Source: Deep Learning on Medium
Thank you very much for your nice article. I would like to discuss some issues. I am trying to convert keras model (.h5) into tensorflow (.pb) for opencv DNN. I converted into .pb and also applied your steps up to Transforming Graph. I can not go further after graph transformation due to some confusion. Would you help me please by answering my questions:
- What about your steps for multiple inputs and outputs; because in your case there is only one input node.
- Which placeholders will I replace from my graph? How can I identify those placeholders?
- In batch normalization, which one I remove and which one I keep? That means How can I identify those ops from graph?
Actually, after transforming, when I used the model with opencv DNN, it gives the following error:
error: OpenCV(4.1.0) opencv\modules\dnn\src\tensorflow\tf_importer.cpp:497: error: (-2:Unspecified error) Input layer not found: ROI_5/split/value in function ‘cv::dnn::dnn4_v20190122::`anonymous-namespace’::TFImporter::connect’