Hi there
Cannot use the following code
from torch.autograd import Variable
import torch.onnx
import torch
dummy_input = Variable(torch.randn(1, 3, 256, 256))
model = torch.load(‘Paramecium.pth’, map_location=torch.device(‘cpu’))
torch.onnx.export(model, dummy_input, “fromTorch.onnx”)
to convert the pth to onnx. It throws the ‘collections.OrderedDict’ object has no attribute ‘state_dict’ at the last line execution.
The way I create the .pth file is following (Colab):
from detecto import core, utils, visualize
dataset = core.Dataset(’/content/drive/My Drive/Paramecium’)
model = core.Model([‘Paramecium’, ‘Noise’])
model.fit(dataset)
model.save(‘Paramecium.pth’)
The model looks good, works and can be visualized with visualize.show_labeled_image
How come it throws the error when converting to ONNX?
Thanks.
The model is actually the weights.!
You need to load the weights inside an existing network.
I don’t see any code where u create a network and then load the weights into it.
For example, you need to something like this.!
weights = torch.load("Paramecium.pth", map_location = torch.device(‘cpu’))
model = core.Model([‘Paramecium’, ‘Noise’])
model.load_state_dict(weights)
torch.onnx.export(model, dummy_input, "*.onnx")
You are actually missing the model.load_state_dict()
The Paramecium.pth just contains the weights, so you need to load these weights inside a model core.Model([‘Paramecium’, ‘Noise’])
Hm… What is core? core.Model - where it comes from?
My imports are as following:
from torch.autograd import Variable
import torch.onnx
import torchvision
import torch
My Visual Studio Python environment is shown in the picture.
Thoughts?
Thanks.
Ooops… I found it, it is just a part of detecto.
Anyway, the line
model.load_state_dict(weights)
throws the following exception
Thoughts?
Thanks.
Try
model= core.Model.load("*pth", ["Paramecium", "Noise"] ) instead of those two lines involving model
Hey, use Paramecium.pth instead of *pth
Tried it already. It brings me back to the initial problem:
It looks like I’m getting more and more involved in voodoo programming.
Let me try to get both tasks (model training and ONNX conversion) done in one place - Colab. So far once the model trained in Colab, I switch to my local environment to Visual Studio finishing the conversion.
Let’s see.
No. The Colab code below
is not working either throwing
AttributeError: ‘Model’ object has no attribute ‘state_dict’
I’m stuck. Help!
I’m not sure what is exactly wrong, but you can refer to this link and see what’s the mistake
https://detecto.readthedocs.io/en/latest/api/core.html
Hi, after
model = core.Model.load("Paramecium.pth", ["Paramecium", "Noise"])
could you try this line,
torch.onnx.export(model.get_internal_model(), dummy_input, "from_onnx.onnx") ?
It moved somewhere ahead but throws the exception after:
RuntimeError: Failed to export an ONNX attribute ‘onnx::Sub’, since it’s not constant, please try to make things (e.g., kernel size) static if possible
Stack:
Thoughts?
One more peace of information from the Colab debugger:
/usr/local/lib/python3.6/dist-packages/torch/onnx/symbolic_helper.py:243: UserWarning: You are trying to export the model with onnx:Upsample for ONNX opset version 9.
This operator might cause results to not match the expected results by PyTorch.
ONNX’s Upsample/Resize operator did not match Pytorch’s Interpolation until opset 11.
Attributes to determine how to transform the input were added in onnx:Resize in opset 11
to support Pytorch’s behavior (like coordinate_transformation_mode and nearest_mode).
We recommend using opset 11 and above for models using this operator.
“” + str(_export_onnx_opset_version) + ". "
How can I get the opset 11+ ?
Thanks.
It looks like I fixed it:
Thanks.
