mirror of
https://github.com/babalae/better-genshin-impact.git
synced 2026-04-15 21:03:42 +08:00
4787f4f2f9
* 增加RodNet必须支持训练的单元测试；RodNet的torch链路改造成全张量计算，由此把之前忽略的参数dz、h_coeff、offset变得可学习 * 了解到损失函数CrossEntropyLoss内置了softmax，因此从forward方法中移出；offset是手动指定的偏置值，不是学习得到的，因此也移出到单独的PostProcess方法中 * 根据得到的源码整理RodNet；新增在数据集上达到一定准确率的单元测试
261 lines
10 KiB
C#
261 lines
10 KiB
C#
using System;
|
||
using System.Linq;
|
||
using static TorchSharp.torch.nn;
|
||
using static TorchSharp.torch;
|
||
using TorchSharp.Modules;
|
||
using TorchSharp;
|
||
using System.Collections.Generic;
|
||
|
||
namespace BetterGenshinImpact.GameTask.AutoFishing;
|
||
|
||
/// <summary>
|
||
/// copy from https://github.com/myHuTao-qwq/HutaoFisher/blob/master/src/rodnet.cpp
|
||
///
|
||
/// 以下是hutaofisher的访谈：
|
||
///
|
||
/// 修改我的qq昵称: 03-20 00:22:48
|
||
/// 有任何问题可以问我 只要我在工位而不是在实验室里跟XXXX斗智斗勇就回答你
|
||
///
|
||
/// 额 我的算法有几个感觉挺符合直觉的假设
|
||
///
|
||
/// 就是 你抛竿的时候鱼是否会咬钩取决于鱼饵落点到鱼的距离
|
||
///
|
||
/// 这里的距离是鱼饵落点的(x, y, z)坐标与鱼的(x, y, z)坐标之差在xy平面(水平面)上的投影
|
||
///
|
||
/// 现在问题转化成了如何估计这两个坐标
|
||
///
|
||
/// 这里我们先约定x轴就是水平面上与屏幕长边平行的线 y轴就是屏幕短边的对于 剩下一个z就是游戏里面往天上的方向
|
||
///
|
||
/// 然后经验告诉我鱼饵落点环的x几乎为0
|
||
///
|
||
/// 剩下我们可以用鱼饵落点环的长和宽算出y和z
|
||
///
|
||
/// 再近似鱼的bouding box的中心就是鱼计算咬钩的那个点 同时近似一个大类里面的所有鱼的z是相同的(这个参数可学习) 于是我们也可以计算出鱼的xyz
|
||
///
|
||
/// 最后一步就是把对应的投影距离算出来 然后线性回归一下得到太近 刚好 太远三个类
|
||
///
|
||
/// tmd 今天我意识到 XXXX可不就是XXXX
|
||
///
|
||
/// 哦 到这一步以后剩下的就很弱智了 远了挪近一点 近了挪远一点 调调参差不多得了
|
||
///
|
||
/// *后来又新增了一些访谈内容：
|
||
///
|
||
/// 额 总之就是要求不能把不咬钩的识别成咬钩的 但是咬钩的可以识别成不咬钩的
|
||
///
|
||
/// 然后就可视化一下onehot在不同距离的结果 加一个offset使得模型输出的结果在保证可以predict距离正好的结果的同时距离范围尽可能小
|
||
/// </summary>
|
||
public class RodNet : Module<Tensor, Tensor, Tensor, Tensor, Tensor>
|
||
{
|
||
const double alpha = 1734.34 / 2.5;
|
||
// fitted parameters
|
||
static readonly double[] dz = {1.0307939, 1.5887239, 1.4377865, 0.8548809,
|
||
1.8640924, -0.1687729, 1.8621461, 0.7167622,
|
||
1.7071064, 1.8727832, 0.5531539};
|
||
static readonly double[] h_coeff = {0.5840698, 0.8029298, 0.6090596,
|
||
-0.1390072, 0.7214464, -0.6076725,
|
||
0.3286690, -0.2991239, 0.6072225,
|
||
0.7662407, -0.3689651};
|
||
static readonly double[,] weight = {{0.7779633, -1.7124480, 2.7366412},
|
||
{-0.0381155, -1.6536976, 3.5904298},
|
||
{0.1947731, -0.0445049, 0.8416666},
|
||
{-0.0331017, -1.3641578, 1.2834741},
|
||
{1.0268835, -1.6553984, 2.9930501},
|
||
{0.0108103, -0.8515291, 1.0032536},
|
||
{-0.0746362, -0.9677668, 0.7450780},
|
||
{0.7382144, -9.5275803, 2.6134675},
|
||
{-0.3597502, -1.7422760, 1.4354013},
|
||
{-0.0578425, -2.0274212, 1.7173727},
|
||
{-0.1225260, -1.0630554, 1.2958838}};
|
||
static readonly double[,] bias = {{3.1733532, 9.3601589, -11.0612173},
|
||
{6.4961057, 11.2683334, -13.7752209},
|
||
{2.3662698, 2.4709859, -2.5402584},
|
||
{2.4701204, 8.5112562, -7.6070199},
|
||
{0.9597272, 8.9189463, -11.9037018},
|
||
{2.1239815, 5.8446727, -5.7748013},
|
||
{2.1403685, 5.5432696, -4.0048418},
|
||
{-9.0128260, 28.4402637, -24.2205143},
