site-packages\torch\jit\_trace.py", line 118, in wrapper
File "c:/Users/dufei/codes/ai/ai-lab/deploy/onnx/1-srcnn-onnxRuntime copy.py", line 53, in forward
align_corners=False) <--------------------------------
File "C:\Users\dufei\.conda\envs\pt10\lib\site-packages\torch\nn\functional.py", line 3737, in interpolate
return torch._C._nn.upsample_bicubic2d(input, output_size, align_corners, scale_factors)TypeError: upsample_bicubic2d() received an invalid combination of arguments - got (Tensor, NoneType, bool, list), but expected one of:
* (Tensor input, tuple of ints output_size, bool align_corners, tuple of floats scale_factors)
didn't match because some of the arguments have invalid types: (Tensor, NoneType, bool, list)
* (Tensor input, tuple of ints output_size, bool align_corners, float scales_h, float scales_w, *, Tensor out)
~~~
原因：
刚刚的报错是因为 PyTorch 模型在导出到 ONNX 模型时，模型的输入参数的类型必须全部是 torch.Tensor。而实际上我们传入的第二个参数" 3 "是一个整形变量。这不符合 PyTorch 转 ONNX 的规定。我们必须要修改一下原来的模型的输入。为了保证输入的所有参数都是 torch.Tensor 类型的，我们做如下修改：
## 2, TraceWarning 的警告
~~~
c:/Users/dufei/codes/ai/ai-lab/deploy/onnx/1-srcnn-onnxRuntime copy.py:53: TracerWarning: Converting a tensor to a Python number might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
scale_factor=upscale_factor.item(),
~~~
查看导出的onnx ，是有问题的(没有缩放因子输入)：
解决方法：自定义算子
我们得自己定义一个实现插值的 PyTorch 算子，然后让它映射到一个我们期望的 ONNX Resize 算子上。
~~~python
class NewInterpolate(torch.autograd.Function):
@staticmethod
def symbolic(g, input, scales):
return g.op("Resize",
input,
g.op("Constant",
value_t=torch.tensor([], dtype=torch.float32)),
scales,
coordinate_transformation_mode_s="pytorch_half_pixel",
cubic_coeff_a_f=-0.75,
mode_s='cubic',
nearest_mode_s="floor")
@staticmethod
def forward(ctx, input, scales):
scales = scales.tolist()[-2:]
return interpolate(input,
scale_factor=scales,
mode='bicubic',
align_corners=False)
~~~
经过自定义算子，导出的onnx 模型效果如下：
参考：
- [第二章：解决模型部署中的难题](https://github.com/open-mmlab/mmdeploy/blob/master/docs/zh_cn/tutorial/02_challenges.md)
-
深度学习研习，包括训练，部署等.zip （16个子文件） 
lern_2 pic 
a1.png 190KB basics 
demo_train.py 4KB AdamOptimizer.ipynb 1010KB llm Untitled1.ipynb 5KB test.py 1KB deploy torch-script demo.py 2KB onnx 5-.py 960B my_add.cpp 191B 3-export_onnx.py 4KB setup.py 236B 4-define_ops.py 6KB 1-srcnn-onnxRuntime.py 6KB 2-srcnn-dynamic-onnxRuntime.py 7KB README.md 4KB .gitignore 2KB README.md 2KB资源评论
