I have model generated using the tensorflow for poets tutorial. I have both binary and 3-class models that are exhibiting the same behavior.
The forzen graph performs as expected but soon as I transform that to be used with opencv dnn module, the predictions are way off, usually favoring one class over others.
I used the transforms described here: http://answers.opencv.org/question/175699/readnetfromtensorflow-fails-on-retrained-nn/
and here https://www.tensorflow.org/mobile/prepare_models
Below is the output of summarize graph after each transformation. Just in case you see something unusual:
**original**
--------
No inputs spotted.
No variables spotted.
Found 1 possible outputs: (name=final_result, op=Softmax)
Found 21826166 (21.83M) const parameters, 0 (0) variable parameters, and 99 control_edges
Op types used: 489 Const, 101 Identity, 99 CheckNumerics, 94 Relu, 94 BatchNormWithGlobalNormalization, 94 Conv2D, 11 Concat, 9 AvgPool, 5 MaxPool, 1 DecodeJpeg, 1 ExpandDims, 1 Cast, 1 MatMul, 1 Mul, 1 PlaceholderWithDefault, 1 Add, 1 Reshape, 1 ResizeBilinear, 1 Softmax, 1 Sub
-----------------------
**After optimize_for_inference**
----------------------------
Found 1 possible inputs: (name=DecodeJpeg/contents, type=float(1), shape=None)
No variables spotted.
Found 1 possible outputs: (name=final_result, op=Softmax)
Found 21774517 (21.77M) const parameters, 0 (0) variable parameters, and 0 control_edges
Op types used: 206 Const, 94 BiasAdd, 94 Conv2D, 94 Relu, 11 Concat, 9 AvgPool, 5 MaxPool, 1 Sub, 1 Add, 1 Softmax, 1 ResizeBilinear, 1 Reshape, 1 PlaceholderWithDefault, 1 Placeholder, 1 Mul, 1 MatMul, 1 ExpandDims, 1 DecodeJpeg, 1 Cast
-----------------------------
**transform_graph with strip_unused**
Found 1 possible inputs: (name=Mul, type=float(1), shape=[1,299,299,3])
No variables spotted.
Found 1 possible outputs: (name=final_result, op=Softmax)
Found 21826160 (21.83M) const parameters, 0 (0) variable parameters, and 99 control_edges Op types used: 484 Const, 101 Identity, 99 CheckNumerics, 94 BatchNormWithGlobalNormalization, 94 Conv2D, 94 Relu, 11 Concat, 9 AvgPool, 5 MaxPool, 1 Add, 1 MatMul, 1 Placeholder, 1 PlaceholderWithDefault, 1 Reshape, 1 Softmax
------------------------
I'm wondering what the cause could be and how to maintain the inference performance when deploying the graph.
