Model freezing

[pay20y]

Model freezing

[Pay20Y]

Sorry I have not tried it. Since I'm still doing experiments to reimplement the effect in paper. When I make it I'll release a frozen model.

[bharatsubedi]

I tried to freeze model but I faced one issue for model freezing. For detection session output of detection is f_score and f_geometry is in a graph so it will easily convert but in cases of recognition session dense_decode which is the output node of recognition is not inside the graph. do you know how to solve and freeze model? could you provide some hints on how can I solve this issue?

[Pay20Y]

Maybe reimplement the NMS and the postprocess of detection(generating box from f_geometry) with tensorflow api.

[bharatsubedi]

could you explain more about model freezing? I tried different way but fails to get dense_decode output inside recog_decode session.

[Pay20Y]

Sorry, I have never tried to freeze the model.

[lerndeep]

hello pay20Y,

I am also doing model freezing. I got output dense_decode but when i tried to run a session for recognition feeding input value for input_box_mask I got the issue.

Due to shape() issue I could not feed value into the session. could you provide some information?

[Pay20Y]

@lerndeep input_box_mask is a list of 1D tensor. It is used to indicate which feature map that rois belonging to. So if you run with batch size of 1 the feeding input value is [0] * num_of_rois

[lerndeep]

Thank you for your reply. Now I solved the tensor size issue but I got another Issue "input_box_masks_.

This is my detail implementation code.

[Alek-dr]

Have somebody success in freezing?