use std::collections::HashMap;
use burn::prelude::*;
use half::bf16;
use safetensors::SafeTensors;
use crate::model::eegpt::EEGPT;
use crate::config::ModelConfig;
pub struct WeightMap {
pub tensors: HashMap<String, (Vec<f32>, Vec<usize>)>,
}
impl WeightMap {
pub fn from_file(path: &str) -> anyhow::Result<Self> {
let bytes = std::fs::read(path)?;
let st = SafeTensors::deserialize(&bytes)?;
let mut tensors = HashMap::with_capacity(st.len());
for (key, view) in st.tensors() {
let key = key.strip_prefix("model.").unwrap_or(&key).to_string();
let shape: Vec<usize> = view.shape().to_vec();
let data = view.data();
let f32s: Vec<f32> = match view.dtype() {
safetensors::Dtype::BF16 => data.chunks_exact(2).map(|b| bf16::from_le_bytes([b[0], b[1]]).to_f32()).collect(),
safetensors::Dtype::F32 => data.chunks_exact(4).map(|b| f32::from_le_bytes([b[0], b[1], b[2], b[3]])).collect(),
safetensors::Dtype::F16 => data.chunks_exact(2).map(|b| half::f16::from_le_bytes([b[0], b[1]]).to_f32()).collect(),
other => anyhow::bail!("unsupported dtype {:?} for key {key}", other),
};
tensors.insert(key, (f32s, shape));
}
Ok(Self { tensors })
}
pub fn take<B: Backend, const N: usize>(&mut self, key: &str, device: &B::Device) -> anyhow::Result<Tensor<B, N>> {
let (data, shape) = self.tensors.remove(key).ok_or_else(|| anyhow::anyhow!("key not found: {key}"))?;
if shape.len() != N { anyhow::bail!("rank mismatch for {key}: expected {N}, got {}", shape.len()); }
Ok(Tensor::<B, N>::from_data(TensorData::new(data, shape), device))
}
pub fn has(&self, key: &str) -> bool { self.tensors.contains_key(key) }
}
fn set_linear_wb<B: Backend>(l: &mut burn::nn::Linear<B>, w: Tensor<B, 2>, b: Tensor<B, 1>) {
l.weight = l.weight.clone().map(|_| w.transpose());
if let Some(ref bias) = l.bias { l.bias = Some(bias.clone().map(|_| b)); }
}
fn set_layernorm<B: Backend>(n: &mut burn::nn::LayerNorm<B>, w: Tensor<B, 1>, b: Tensor<B, 1>) {
n.gamma = n.gamma.clone().map(|_| w);
if let Some(ref beta) = n.beta { n.beta = Some(beta.clone().map(|_| b)); }
}
fn set_conv2d_wb<B: Backend>(c: &mut burn::nn::conv::Conv2d<B>, w: Tensor<B, 4>, b: Tensor<B, 1>) {
c.weight = c.weight.clone().map(|_| w);
if let Some(ref bias) = c.bias { c.bias = Some(bias.clone().map(|_| b)); }
}
pub fn load_model<B: Backend>(cfg: &ModelConfig, path: &str, device: &B::Device) -> anyhow::Result<EEGPT<B>> {
let mut wm = WeightMap::from_file(path)?;
eprintln!("Loading {} weight tensors...", wm.tensors.len());
load_model_from_wm(cfg, &mut wm, device)
}
pub fn load_model_from_wm<B: Backend>(cfg: &ModelConfig, wm: &mut WeightMap, device: &B::Device) -> anyhow::Result<EEGPT<B>> {
let mut model = EEGPT::new(
cfg.n_outputs, cfg.n_chans, cfg.n_times,
cfg.patch_size, cfg.patch_stride, cfg.embed_num, cfg.embed_dim,
cfg.depth, cfg.num_heads, cfg.mlp_ratio, cfg.qkv_bias,
cfg.n_chan_embeddings, cfg.probe_hidden_dim, 1e-6, device,
);
load_weights(wm, &mut model, cfg, device)?;
Ok(model)
}
fn load_weights<B: Backend>(wm: &mut WeightMap, model: &mut EEGPT<B>, cfg: &ModelConfig, device: &B::Device) -> anyhow::Result<()> {
let te = &mut model.target_encoder;
if let Ok(t) = wm.take::<B, 3>("target_encoder.summary_token", device) {
te.summary_token = te.summary_token.clone().map(|_| t);
}
if let (Ok(w), Ok(b)) = (wm.take::<B,4>("target_encoder.patch_embed.proj.weight", device),
wm.take::<B,1>("target_encoder.patch_embed.proj.bias", device)) {
set_conv2d_wb(&mut te.patch_embed.proj, w, b);
}
if let Ok(w) = wm.take::<B, 2>("target_encoder.chan_embed.weight", device) {
te.chan_embed.weight = te.chan_embed.weight.clone().map(|_| w);
}
for i in 0..cfg.depth {
let block = &mut te.blocks[i];
let p = format!("target_encoder.blocks.{i}");
if let (Ok(w), Ok(b)) = (wm.take::<B,1>(&format!("{p}.norm1.weight"), device),
wm.take::<B,1>(&format!("{p}.norm1.bias"), device)) {
set_layernorm(&mut block.norm1, w, b);
}
if let (Ok(w), Ok(b)) = (wm.take::<B,2>(&format!("{p}.attn.qkv.weight"), device),
wm.take::<B,1>(&format!("{p}.attn.qkv.bias"), device)) {
set_linear_wb(&mut block.attn.qkv, w, b);
}
if let (Ok(w), Ok(b)) = (wm.take::<B,2>(&format!("{p}.attn.proj.weight"), device),
wm.take::<B,1>(&format!("{p}.attn.proj.bias"), device)) {
set_linear_wb(&mut block.attn.proj, w, b);
}
if let (Ok(w), Ok(b)) = (wm.take::<B,1>(&format!("{p}.norm2.weight"), device),
wm.take::<B,1>(&format!("{p}.norm2.bias"), device)) {
set_layernorm(&mut block.norm2, w, b);
}
if let (Ok(w), Ok(b)) = (wm.take::<B,2>(&format!("{p}.mlp.fc1.weight"), device),
wm.take::<B,1>(&format!("{p}.mlp.fc1.bias"), device)) {
set_linear_wb(&mut block.mlp_fc1, w, b);
}
if let (Ok(w), Ok(b)) = (wm.take::<B,2>(&format!("{p}.mlp.fc2.weight"), device),
wm.take::<B,1>(&format!("{p}.mlp.fc2.bias"), device)) {
set_linear_wb(&mut block.mlp_fc2, w, b);
}
}
if let (Ok(w), Ok(b)) = (wm.take::<B,1>("target_encoder.norm.weight", device),
wm.take::<B,1>("target_encoder.norm.bias", device)) {
set_layernorm(&mut te.norm, w, b);
}
if let (Ok(w), Ok(b)) = (wm.take::<B,2>("final_layer.probe1.weight", device),
wm.take::<B,1>("final_layer.probe1.bias", device)) {
set_linear_wb(&mut model.probe1, w, b);
}
if let (Ok(w), Ok(b)) = (wm.take::<B,2>("final_layer.probe2.weight", device),
wm.take::<B,1>("final_layer.probe2.bias", device)) {
set_linear_wb(&mut model.probe2, w, b);
}
Ok(())
}