mirror of
https://github.com/babalae/better-genshin-impact.git
synced 2026-05-21 09:45:48 +08:00
尝试复现RodNet的训练结果 (#1688)
* 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
This commit is contained in:
FishmanTheMurloc
GitHub
@@ -0,0 +1,392 @@
|
||||
using System;
|
||||
using System.Collections.Generic;
|
||||
using System.Linq;
|
||||
using System.Text;
|
||||
using System.Threading.Tasks;
|
||||
using static TorchSharp.torch.nn;
|
||||
using static TorchSharp.torch;
|
||||
using TorchSharp;
|
||||
using System.Diagnostics;
|
||||
using System.Collections;
|
||||
using BetterGenshinImpact.GameTask.AutoFishing;
|
||||
|
||||
namespace BetterGenshinImpact.UnitTest.GameTaskTests.AutoFishingTests
|
||||
{
|
||||
public partial class RodNetTests
|
||||
{
|
||||
/// <summary>
|
||||
/// RodNet验证，应在数据集上达到一定准确率
|
||||
/// </summary>
|
||||
[Theory]
|
||||
[InlineData(@"..\..\..\Assets\AutoFishing\data_selected.csv")]
|
||||
public void Training_AccuracyShouldBeOK(string dataLocation)
|
||||
{
|
||||
//
|
||||
using var _ = no_grad();
|
||||
|
||||
var device =
|
||||
torch.cuda.is_available() ? torch.CUDA :
|
||||
torch.mps_is_available() ? torch.MPS :
|
||||
torch.CPU;
|
||||
var loss = CrossEntropyLoss();
|
||||
var sut = new RodNet().to((Device)device);
|
||||
sut.SetWeightsManually();
|
||||
|
||||
using var test_reader = new CSVReader(Enumerable.Repeat(false, 8).Concat(Enumerable.Repeat(true, 2)), Path.GetFullPath(dataLocation), (Device)device);
|
||||
|
||||
//
|
||||
var accuracy = evaluate(test_reader.GetBatches(eval_batch_size), sut, loss);
|
||||
|
||||
//
|
||||
Assert.True(accuracy > 0.8);
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// RodNet必须粗略地支持训练
|
||||
/// </summary>
|
||||
[Fact]
|
||||
public void Training_ShouldBeDifferentiable()
|
||||
{
|
||||
//
|
||||
RodInput input = new RodInput();
|
||||
var (y0, z0, t, u, v, h) = RodNet.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);
|
||||
RodNet sut = new RodNet();
|
||||
|
||||
//
|
||||
Tensor output = sut.forward(fishLabel, uv, y0z0t, h_);
|
||||
output.backward([torch.ones_like(output)]);
|
||||
//
|
||||
}
|
||||
|
||||
#region 训练相关代码
|
||||
// 这部分代码改编自TorchSharpExamples的CSharpExamples.TextClassification
|
||||
private const long batch_size = 32;
|
||||
private const long eval_batch_size = 32;
|
||||
|
||||
internal static RodNet Run(int epochs, int timeout, string dataLocation)
|
||||
{
|
||||
torch.random.manual_seed(1);
|
||||
|
||||
var device =
|
||||
torch.cuda.is_available() ? torch.CUDA :
|
||||
torch.mps_is_available() ? torch.MPS :
|
||||
torch.CPU;
|
||||
|
||||
Console.WriteLine();
|
||||
Console.WriteLine($"\tRunning TextClassification on {device.type.ToString()} for {epochs} epochs, terminating after {TimeSpan.FromSeconds(timeout)}.");
|
||||
Console.WriteLine();
|
||||
|
||||
Console.WriteLine($"\tPreparing training and test data...");
|
||||
|
||||
using (var reader = new CSVReader(Enumerable.Repeat(true, 8).Concat(Enumerable.Repeat(false, 2)), dataLocation, (Device)device))
|
||||
{
|
||||
Console.WriteLine($"\tCreating the model...");
|
||||
Console.WriteLine();
|
||||
|
||||
var model = new RodNet().to((Device)device);
|
||||
|
||||
var loss = CrossEntropyLoss();
|
||||
var lr = 1e-2;
|
||||
var optimizer = torch.optim.SGD(model.parameters(), learningRate: lr);
|
||||
var scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, epochs, eta_min: 0);
|
||||
|
||||
var totalTime = new Stopwatch();
|
||||
totalTime.Start();
|
||||
|
||||
foreach (var epoch in Enumerable.Range(1, epochs))
