Trouble Loading in Python trained model into C#

[oridayne]

I have trained a model in python, and am trying to load it into C#.

I do, Function.Load(path, device), and I printed the inputs and outputs. I noticed that there are actually two inputs, and 5 expected outputs, which really confused me, because the model I trained in Python should only take one input (a one_hot vector), and have one output( also a vector).

When I print various attributes of the model in python, I get: print(model.argument_map) Composite(Sequence[Tensor[1]]) -> Tensor[2] print(mode.arguments) (Input('Input1685923', [#, *], [1]),)

I'm really unfamiliar with the notation, but it seems that it's just a single variable with shape=1, which is what I defined it as earlier: X = C.sequence.input_variable(shape=(1))

As a result, I'm really confused why there are two inputs required for the C# loaded version, and why there are 5 required for the outputs.

I tried inputing and outputing variables into a dictionary and evaluating them in C#, but none of the numbers look remotely familiar to what I should be getting, and they are not even in the same expected format.

Does anyone have any tips loading in models into C# that are trained in python?

[Tixxx]

Hi @oridayne, could you share the model that you are building? I just tried loading a simple python model in c#, the inputs and outputs are expected: Python model: import cntk as C f = C.layers.BatchNormalization(map_rank=1) f.update_signature((3,480,640)) f

output of this line: Composite(BatchNormalization): Input('x', [#], [3 x 480 x 640]) -> Output('Block58_Output_0', [#], [3 x 480 x 640])

f.save("test.model") When i load test.model into c#, it shows 6 inputs and 1 output. The 6 inputs are because the batch normalization is defined to have default parameters like this:

def BatchNormalization(map_rank=default_override_or(None),
init_scale=1, normalization_time_constant=default_override_or(5000), blend_time_constant=0, epsilon=default_override_or(0.00001), use_cntk_engine=default_override_or(False), disable_regularization=default_override_or(False), name=''): Maybe you can check the function you are using and see if they contain default parameters?

[oridayne]

Hey Tix, thanks for replying, and that's a pretty good idea. The current model I was testing on was taken from someone else's notebook so I don't have a complete understanding of it.

I have actually tried to test this on another simpler python cntk model: https://github.com/Microsoft/CNTK/blob/v2.5.1/Tutorials/CNTK_101_LogisticRegression.ipynb

I prefer to show this model over mine because it's well documented, and exhibits similar characteristics to the problem I described above. However, I had trouble finding where these default parameters were set, but I did discover some newer things...

The inputs and outputs of the python model are: Composite(Tensor[2]) -> Tensor[2]. I take the output vector and do C.argmax([x,y]) -> which returns a 0 or 1 as my prediction result.

The C# implementation also shows 2 inputs, and 5 outputs.

I used the code below and instead just took the first input, and the first output `
model = Function.Load(modelPath, DeviceDescriptor.CPUDevice); List vect = vectorizeName(name); // convert to one-hot vector IList inputs = model.Arguments; IList output = model.Outputs;

    // specify input shape
    var inputDataMap = new Dictionary<Variable, Value>();
    var inputVal = Value.CreateSequence(inputs[0].Shape, vect, DeviceDescriptor.CPUDevice);
    inputDataMap.Add(inputs[0], inputVal);

    // specify output shape
    var outputDataMap = new Dictionary<Variable, Value>();
    outputDataMap.Add(output[0], null);

    model.Evaluate(inputDataMap, outputDataMap, DeviceDescriptor.CPUDevice);

    var outputVal = outputDataMap[output[0]];
    var outputData = outputVal.GetDenseData<float>(output[0]);
    var result = outputData[0];`

Result would return me a vector with 2 inputs. While the 2 inputs are different numbers than I would get from the python model, the argmax of the vector returns the expected result (tested 100 times). For example, the python model would return [-28.50, 2.567] while the C# model would return [-3.56, 3.01], but argument max of both is 1.

I haven't tried out using the two inputs and 5 outputs, but I'm not sure how to even use them in the eval functions. I'll try to search through the logistic regression model for what layer exactly defines the default paramters.

[ke1337]

Please note that data in C++/C# are column major, while the data in Python is row-major.

If you are building from source, please comment this line in C# swig binding and use CNTKLib.SetComputationNetworkTraceLevel(1000000) to dump the output of nodes from C#, and compare with the dumps in Python using set_computation_network_trace_level(1000000).

[oridayne]

The C++C# column major issue was not the issue.

In the end my numbers were right, I just had to use CreateBatch instead of Create sequence. The model now evaluates correctly in C#. As long as I took only the first input and output from the C# model for evaluation, it works.

Thank you for commenting!