|
||
{5.2072763, 8.6428480, -9.2946615},
|
||
{4.9253063, 11.4634714, -9.4336052},
|
||
{5.2460732, 7.7711511, -7.5998945}};
|
||
|
||
|
||
static readonly double[] offset = { 0.8, 0.4, 0.35, 0.35, 0.6, 0.3, 0.3, 0.8, 0.8, 0.8, 0.8 };
|
||
|
||
private Parameter thetaParameter;
|
||
private Parameter bParameter;
|
||
private Parameter dzParameter;
|
||
private Parameter hCoeffParameter;
|
||
|
||
public RodNet() : base("RodNet")
|
||
{
|
||
long num_embeddings = RodNet.weight.GetLength(0);
|
||
long embedding_dim = 3;
|
||
|
||
this.thetaParameter = new Parameter(torch.randn(num_embeddings, embedding_dim, dtype: ScalarType.Float64));
|
||
this.bParameter = new Parameter(torch.randn(num_embeddings, embedding_dim, dtype: ScalarType.Float64));
|
||
|
||
this.dzParameter = new Parameter(torch.zeros(num_embeddings, 1, dtype: ScalarType.Float64));
|
||
this.hCoeffParameter = new Parameter(torch.zeros(num_embeddings, 1, dtype: ScalarType.Float64));
|
||
|
||
RegisterComponents();
|
||
}
|
||
|
||
static void Softmax(double[] dst, double[] x, int n)
|
||
{
|
||
double sum = 0;
|
||
for (int i = 0; i < n; i++)
|
||
{
|
||
dst[i] = Math.Exp(x[i]);
|
||
sum += dst[i];
|
||
}
|
||
for (int i = 0; i < n; i++)
|
||
{
|
||
dst[i] /= sum;
|
||
}
|
||
}
|
||
|
||
internal static int GetRodState(RodInput input)
|
||
{
|
||
double[] pred = ComputeScores(input);
|
||
|
||
return Array.IndexOf(pred, pred.Max());
|
||
}
|
||
|
||
public static double[] ComputeScores(RodInput input)
|
||
{
|
||
var (y0, z0, t, u, v, h) = GetRodStatePreProcess(input);
|
||
|
||
v -= h * h_coeff[input.fish_label];
|
||
double x, y, dist;
|
||
|
||
x = u * (z0 + dz[input.fish_label]) * Math.Sqrt(1 + t * t) / (t - v);
|
||
y = (z0 + dz[input.fish_label]) * (1 + t * v) / (t - v);
|
||
dist = Math.Sqrt(x * x + (y - y0) * (y - y0));
|
||
|
||
int fish_label = input.fish_label;
|
||
|
||
double[] logits = new double[3];
|
||
for (int i = 0; i < 3; i++)
|
||
{
|
||
logits[i] = weight[fish_label, i] * dist + bias[fish_label, i];
|
||
}
|
||
|
||
double[] pred = new double[3];
|
||
Softmax(pred, logits, 3);
|
||
pred[0] -= offset[fish_label]; // to make the prediction more precise when deployed
|
||
|
||
return pred;
|
||
}
|
||
|
||
internal int GetRodState_Torch(RodInput input)
|
||
{
|
||
using var _ = no_grad();
|
||
Tensor outputTensor = ComputeScores_Torch(input);
|
||
|
||
var max = argmax(outputTensor);
|
||
return (int)max.item<long>();
|
||
}
|
||
|
||
public Tensor ComputeScores_Torch(RodInput input)
|
||
{
|
||
using var _ = no_grad();
|
||
this.SetWeightsManually();
|
||
|
||
var (y0, z0, t, u, v, h) = GetRodStatePreProcess(input);
|
||
|
||
Tensor fishLabel = tensor(new double[] { input.fish_label }, dtype: ScalarType.Int32);
|
||
Tensor uv = tensor(new double[,] { { u, v } }, dtype: ScalarType.Float64);
|
||
Tensor y0z0t = tensor(new double[,] { { y0, z0, t } }, dtype: ScalarType.Float64);
|
||
Tensor h_ = tensor(new double[,] { { h } }, dtype: ScalarType.Float64);
|
||
|
||
