Pytorch model does not seems to return the correct result

[carlosuc3m]

Hello everyone,
I am working with the Pytorch presets trying to load and execute a model saved in torchscript jit format.
However the resulting tensor is not what it should be expected.

This is the code that I am using:

public static void main(String[] args) {
    	// Load a Torchscript model
    	String modelSource = "weights-torchscript.pt";
    	JitModule model = org.bytedeco.pytorch.global.torch.load(modelSource);
    	
    	// Create an input tensor
    	org.bytedeco.pytorch.Tensor inputTensor = 
    			org.bytedeco.pytorch.Tensor.create(new float[512* 512], new long[] {1, 1, 512, 512});
    	
    	// Run the model
		model.eval();
    	IValue outputTensor = model.forward(new IValueVector(new IValue(inputTensor)));
    	
    	// Print shape of the output tensor
    	System.out.println(Arrays.toString(outputTensor.toTensor().shape()));
    	
    	// Retrieve result
    	float[] array = new float[512 * 512 * 2];
    	outputTensor.toTensor().data_ptr_float().get(array);
    	
    	System.out.println("Done");
    }

The produced outputTensor only has to unique values repeated , and it shouldn't be the case.

Is that the correct way to load and execute a model? What am I missing?
Regards,
Carlos

[saudet]

Can you try to call createIndexer() instead of data_ptr_float()?

[carlosuc3m]

It still does not work.
Here are the dependencies used:

    <dependencies>
            <groupId>org.bytedeco</groupId>
            <artifactId>pytorch-platform</artifactId>
            <version>1.13.1-1.5.9-SNAPSHOT</version>
        </dependency>
		<dependency>
            <groupId>org.bytedeco</groupId>
            <artifactId>mkl-platform-redist</artifactId>
            <version>2022.2-1.5.9-SNAPSHOT</version>
        </dependency>
    </dependencies>

[saudet]

Does it do the same in C++ without JavaCPP? It's really not complicated code, so I'm guessing this is an issue with the C++ API of PyTorch, if anything.

[carlosuc3m]

I will check it and get back with the result

[carlosuc3m]

I was wrong the whole time the results were perfect. After checking against the python implementation I realised that the results were fine all the time.