Different errors on conversion

[edward9112]

Can't convert layer : id = 210 , name = Gather_199549/Cast_122100_const , type = Const , version = opset1 !
Can't convert IE model v10!

Unknown element_type = i32 !
Can't convert layer : id = 383 , name = up_sampling2d/Shape/GatherNCHWtoNHWC/Cast_123490_const , type = Const , version = opset1 !
Can't convert IE model v10!

Basically I can convert only the oldest, early 2019 models.

[ermig1979]

Hello. The support of IE model v10 is not full. If you give me your model I will fix the error when I have free time.

[edward9112]

Awesome! Just sent you the model via email.

[ermig1979]

The bug was fixed.

[edward9112]

Thank you!

Can you as well add the yolo models support? (they produce error now) Can't convert layer : id = 59 , name = predict_conv/BiasAdd/YoloRegion , type = RegionYolo , version = opset1 ! Can't convert IE model v10!

[ermig1979]

Ок. I will check it next week.

[edward9112]

Awesome! And, by the way, person-vehicle-bike-detection-2002 does not do the inference properly - it just generates a bunch of rectangles, although conversion works well. Not sure if it's a model or conversion issue.

[ermig1979]

I fixed this error today. You have to perform conversion again.

[edward9112]

By the way, FP16 models also do not covert, here is what it says:

Unknown element_type = f16 !
Can't convert layer : id = 1 , name = data_mul_/copy_const , type = Const , version = opset1 !
Can't convert IE model v10!

[ermig1979]

Synet does not support FP16 models (they are for GPU only).

[edward9112]

Synet does not support FP16 models (they are for GPU only).

Got it. What about INT8?

[ermig1979]

INT8 is supported, but network output is not equal to openvino output. There is a rounding error (FP32 to INT8 conversion).

[edward9112]

What about FP32-INT8 models?

[ermig1979]

I am sorry. Heretofore I kept in mind FP32-INT8 models of course. So Synet support FP32-INT8 models, not INT8.

[edward9112]

I get the following errors while converting different FP32-INT8 models: Can't convert layer : id = 9 , name = init_block1/dim_inc/conv/fq_input_0 , type = FakeQuantize , version = opset1 !

Can't convert layer : id = 13 , name = L0008_ActivationBin-back_bone_seq.conv2_2_sep_relubin_bin_convBIN01/Quantize , type = Quantize !

The models are: person-detection-retail-0013 pedestrian-detection-adas-binary-0001

[ermig1979]

I am developing own method of quantization of FP32-INT8 models. Support of conversion of FP32-INT8 Openvino models is dead end way. So I will suport only conversion of FP32 openvino models.

[edward9112]

Got it. Is your method also open source?

[ermig1979]

Of course. See Synet/src/Test/TestQuantization.cpp

[edward9112]

TestQuantization says "Can't load param.xml". Is there a list of required arguments?

[ermig1979]

File 'param.xml' contains additional information to run network in tests (input scale range, etc.). See declaration of TestParam structure in file TestNetwork.h. Each Synet's test contain such file.

[edward9112]

Is there a typical sample of the param file? The declaration has many parameters that require some knowledge about the quantizer which most people don't have.

[ermig1979]

Quantization is complicated process and not fully automized. This operation leads to accuracy degradation and requires test samples. Also it requires independent verification of final accuracy.

[edward9112]

Got it, thanks! Any updates on the yolo models? By the way, I tried to convert the yolo v3&4 models from here https://github.com/AlexeyAB/darknet The conversion went well but the inference just didn't work.

[Berzhik]

Hello, I found your framework recently, and it's amazing! But I have some problem with conversion this model I have the following errors:

Not implemented layer : name = tile1033 ; type = Tile
Not implemented layer : name = decode ; type = CTCGreedyDecoder

Could you help me, please?

[ermig1979]

Hello, I found your framework recently, and it's amazing! But I have some problem with conversion this model I have the following errors:

Not implemented layer : name = tile1033 ; type = Tile
Not implemented layer : name = decode ; type = CTCGreedyDecoder

Could you help me, please?

Thank you for bug report! The bug was fixed. See test_014.

[Berzhik]

Hello, I have the question on how to use the license plate recognition model That model has 2 inputs and Network::SetInput doesn't work, because there is a condition on 1 input in Synet::SetInput

[ermig1979]

You have to use function Network::Src and set inputs manually. As an example you can look at SynetNetwork::SetInput in file TestSynet.h.