|
||||
{
|
||||
|
||||
var sw = new Stopwatch();
|
||||
sw.Start();
|
||||
|
||||
train(epoch, reader.GetBatches(batch_size), model, loss, optimizer);
|
||||
|
||||
sw.Stop();
|
||||
|
||||
Console.WriteLine($"\nEnd of epoch: {epoch} | lr: {optimizer.ParamGroups.First().LearningRate:0.0000} | time: {sw.Elapsed.TotalSeconds:0.0}s\n");
|
||||
scheduler.step();
|
||||
|
||||
if (totalTime.Elapsed.TotalSeconds > timeout) break;
|
||||
}
|
||||
|
||||
totalTime.Stop();
|
||||
|
||||
using (var test_reader = new CSVReader(Enumerable.Repeat(false, 8).Concat(Enumerable.Repeat(true, 2)), dataLocation, (Device)device))
|
||||
{
|
||||
|
||||
var sw = new Stopwatch();
|
||||
sw.Start();
|
||||
|
||||
var accuracy = evaluate(test_reader.GetBatches(eval_batch_size), model, loss);
|
||||
|
||||
sw.Stop();
|
||||
|
||||
Console.WriteLine($"\nEnd of training: test accuracy: {accuracy:0.00} | eval time: {sw.Elapsed.TotalSeconds:0.0}s\n");
|
||||
scheduler.step();
|
||||
}
|
||||
|
||||
foreach (var (name, param) in model.named_parameters())
|
||||
{
|
||||
switch (param.dtype)
|
||||
{
|
||||
case ScalarType.Int64:
|
||||
Console.WriteLine($"参数{name}={String.Join(", ", param.data<long>())}");
|
||||
break;
|
||||
case ScalarType.Float32:
|
||||
Console.WriteLine($"参数{name}={String.Join(", ", param.data<float>())}");
|
||||
break;
|
||||
case ScalarType.Float64:
|
||||
Console.WriteLine($"参数{name}={String.Join(", ", param.data<double>())}");
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
return model;
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
static void train(int epoch, IEnumerable<(Tensor, Tensor, Tensor, Tensor, Tensor)> train_data, RodNet model, Loss<Tensor, Tensor, Tensor> criterion, torch.optim.Optimizer optimizer)
|
||||
{
|
||||
model.train();
|
||||
|
||||
double total_acc = 0.0;
|
||||
long total_count = 0;
|
||||
long log_interval = 1;
|
||||
|
||||
var batch = 0;
|
||||
|
||||
var batch_count = train_data.Count();
|
||||
|
||||
using (var d = torch.NewDisposeScope())
|
||||
{
|
||||
foreach (var (y0z0t, uv, h, fish_label, success) in train_data)
|
||||
{
|
||||
|
||||
optimizer.zero_grad();
|
||||
|
||||
using (var predicted_labels = model.forward(fish_label, uv, y0z0t, h))
|
||||
{
|
||||
var loss = criterion.forward(predicted_labels, success.to(ScalarType.Int64));
|
||||
loss.backward();
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1);
|
||||
optimizer.step();
|
||||
|
||||
total_acc += (predicted_labels.argmax(1) == success).sum().to(torch.CPU).item<long>();
|
||||
total_count += success.size(0);
|
||||
}
|
||||
|
||||
batch += 1;
|
||||
if (batch % log_interval == 0)
|
||||
{
|
||||
var accuracy = total_acc / total_count;
|
||||
Console.WriteLine($"epoch: {epoch} | batch: {batch} / {batch_count} | accuracy: {accuracy:0.00}");
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static double evaluate(IEnumerable<(Tensor, Tensor, Tensor, Tensor, Tensor)> test_data, RodNet model, Loss<Tensor, Tensor, Tensor> criterion)
|
||||
{
|
||||
model.eval();
|
||||
|
||||
double total_acc = 0.0;
|
||||
long total_count = 0;
|
||||
|
||||
using (var d = torch.NewDisposeScope())
|
||||
{
|
||||
foreach (var (y0z0t, uv, h, fish_label, success) in test_data)
|
||||
{
|
||||
using (var predicted_labels = model.forward(fish_label, uv, y0z0t, h))
|
||||
{
|
||||
var loss = criterion.forward(predicted_labels, success.to(ScalarType.Int64));
|
||||
|
||||
total_acc += (predicted_labels.argmax(1) == success).sum().to(torch.CPU).item<long>();
|
||||
total_count += success.size(0);
|
