Struct SnakeCNNInference 

pub struct SnakeCNNInference {

    pub grid_width: usize,
    pub grid_height: usize,
    pub input_channels: usize,
    pub num_actions: usize,
    pub conv1_weight: Vec<Vec<Vec<Vec<f32>>>>,
    pub conv1_bias: Vec<f32>,
    pub conv2_weight: Vec<Vec<Vec<Vec<f32>>>>,
    pub conv2_bias: Vec<f32>,
    pub conv3_weight: Vec<Vec<Vec<Vec<f32>>>>,
    pub conv3_bias: Vec<f32>,
    pub fc_common_weight: Vec<Vec<f32>>,
    pub fc_common_bias: Vec<f32>,
    pub fc_policy_weight: Vec<Vec<f32>>,
    pub fc_policy_bias: Vec<f32>,
    pub fc_value_weight: Vec<Vec<f32>>,
    pub fc_value_bias: Vec<f32>,
    pub metadata: Option<TrainingMetadata>,
}

A serializable CNN model for Snake inference

This struct contains all the weights and biases needed to run CNN inference in pure Rust (no Burn/tensor-stack dependency).

Fields

grid_width: usize

Grid width

grid_height: usize

Grid height

input_channels: usize

Number of input channels (should be 5 for Snake)

num_actions: usize

Number of actions (should be 4 for Snake)

conv1_weight: Vec<Vec<Vec<Vec<f32>>>>

Conv1 kernel weights, shape [out=32, in=input_channels, kh=3, kw=3]. Applied first in SnakeCNNInference::forward with padding=1.

conv1_bias: Vec<f32>

Conv1 per-output-channel bias, shape [32].

conv2_weight: Vec<Vec<Vec<Vec<f32>>>>

Conv2 kernel weights, shape [out=64, in=32, kh=3, kw=3]. Applied after Conv1 + ReLU with padding=1.

conv2_bias: Vec<f32>

Conv2 per-output-channel bias, shape [64].

conv3_weight: Vec<Vec<Vec<Vec<f32>>>>

Conv3 kernel weights, shape [out=64, in=64, kh=3, kw=3]. Applied after Conv2 + ReLU with padding=1.

conv3_bias: Vec<f32>

Conv3 per-output-channel bias, shape [64].

fc_common_weight: Vec<Vec<f32>>

Shared fully-connected weights, shape [256, 64 * grid_width * grid_height], applied to the flattened Conv3 output.

fc_common_bias: Vec<f32>

Shared fully-connected bias, shape [256].

fc_policy_weight: Vec<Vec<f32>>

Policy head weights, shape [num_actions, 256]. Output is raw logits (no softmax applied here — callers do that in SnakeCNNInference::get_action or externally).

fc_policy_bias: Vec<f32>

Policy head bias, shape [num_actions].

fc_value_weight: Vec<Vec<f32>>

Value head weights, shape [1, 256]. Produces a scalar state-value estimate (no activation).

fc_value_bias: Vec<f32>

Value head bias, shape [1].

metadata: Option<TrainingMetadata>

Optional training metadata

Implementations

impl SnakeCNNInference

pub fn forward(&self, grid: &[f32]) -> (Vec<f32>, f32)

Forward pass: compute action logits and value

Arguments

• grid - Input grid [channels, height, width] flattened as [c0_pixels…, c1_pixels…, …]

Returns

• (logits, value) - Action logits [num_actions] and state value (scalar)

pub fn get_action(&self, grid: &[f32]) -> usize

Get action from grid using argmax (deterministic)

pub fn save_json(&self, path: &str) -> Result<()>

Save model to JSON file

pub fn load_json(path: &str) -> Result<Self>

Load model from JSON file

Trait Implementations

impl Clone for SnakeCNNInference

fn clone(&self) -> SnakeCNNInference

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for SnakeCNNInference

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for SnakeCNNInference

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for SnakeCNNInference

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.