OpenVino support variable Input image, but results are different

[TonyWong]

Dear Intel Guru,

Hi, I have tried OpenVino, and it is awesome.

I'm having another question.

My input model size is [1,3, height, width]

I just tried, and it seems it does support variable input size. 

However: 

Using the mo.py,  though the warming still mentioned that it is still strongly suggested to use static input size.

The results are different with original ONNX format. 

Anyone can help?

Here is the  model link:

and the command  is :

$python mo.py --input_model model.onnx --output_dir ~/Downloads

Tony

[TonyWong]

The key is: results are different.

Even the dimension   

from 48*80*606

to: 30*40*606

[Iffa_Intel]

Hi @TonyWong ,

From the Model Optimizer, you can feed the input shape by using --input_shape parameter.

Example:  python mo.py --input_model model.onnx --output_dir ~/Downloads --input_shape [1,3,227,227]

Note that the order of dimension depends on the framework input layout of the model.

For instance:

1. Caffe model: [N,C,H,W]

2. Tensorflow model: [N,H,W,C]

Model Optimizer requires [N,C,H,W] layout and would perform the necessary transformation to the model. The shape should not contain undefined dimensions such as (? or -1).

Depending on your model, you need to consider these:

1. When to specify input shape

2. When to reverse input channel

Hope this helps!

Sincerely,

Iffa

[Iffa_Intel]

Greetings,

Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question. 

Sincerely,

Iffa