||||
}
|
||||
}
|
||||
|
||||
return total_acc / total_count;
|
||||
}
|
||||
}
|
||||
#endregion
|
||||
}
|
||||
|
||||
internal class CSVReader : IDisposable
|
||||
{
|
||||
/// <summary>
|
||||
///
|
||||
/// </summary>
|
||||
/// <param name="takeMask">按长度分组，布尔值代表每组内此序号元素是否被读取。例如8个true2个false就是将近80%进入训练集</param>
|
||||
/// <param name="path"></param>
|
||||
/// <param name="device"></param>
|
||||
public CSVReader(IEnumerable<bool> takeMask, string path, Device device)
|
||||
{
|
||||
this.takeMask = takeMask.ToArray();
|
||||
_path = path;
|
||||
_device = device;
|
||||
}
|
||||
|
||||
private readonly bool[] takeMask;
|
||||
private readonly string _path;
|
||||
private readonly Device _device;
|
||||
|
||||
|
||||
public IEnumerable<Data> Enumerate()
|
||||
{
|
||||
var all = File.ReadLines(_path).Skip(1); // 跳过首行列名
|
||||
int count = takeMask.Length;
|
||||
int maskCount = all.Count() / count;
|
||||
for (int i = 0; i < maskCount + 1; i++)
|
||||
{
|
||||
int lastGroupCount = (i == maskCount) ? (all.Count() % count) : count;
|
||||
for (int j = 0; j < lastGroupCount; j++)
|
||||
{
|
||||
if (takeMask[j])
|
||||
{
|
||||
yield return ParseLine(all.Skip(i * count + j).First());
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
public IEnumerable<(Tensor, Tensor, Tensor, Tensor, Tensor)> GetBatches(long batch_size)
|
||||
{
|
||||
// This data set fits in memory, so we will simply load it all and cache it between epochs.
|
||||
|
||||
var inputs = new List<Data>();
|
||||
|
||||
if (_data == null)
|
||||
{
|
||||
|
||||
_data = new List<(Tensor, Tensor, Tensor, Tensor, Tensor)>();
|
||||
|
||||
var counter = 0;
|
||||
var lines = Enumerate().ToList();
|
||||
var left = lines.Count;
|
||||
|
||||
foreach (var line in lines)
|
||||
{
|
||||
|
||||
inputs.Add(line);
|
||||
left -= 1;
|
||||
|
||||
if (++counter == batch_size || left == 0)
|
||||
{
|
||||
_data.Add(Batchifier(inputs));
|
||||
inputs.Clear();
|
||||
counter = 0;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return _data;
|
||||
}
|
||||
|
||||
private List<(Tensor, Tensor, Tensor, Tensor, Tensor)> _data;
|
||||
private bool disposedValue;
|
||||
|
||||
/// <summary>
|
||||
/// 将csv中的数据进行初步转换
|
||||
/// </summary>
|
||||
/// <param name="input"></param>
|
||||
/// <returns>y0z0t、uv、h、fish_label、success张量</returns>
|
||||
private (Tensor, Tensor, Tensor, Tensor, Tensor) Batchifier(IEnumerable<Data> input)
|
||||
{
|
||||
var y0List = new List<double>();
|
||||
var z0List = new List<double>();
|
||||
var tList = new List<double>();
|
||||
var uList = new List<double>();
|
||||
var vList = new List<double>();
|
||||
var hList = new List<double>();
|
||||
var labelList = new List<int>();
|
||||
var successList = new List<int>();
|
||||
|
||||
foreach (var line in input)
|
||||
{
|
||||
int fish_label = line.fish_label;
|
||||
int success = line.success;
|
||||
RodInput rodInput = new RodInput()
|
||||
{
|
||||
rod_x1 = line.rod_x1,
|
||||
rod_x2 = line.rod_x2,
|
||||
rod_y1 = line.rod_y1,
|
||||
rod_y2 = line.rod_y2,
|
||||
fish_x1 = line.fish_x1,
|
||||
fish_x2 = line.fish_x2,
|
||||
fish_y1 = line.fish_y1,
|
||||
fish_y2 = line.fish_y2
|
||||
};
|
||||
var (y0, z0, t, u, v, h) = RodNet.GetRodStatePreProcess(rodInput);
|
||||
|
||||
y0List.Add(y0);
|
||||
z0List.Add(z0);
|
||||
tList.Add(t);
|
||||
uList.Add(u);
|
||||
vList.Add(v);
|
||||
hList.Add(h);
|
||||
labelList.Add(fish_label);
|
||||
successList.Add(success);
|
||||
}
|
||||
|
||||
Tensor y0Tensor = tensor(y0List, dtype: ScalarType.Float64).to(_device);