var logits = forward(fishLabel, uv, y0z0t, h_);
|
||
var output = PostProcess(logits, fishLabel);
|
||
|
||
return output;
|
||
}
|
||
|
||
/// <summary>
|
||
/// 使用时直接赋值已知权重
|
||
/// </summary>
|
||
public void SetWeightsManually()
|
||
{
|
||
var weightTensor = tensor(RodNet.weight, ScalarType.Float64);
|
||
var biasTensor = tensor(RodNet.bias, ScalarType.Float64);
|
||
var dzTensor = tensor(RodNet.dz, ScalarType.Float64).reshape([RodNet.dz.Length, 1]);
|
||
var h_coeffTensor = tensor(RodNet.h_coeff, ScalarType.Float64).reshape([RodNet.h_coeff.Length, 1]);
|
||
this.thetaParameter = new Parameter(weightTensor);
|
||
this.bParameter = new Parameter(biasTensor);
|
||
this.dzParameter = new Parameter(dzTensor);
|
||
this.hCoeffParameter = new Parameter(h_coeffTensor);
|
||
}
|
||
|
||
public override Tensor forward(Tensor fishLabel, Tensor uv, Tensor y0z0t, Tensor h)
|
||
{
|
||
var uvSplit = uv.split([1, 1], dim: 1);
|
||
Tensor u = uvSplit[0];
|
||
Tensor v = uvSplit[1];
|
||
|
||
var y0z0tSplit = y0z0t.split([1, 1, 1], dim: 1);
|
||
Tensor y0 = y0z0tSplit[0];
|
||
Tensor z0 = y0z0tSplit[1];
|
||
Tensor t = y0z0tSplit[2];
|
||
|
||
v = v - h * hCoeffParameter[fishLabel];
|
||
|
||
Tensor x, y, dist;
|
||
|
||
var dz = dzParameter[fishLabel];
|
||
x = u * (z0 + dz) * torch.sqrt(1 + t * t) / (t - v);
|
||
y = (z0 + dz) * (1 + t * v) / (t - v);
|
||
dist = torch.sqrt(x * x + (y - y0) * (y - y0));
|
||
|
||
Tensor logits = this.thetaParameter[fishLabel] * dist + this.bParameter[fishLabel];
|
||
|
||
return logits;
|
||
}
|
||
|
||
public Tensor PostProcess(Tensor logits, Tensor fishLabel)
|
||
{
|
||
var x_softmax = torch.nn.functional.softmax(logits, 1);
|
||
|
||
Tensor x_offset = tensor(fishLabel.data<int>().Select(l => RodNet.offset[l]).ToArray());
|
||
|
||
x_softmax[torch.arange(x_offset.shape[0]), 0] -= x_offset;
|
||
return x_softmax;
|
||
}
|
||
|
||
/// <summary>
|
||
/// 根据rod和fish的坐标计算y0z0t、uv、h
|
||
/// </summary>
|
||
/// <param name="input"></param>
|
||
/// <returns>y0, z0, t, u, v, h</returns>
|
||
public static (double, double, double, double, double, double) GetRodStatePreProcess(RodInput input)
|
||
{
|
||
/*
|
||
* 以下为hutaofisher代码中关于部分变量的意义的注释
|
||
# uv: screen coordinate of bbox center of the fish
|
||
# abv0: rod shape and center coordinate in screen
|
||
*/
|
||
double a, b, v0, u, v, h;
|
||
|
||
a = (input.rod_x2 - input.rod_x1) / 2 / alpha;
|
||
b = (input.rod_y2 - input.rod_y1) / 2 / alpha;
|
||
h = (input.fish_y2 - input.fish_y1) / 2 / alpha;
|
||
|
||
if (a < b)
|
||
{
|
||
b = Math.Sqrt(a * b);
|
||
a = b + 1e-6;
|
||
}
|
||
v0 = (288 - (input.rod_y1 + input.rod_y2) / 2) / alpha;
|
||
u = (input.fish_x1 + input.fish_x2 - input.rod_x1 - input.rod_x2) / 2 / alpha;
|
||
v = (288 - (input.fish_y1 + input.fish_y2) / 2) / alpha;
|
||
double y0, z0, t;
|
||
|
||
y0 = Math.Sqrt(Math.Pow(a, 4) - b * b + a * a * (1 - b * b + v0 * v0)) / (a * a);
|
||
z0 = b / (a * a);
|
||
t = a * a * (y0 * b + v0) / (a * a - b * b);
|
||
|
||
return (y0, z0, t, u, v, h);
|
||
}
|
||
} |