1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352
// This file is part of ICU4X. For terms of use, please see the file
// called LICENSE at the top level of the ICU4X source tree
// (online at: https://github.com/unicode-org/icu4x/blob/main/LICENSE ).
//! Data provider struct definitions for the lstm
// Provider structs must be stable
#![allow(clippy::exhaustive_structs, clippy::exhaustive_enums)]
use icu_provider::prelude::*;
use zerovec::{ule::UnvalidatedStr, ZeroMap, ZeroVec};
// We do this instead of const generics because ZeroFrom and Yokeable derives, as well as serde
// don't support them
macro_rules! lstm_matrix {
($name:ident, $generic:literal) => {
/// The struct that stores a LSTM's matrix.
///
/// <div class="stab unstable">
/// 🚧 This code is considered unstable; it may change at any time, in breaking or non-breaking ways,
/// including in SemVer minor releases. While the serde representation of data structs is guaranteed
/// to be stable, their Rust representation might not be. Use with caution.
/// </div>
#[derive(PartialEq, Debug, Clone, zerofrom::ZeroFrom, yoke::Yokeable)]
#[cfg_attr(feature = "datagen", derive(serde::Serialize))]
pub struct $name<'data> {
// Invariant: dims.product() == data.len()
#[allow(missing_docs)]
pub(crate) dims: [u16; $generic],
#[allow(missing_docs)]
pub(crate) data: ZeroVec<'data, f32>,
}
impl<'data> $name<'data> {
#[cfg(any(feature = "serde", feature = "datagen"))]
/// Creates a LstmMatrix with the given dimensions. Fails if the dimensions don't match the data.
pub fn from_parts(
dims: [u16; $generic],
data: ZeroVec<'data, f32>,
) -> Result<Self, DataError> {
if dims.iter().map(|&i| i as usize).product::<usize>() != data.len() {
Err(DataError::custom("Dimension mismatch"))
} else {
Ok(Self { dims, data })
}
}
#[doc(hidden)] // databake
pub const fn from_parts_unchecked(
dims: [u16; $generic],
data: ZeroVec<'data, f32>,
) -> Self {
Self { dims, data }
}
}
#[cfg(feature = "serde")]
impl<'de: 'data, 'data> serde::Deserialize<'de> for $name<'data> {
fn deserialize<S>(deserializer: S) -> Result<Self, S::Error>
where
S: serde::de::Deserializer<'de>,
{
#[derive(serde::Deserialize)]
struct Raw<'data> {
dims: [u16; $generic],
#[serde(borrow)]
data: ZeroVec<'data, f32>,
}
let raw = Raw::deserialize(deserializer)?;
use serde::de::Error;
Self::from_parts(raw.dims, raw.data)
.map_err(|_| S::Error::custom("Dimension mismatch"))
}
}
#[cfg(feature = "datagen")]
impl databake::Bake for $name<'_> {
fn bake(&self, env: &databake::CrateEnv) -> databake::TokenStream {
let dims = self.dims.bake(env);
let data = self.data.bake(env);
databake::quote! {
icu_segmenter::provider::$name::from_parts_unchecked(#dims, #data)
}
}
}
};
}
lstm_matrix!(LstmMatrix1, 1);
lstm_matrix!(LstmMatrix2, 2);
lstm_matrix!(LstmMatrix3, 3);
#[derive(PartialEq, Debug, Clone, Copy)]
#[cfg_attr(
feature = "datagen",
derive(serde::Serialize,databake::Bake),
databake(path = icu_segmenter::provider),
)]
#[cfg_attr(feature = "serde", derive(serde::Deserialize))]
/// The type of LSTM model
///
/// <div class="stab unstable">
/// 🚧 This code is considered unstable; it may change at any time, in breaking or non-breaking ways,
/// including in SemVer minor releases. While the serde representation of data structs is guaranteed
/// to be stable, their Rust representation might not be. Use with caution.
/// </div>
pub enum ModelType {
/// A model working on code points
Codepoints,
/// A model working on grapheme clusters
GraphemeClusters,
}
/// The struct that stores a LSTM model.
///
/// <div class="stab unstable">
/// 🚧 This code is considered unstable; it may change at any time, in breaking or non-breaking ways,
/// including in SemVer minor releases. While the serde representation of data structs is guaranteed
/// to be stable, their Rust representation might not be. Use with caution.
