mega epoch

Cargo.lock

@@ -5451,6 +5451,7 @@ name = "strafe-ai"
 version = "0.1.0"
 dependencies = [
  "burn",
+ "glam",
  "pollster",
  "strafesnet_common",
  "strafesnet_graphics",

Cargo.toml

@@ -14,3 +14,4 @@ strafesnet_roblox_bot_file = { version = "0.9.4", registry = "strafesnet" }
 strafesnet_roblox_bot_player = { version = "=0.6.2-depth2", registry = "strafesnet" }
 strafesnet_snf = { version = "0.4.0", registry = "strafesnet" }
 pollster = "0.4.0"
+glam = "0.32.1"

src/main.rs

@@ -1,7 +1,7 @@
 use burn::backend::Autodiff;
 use burn::nn::loss::{MseLoss, Reduction};
 use burn::nn::{Linear, LinearConfig, Relu};
-use burn::optim::{GradientsParams, Optimizer, AdamConfig};
+use burn::optim::{AdamConfig, GradientsParams, Optimizer};
 use burn::prelude::*;
 
 type InferenceBackend = burn::backend::Cuda<f32>;
@@ -14,10 +14,7 @@ use strafesnet_roblox_bot_file::v0;
 const SIZE_X: usize = 64;
 const SIZE_Y: usize = 36;
 const INPUT: usize = SIZE_X * SIZE_Y;
-const HIDDEN: [usize; 2] = [
-	INPUT >> 3,
-	INPUT >> 7,
-];
+const HIDDEN: [usize; 2] = [INPUT >> 3, INPUT >> 7];
 // MoveForward
 // MoveLeft
 // MoveBack
@@ -149,7 +146,7 @@ fn training() {
 	// training data
 	let training_samples = timelines.input_events.len() - 1;
 
-	let input_size = (size.x * size.y) as usize * size_of::<f32>();
+	let input_size = INPUT * size_of::<f32>();
 	let mut inputs = Vec::with_capacity(input_size * training_samples);
 	let mut targets = Vec::with_capacity(OUTPUT * training_samples);
 
@@ -227,27 +224,22 @@ fn training() {
 				.unwrap(),
 		};
 
-		fn p(v: v0::Vector3) -> [f32; 3] {
-			[v.x, v.y, v.z]
+		fn vec3(v: v0::Vector3) -> glam::Vec3 {
+			glam::vec3(v.x, v.y, v.z)
 		}
-		fn a(a: v0::Vector3) -> [f32; 2] {
-			[a.y, a.x]
-		}
-		fn sub<T: core::ops::Sub>(lhs: T, rhs: T) -> T::Output {
-			lhs - rhs
+		fn angles(a: v0::Vector3) -> glam::Vec2 {
+			glam::vec2(a.y, a.x)
 		}
 
+		let pos = vec3(output_event.event.position) - world_offset;
+		let angles = angles(output_event.event.angles);
+
 		let mut encoder = device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
 			label: Some("wgpu encoder"),
 		});
 
 		// render!
-		graphics.encode_commands(
-			&mut encoder,
-			&graphics_texture_view,
-			sub(p(output_event.event.position).into(), world_offset),
-			a(output_event.event.angles).into(),
-		);
+		graphics.encode_commands(&mut encoder, &graphics_texture_view, pos, angles);
 
 		// copy the depth texture into ram
 		encoder.copy_texture_to_buffer(
@@ -302,6 +294,21 @@ fn training() {
 		texture_data.clear();
 	}
 
+	// normalize inputs
+	let global_min = *inputs
+		.iter()
+		.min_by(|a, b| a.partial_cmp(b).unwrap())
+		.unwrap();
+	let global_max = *inputs
+		.iter()
+		.max_by(|a, b| a.partial_cmp(b).unwrap())
+		.unwrap();
+	let global_range = global_max - global_min;
+	println!("Normalizing to range {global_min} - {global_max}");
+	inputs.iter_mut().for_each(|value| {
+		*value = 1.0 - (*value - global_min) / global_range;
+	});
+
 	let device = burn::backend::cuda::CudaDevice::new(gpu_id);
 
 	let mut model: Net<TrainingBackend> = Net::init(&device);
@@ -310,10 +317,7 @@ fn training() {
 	let mut optim = AdamConfig::new().init();
 
 	let inputs = Tensor::from_data(
-		TensorData::new(
-			inputs,
-			Shape::new([training_samples, (size.x * size.y) as usize]),
-		),
+		TensorData::new(inputs, Shape::new([training_samples, INPUT])),
 		&device,
 	);
 	let targets = Tensor::from_data(
@@ -321,8 +325,8 @@ fn training() {
 		&device,
 	);
 
-	const LEARNING_RATE: f64 = 0.5;
-	const EPOCHS: usize = 10000;
+	const LEARNING_RATE: f64 = 0.001;
+	const EPOCHS: usize = 100000;
 
 	for epoch in 0..EPOCHS {
 		let predictions = model.forward(inputs.clone());