javacpp-pytorch Graph Neural Networks we have. implement , and I think we need Parameter class

[mullerhai]

HI @saudet

I have one big news want to share with you , now I have implemnet gym framework with javacpp-pytorch by pure java code, it really could work , https://pyg.org/ Graph Neural Networks https://github.com/pyg-team/pytorch_geometric

so future. we could. do Graph Neural Networks build by javacpp-pytorch , it is very cool . these days. I. just clean code and want to release with javacpp-pytorch 2.9.1-1.5.13 . after I will git push the code

and. when I. write the Graph Neural Networks ,them may need Parameter object , I know in libtorch not implement Parameter , but Parameter is often use in python .
https://github.com/pytorch/pytorch/blob/main/torch/nn/parameter.py.
so could. we. implement nn.parameter in javacpp-pytorch. we also have ParameterListImpl and ParameterDictImpl ,

maybe we could write code like that

package org.example.torch.nn; 

import org.bytedeco.pytorch.Tensor;
import org.bytedeco.pytorch.global.torch;

/**
 * 模仿 Python PyTorch 的 torch.nn.Parameter 实现。
 * <p>
 * Parameter 是 Tensor 的子类，当它们作为属性分配给 Module 时，
 * 它们会被自动添加到参数列表中。
 */
public class Parameter extends Tensor {


    public Parameter(Tensor data) {
        this(data, true);
    }


    public Parameter(Tensor data, boolean requiresGrad) {
      
        super(data != null ? data.detach() : new Tensor());

        if (requiresGrad) {
            this.set_requires_grad(true);
        }
    }

    public Tensor data() {
        return this;
    }

    @Override
    public String toString() {
      
        return "Parameter containing:\n" + super.toString();
    }
  
}

import org.bytedeco.pytorch.Tensor;
import org.bytedeco.pytorch.global.torch;
import java.lang.reflect.Field;
import java.util.ArrayList;
import java.util.List;

public class MyLinearLayer {
    // 使用我们需要转译的 Parameter 类型
    public Parameter weight;
    public Parameter bias;

    public MyLinearLayer(int inFeatures, int outFeatures) {
        // 初始化权重 (Xavier/Kaiming init 均可，这里用 randn 演示)
        Tensor w = torch.randn(outFeatures, inFeatures);
        Tensor b = torch.zeros(outFeatures);

        // 包装成 Parameter
        this.weight = new Parameter(w, true);
        this.bias = new Parameter(b, true);
    }

    // 模拟 PyTorch 的 parameters() 方法
    public List<Parameter> parameters() {
        List<Parameter> params = new ArrayList<>();
        // 使用反射扫描当前类中所有的 Parameter 字段
        for (Field field : this.getClass().getDeclaredFields()) {
            if (Parameter.class.isAssignableFrom(field.getType())) {
                try {
                    params.add((Parameter) field.get(this));
                } catch (IllegalAccessException e) {
                    e.printStackTrace();
                }
            }
        }
        return params;
    }
    
    public Tensor forward(Tensor input) {
        // Parameter 可以直接当作 Tensor 使用
        return torch.addmm(this.bias, input, this.weight.t());
    }
}