/// </div>
#[derive(PartialEq, Debug, Clone, yoke::Yokeable, zerofrom::ZeroFrom)]
#[cfg_attr(feature = "datagen", derive(serde::Serialize))]
#[yoke(prove_covariance_manually)]
pub struct LstmDataFloat32<'data> {
/// Type of the model
pub(crate) model: ModelType,
/// The grapheme cluster dictionary used to train the model
pub(crate) dic: ZeroMap<'data, UnvalidatedStr, u16>,
/// The embedding layer. Shape (dic.len + 1, e)
pub(crate) embedding: LstmMatrix2<'data>,
/// The forward layer's first matrix. Shape (h, 4, e)
pub(crate) fw_w: LstmMatrix3<'data>,
/// The forward layer's second matrix. Shape (h, 4, h)
pub(crate) fw_u: LstmMatrix3<'data>,
/// The forward layer's bias. Shape (h, 4)
pub(crate) fw_b: LstmMatrix2<'data>,
/// The backward layer's first matrix. Shape (h, 4, e)
pub(crate) bw_w: LstmMatrix3<'data>,
/// The backward layer's second matrix. Shape (h, 4, h)
pub(crate) bw_u: LstmMatrix3<'data>,
/// The backward layer's bias. Shape (h, 4)
pub(crate) bw_b: LstmMatrix2<'data>,
/// The output layer's weights. Shape (2, 4, h)
pub(crate) time_w: LstmMatrix3<'data>,
/// The output layer's bias. Shape (4)
pub(crate) time_b: LstmMatrix1<'data>,
}
impl<'data> LstmDataFloat32<'data> {
#[doc(hidden)] // databake
#[allow(clippy::too_many_arguments)] // constructor
pub const fn from_parts_unchecked(
model: ModelType,
dic: ZeroMap<'data, UnvalidatedStr, u16>,
embedding: LstmMatrix2<'data>,
fw_w: LstmMatrix3<'data>,
fw_u: LstmMatrix3<'data>,
fw_b: LstmMatrix2<'data>,
bw_w: LstmMatrix3<'data>,
bw_u: LstmMatrix3<'data>,
bw_b: LstmMatrix2<'data>,
time_w: LstmMatrix3<'data>,
time_b: LstmMatrix1<'data>,
) -> Self {
Self {
model,
dic,
embedding,
fw_w,
fw_u,
fw_b,
bw_w,
bw_u,
bw_b,
time_w,
time_b,
}
}
#[cfg(any(feature = "serde", feature = "datagen"))]
/// Creates a LstmDataFloat32 with the given data. Fails if the matrix dimensions are inconsistent.
#[allow(clippy::too_many_arguments)] // constructor
pub fn try_from_parts(
model: ModelType,
dic: ZeroMap<'data, UnvalidatedStr, u16>,
embedding: LstmMatrix2<'data>,
fw_w: LstmMatrix3<'data>,
fw_u: LstmMatrix3<'data>,
fw_b: LstmMatrix2<'data>,
bw_w: LstmMatrix3<'data>,
bw_u: LstmMatrix3<'data>,
bw_b: LstmMatrix2<'data>,
time_w: LstmMatrix3<'data>,
time_b: LstmMatrix1<'data>,
) -> Result<Self, DataError> {
let dic_len = u16::try_from(dic.len())
.map_err(|_| DataError::custom("Dictionary does not fit in u16"))?;
let num_classes = embedding.dims[0];
let embedd_dim = embedding.dims[1];
let hunits = fw_u.dims[2];
if num_classes - 1 != dic_len
|| fw_w.dims != [4, hunits, embedd_dim]
|| fw_u.dims != [4, hunits, hunits]
|| fw_b.dims != [4, hunits]
|| bw_w.dims != [4, hunits, embedd_dim]
|| bw_u.dims != [4, hunits, hunits]
|| bw_b.dims != [4, hunits]
|| time_w.dims != [2, 4, hunits]
|| time_b.dims != [4]
{
return Err(DataError::custom("LSTM dimension mismatch"));
}
#[cfg(debug_assertions)]
if !dic.iter_copied_values().all(|(_, g)| g < dic_len) {
return Err(DataError::custom("Invalid cluster id"));
}
Ok(Self {
model,
dic,
embedding,
fw_w,
fw_u,
fw_b,
bw_w,
bw_u,
bw_b,
time_w,
time_b,
})
}
}
#[cfg(feature = "serde")]
impl<'de: 'data, 'data> serde::Deserialize<'de> for LstmDataFloat32<'data> {
fn deserialize<S>(deserializer: S) -> Result<Self, S::Error>
