dnn: inputs.size() in function 'getMemoryShapes' #16660
Description
- OpenCV => 4.1.2 / 4.2.0
- Operating System / Platform => google colab
- Compiler => python3
trying to export a pytorch unet model via onnx.
when importing it into cv2.dnn i met:
error: OpenCV(4.1.2) /io/opencv/modules/dnn/src/dnn.cpp:3873: error: (-215:Assertion failed) inputs.size() in function 'getMemoryShapes'
problem seems to come from this line:
code to reproduce
import torch
import torch.nn as nn
import cv2
def get_incoming_shape(incoming):
size = incoming.size()
# returns the incoming data shape as a list
return [size[0], size[1], size[2], size[3]]
def interleave(tensors, axis):
old_shape = get_incoming_shape(tensors[0])
# change the first element (batch_size to 1)
new_shape = [1] + old_shape[1:]
# double 1 dimension
new_shape[axis] *= len(tensors) #### PROBLEM HERE
# pack the tensors on top of each other
stacked = torch.stack(tensors, axis+1)
# reshape and return
reshaped = stacked.reshape(new_shape)
return reshaped
class Problem(nn.Module):
def __init__(self):
super(Problem, self).__init__()
def forward(self,x):
return interleave([x,x],2)
def convert_to_onnx(net, output_name):
input = torch.randn(1, 3, 256, 320)
input_names = ['data']
output_names = ['output']
net.eval()
torch.onnx.export(net, input, output_name, verbose=True, input_names=input_names, output_names=output_names)
model = Problem()
convert_to_onnx(model, "problem.onnx")
net = cv2.dnn.readNet("problem.onnx")
\```
</p>
</details>
<p>.
<details><summary>onnx output</summary>
<p>
graph(%data : Float(1, 3, 256, 320)):
%1 : Tensor = onnx::Shape(%data)
%2 : Tensor = onnx::Constantvalue={1}
%3 : Long() = onnx::Gather[axis=0](%1, %2) # :6:0
%4 : Tensor = onnx::Shape(%data)
%5 : Tensor = onnx::Constantvalue={2}
%6 : Long() = onnx::Gather[axis=0](%4, %5) # :6:0
%7 : Tensor = onnx::Shape(%data)
%8 : Tensor = onnx::Constantvalue={3}
%9 : Long() = onnx::Gather[axis=0](%7, %8) # :6:0
%10 : Long() = onnx::Constantvalue={2}
%11 : Long() = onnx::Mul(%6, %10)
%12 : Tensor = onnx::Unsqueezeaxes=[3]
%13 : Tensor = onnx::Unsqueezeaxes=[3]
%14 : Float(1, 3, 256, 2, 320) = onnx::Concat[axis=3](%12, %13) # :18:0
%15 : Long() = onnx::Constantvalue={1}
%16 : Tensor = onnx::Unsqueezeaxes=[0]
%17 : Tensor = onnx::Unsqueezeaxes=[0]
%18 : Tensor = onnx::Unsqueezeaxes=[0]
%19 : Tensor = onnx::Unsqueezeaxes=[0]
%20 : Tensor = onnx::Concat[axis=0](%16, %17, %18, %19)
%output : Float(1, 3, 512, 320) = onnx::Reshape(%14, %20) # :20:0
return (%output)
```
%11 : Long() = onnx::Mul(%6, %10)
is the one with missing inputs here. are "plain maths ops" in a pytorch graph a problem ?
you can check the exported model here: problem.onnx
p.s. i was able to solve it, by letting pytoch infer the reshape param like
new_shape = [1, old_shape[1], -1, old_shape[3]]
y = x.reshape(new_shape)
thus avoiding the multiplication