Request for input shapes to be specified

[danielliuce]

Feature request

Currently,
optimum-cli does not provide a way to specify static input shapes, it defaults to dynamic shapes. Is there a way to make it possible to specify the input shape? If not, why do we not allow this?

An example would be:
optimum-cli export openvino --model microsoft/resnet-50 graph_convert -> optimum-cli export openvino --model microsoft/resnet-50 graph_convert --input [1, 3, 224, 224]

Motivation

Specifying a static shape in OpenVINO IR is nice to have for the Intel/Altera FPGA AI Suite toolchain which does not support dynamic input shapes of OpenVINO IR at the moment

Your contribution

Yes if possible or the green light is given that this is allowed.
Some modifications to the optimum_cli.py file here
would probably be needed

[IlyasMoutawwakil]

Hi ! is this related to openvino or onnx export ? in onnx I believe you can disable dynamic axes with https://github.com/huggingface/optimum-onnx/blob/main/optimum/commands/export/onnx.py#L154 and specify the input shapes with https://github.com/huggingface/optimum-onnx/blob/main/optimum/commands/export/onnx.py#L171-L245

[danielliuce]

Hi ! is this related to openvino or onnx export ? in onnx I believe you can disable dynamic axes with https://github.com/huggingface/optimum-onnx/blob/main/optimum/commands/export/onnx.py#L154 and specify the input shapes with https://github.com/huggingface/optimum-onnx/blob/main/optimum/commands/export/onnx.py#L171-L245

This is a feature request for the OpenVINO export. Thanks for showing me that this is posible with an onnx support, I believe the feature would be the exact same but for an OpenVINO export

[danielliuce]

Hi ! is this related to openvino or onnx export ? in onnx I believe you can disable dynamic axes with https://github.com/huggingface/optimum-onnx/blob/main/optimum/commands/export/onnx.py#L154 and specify the input shapes with https://github.com/huggingface/optimum-onnx/blob/main/optimum/commands/export/onnx.py#L171-L245

The FPGA AI Suite supports OpenVINO as a plugin inference accelerator. But this plugin does not support dynamic shapes, so we need a way to create OpenVINO IR without the dynamic shapes