|
||||
Tensor z0Tensor = tensor(z0List, dtype: ScalarType.Float64).to(_device);
|
||||
Tensor tTensor = tensor(tList, dtype: ScalarType.Float64).to(_device);
|
||||
Tensor uTensor = tensor(uList, dtype: ScalarType.Float64).to(_device);
|
||||
Tensor vTensor = tensor(vList, dtype: ScalarType.Float64).to(_device);
|
||||
Tensor hTensor = tensor(hList, dtype: ScalarType.Float64).to(_device);
|
||||
Tensor fish_labelTensor = tensor(labelList, dtype: ScalarType.Int32).to(_device);
|
||||
Tensor successTensor = tensor(successList, dtype: ScalarType.Int32).to(_device);
|
||||
|
||||
return (torch.stack([y0Tensor, z0Tensor, tTensor], dim: 1), torch.stack([uTensor, vTensor], dim: 1), hTensor.unsqueeze(1), fish_labelTensor, successTensor);
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// csv的列定义默认为time,bite_time,rod_x1,rod_x2,rod_y1,rod_y2,fish_x1,fish_x2,fish_y1,fish_y2,fish_label,success
|
||||
/// </summary>
|
||||
/// <param name="line"></param>
|
||||
/// <returns></returns>
|
||||
public Data ParseLine(string line)
|
||||
{
|
||||
var columns = line.Split(",").ToArray();
|
||||
|
||||
return new Data(float.Parse(columns[2]), float.Parse(columns[3]), float.Parse(columns[4]), float.Parse(columns[5]),
|
||||
float.Parse(columns[6]), float.Parse(columns[7]), float.Parse(columns[8]), float.Parse(columns[9]),
|
||||
int.Parse(columns[10]), int.Parse(columns[11]));
|
||||
}
|
||||
|
||||
protected virtual void Dispose(bool disposing)
|
||||
{
|
||||
if (!disposedValue)
|
||||
{
|
||||
if (disposing && _data != null)
|
||||
{
|
||||
foreach (var (y0z0t, uv, h, label, success) in _data)
|
||||
{
|
||||
y0z0t.Dispose();
|
||||
uv.Dispose();
|
||||
h.Dispose();
|
||||
label.Dispose();
|
||||
success.Dispose();
|
||||
}
|
||||
}
|
||||
|
||||
disposedValue = true;
|
||||
}
|
||||
}
|
||||
|
||||
public void Dispose()
|
||||
{
|
||||
// Do not change this code. Put cleanup code in 'Dispose(bool disposing)' method
|
||||
Dispose(disposing: true);
|
||||
GC.SuppressFinalize(this);
|
||||
}
|
||||
}
|
||||
|
||||
internal record Data(float rod_x1, float rod_x2, float rod_y1, float rod_y2, float fish_x1, float fish_x2, float fish_y1, float fish_y2, int fish_label, int success);
|
||||
}
|
||||
@@ -1,17 +1,16 @@
|
||||
using BetterGenshinImpact.GameTask.AutoFishing;
|
||||
using BetterGenshinImpact.UnitTest.GameTaskTests.AutoFishingTests;
|
||||
using System;
|
||||
using System.Collections.Generic;
|
||||
using System.Linq;
|
||||
using System.Text;
|
||||
using System.Threading.Tasks;
|
||||
using static BetterGenshinImpact.GameTask.AutoFishing.RodNet;
|
||||
using TorchSharp;
|
||||
using static TorchSharp.torch;
|
||||
|
||||
namespace BetterGenshinImpact.UnitTest.GameTaskTests.AutoFishingTests
|
||||
{
|
||||
[Collection("Init Collection")]
|
||||
public class RodNetTests
|
||||
public partial class RodNetTests
|
||||
{
|
||||
public RodNetTests(TorchFixture torch)
|
||||
{
|
||||
@@ -42,9 +41,8 @@ namespace BetterGenshinImpact.UnitTest.GameTaskTests.AutoFishingTests
|
||||
RodNet sut = new RodNet();
|
||||
|
||||
//
|
||||
NetInput netInput = GeometryProcessing(rodInput) ?? throw new NullReferenceException();
|
||||
Tensor outputTensor = sut.ComputeScores_Torch(netInput);
|
||||
double[] pred = ComputeScores(netInput);
|
||||
Tensor outputTensor = sut.ComputeScores_Torch(rodInput);
|
||||
double[] pred = RodNet.ComputeScores(rodInput);
|
||||
|
||||
//
|
||||
Assert.Equal((float)pred[0], (float)outputTensor.data<double>()[0]); // 对比时降低精度，差不多就行
|
||||
|
||||
Reference in New Issue
Block a user