284_CREStereo onnx model generation query
ashank-art opened this issue · comments
Issue Type
Support, Documentation Feature Request
OS
Ubuntu
OS architecture
x86_64
Programming Language
Python
Framework
PyTorch, ONNX
Model name and Weights/Checkpoints URL
https://github.com/PINTO0309/PINTO_model_zoo/blob/main/284_CREStereo/download_iter10.sh
Description
Hi @PINTO0309 Thanks for the provided models in onnx.
I have one query with the model conversion from PyTorch to onnx.
For Crestereo model conversion to onnx,
We need to convert the model twice, once without the init_flow and once with the init_flow. So the commented out part in the (https://github.com/ibaiGorordo/CREStereo-Pytorch/blob/main/convert_to_onnx.py) are not options, but two separate runs.
Then use the merging tool PINTO0309 developed to merge the two parts into one. Or, just use them separately in sequence.
For generating the merged model, we need to combine init and next models, with respect to the resolution for init and next models
Assume for generating the combined model of resolution 480 * 640, we need to merge init model of resolution 480 * 640 and next model with resolution 480 * 640, but what is the input resolution used for generating the init model ?
Is that 480 * 640 or 960 * 1280? Since according to this script (https://github.com/ibaiGorordo/CREStereo-Pytorch/blob/main/convert_to_onnx.py) init model is generated with half of the input resolution.
Relevant Log Output
No response
URL or source code for simple inference testing code
No response