where
S: serde::de::Deserializer<'de>,
{
#[derive(serde::Deserialize)]
struct Raw<'data> {
model: ModelType,
#[cfg_attr(feature = "serde", serde(borrow))]
dic: ZeroMap<'data, UnvalidatedStr, u16>,
#[cfg_attr(feature = "serde", serde(borrow))]
embedding: LstmMatrix2<'data>,
#[cfg_attr(feature = "serde", serde(borrow))]
fw_w: LstmMatrix3<'data>,
#[cfg_attr(feature = "serde", serde(borrow))]
fw_u: LstmMatrix3<'data>,
#[cfg_attr(feature = "serde", serde(borrow))]
fw_b: LstmMatrix2<'data>,
#[cfg_attr(feature = "serde", serde(borrow))]
bw_w: LstmMatrix3<'data>,
#[cfg_attr(feature = "serde", serde(borrow))]
bw_u: LstmMatrix3<'data>,
#[cfg_attr(feature = "serde", serde(borrow))]
bw_b: LstmMatrix2<'data>,
#[cfg_attr(feature = "serde", serde(borrow))]
time_w: LstmMatrix3<'data>,
#[cfg_attr(feature = "serde", serde(borrow))]
time_b: LstmMatrix1<'data>,
}
let raw = Raw::deserialize(deserializer)?;
use serde::de::Error;
Self::try_from_parts(
raw.model,
raw.dic,
raw.embedding,
raw.fw_w,
raw.fw_u,
raw.fw_b,
raw.bw_w,
raw.bw_u,
raw.bw_b,
raw.time_w,
raw.time_b,
)
.map_err(|_| S::Error::custom("Invalid dimensions"))
}
}
#[cfg(feature = "datagen")]
impl databake::Bake for LstmDataFloat32<'_> {
fn bake(&self, env: &databake::CrateEnv) -> databake::TokenStream {
let model = self.model.bake(env);
let dic = self.dic.bake(env);
let embedding = self.embedding.bake(env);
let fw_w = self.fw_w.bake(env);
let fw_u = self.fw_u.bake(env);
let fw_b = self.fw_b.bake(env);
let bw_w = self.bw_w.bake(env);
let bw_u = self.bw_u.bake(env);
let bw_b = self.bw_b.bake(env);
let time_w = self.time_w.bake(env);
let time_b = self.time_b.bake(env);
databake::quote! {
icu_segmenter::provider::LstmDataFloat32::from_parts_unchecked(
#model,
#dic,
#embedding,
#fw_w,
#fw_u,
#fw_b,
#bw_w,
#bw_u,
#bw_b,
#time_w,
#time_b,
)
}
}
}
/// The data to power the LSTM segmentation model.
///
/// This data enum is extensible: more backends may be added in the future.
/// Old data can be used with newer code but not vice versa.
///
/// Examples of possible future extensions:
///
/// 1. Variant to store data in 16 instead of 32 bits
/// 2. Minor changes to the LSTM model, such as different forward/backward matrix sizes
///
/// <div class="stab unstable">
/// 🚧 This code is considered unstable; it may change at any time, in breaking or non-breaking ways,
/// including in SemVer minor releases. While the serde representation of data structs is guaranteed
/// to be stable, their Rust representation might not be. Use with caution.
/// </div>
#[icu_provider::data_struct(LstmForWordLineAutoV1Marker = "segmenter/lstm/wl_auto@1")]
#[derive(Debug, PartialEq, Clone)]
#[cfg_attr(
feature = "datagen",
derive(serde::Serialize, databake::Bake),
databake(path = icu_segmenter::provider),
)]
#[cfg_attr(feature = "serde", derive(serde::Deserialize))]
#[yoke(prove_covariance_manually)]
#[non_exhaustive]
pub enum LstmDataV1<'data> {
/// The data as matrices of zerovec f32 values.
Float32(#[cfg_attr(feature = "serde", serde(borrow))] LstmDataFloat32<'data>),
// new variants should go BELOW existing ones
// Serde serializes based on variant name and index in the enum
// https://docs.rs/serde/latest/serde/trait.Serializer.html#tymethod.serialize_unit_variant
}