Export Detectron2 model before calling model.eval()

vision
1

I trained a detectron2 model on some data that I plan to use, but unfortunately the team I’m working with is requiring that I use another software that they have already implemented (its a long story). Anyways now I’m having to try to convert this model into a software known as MVtec Halcon. The issue is that it gives the following error when trying to upload the onnx model of the detectron2 that I trained:

Unhandled program exception:

HALCON operator error

while calling ‘read_dl_model’ in procedure ‘main’ line: 4.

Error occurred while reading an ONNX model (HALCON error code: 7801)

Extended error code: 5:

Shape dimension for graph input must be 4

What I have come to understand from this issue is that it requires me to convert the detectron2 model into onnx before calling the model.eval() method which removes the batch_size from the input requirements for the detectron2 model.

I imagine this isn’t common, but I was curious to if anyone has had any experience with this and could provide any tips. I can’t seem to figure it out and I can’t seem to find anything online relating to it.

ADDITIONAL INFO: The batch_size that I trained on is 1 (the images are massive), so maybe there is a way I can export the model in the typical way and then just unsqueeze the input? Open to any ideas.

2

So I have been continuing to try to figure this out and arrived at a potential idea. I was able to employ the following code:

import onnx
from onnx import helper, TensorProto

Load your ONNX model

model = onnx.load(model_path_v1)

Define the input tensor with dynamic batch size

input_tensor = helper.make_tensor_value_info(‘input’, TensorProto.FLOAT, [1, 3, 1333, 348])

Update the input of the model

model.graph.input[0].CopyFrom(input_tensor)

Iterate through all nodes and update dimensions if necessary

for node in model.graph.node:
for i, input_name in enumerate(node.input):
if input_name == ‘x.1’:
node.input[i] = ‘input’

Save the modified model

onnx.save(model, ‘modified_model.onnx’)

It increases the input to contain a fixed batch_size; However, this just gives me more issues (which was kind of expected). Here is the netron of my model, hopefully that can help someone who is trying to understand my issue.

image
image735×923 24.9 KB