|
| 1 | +use std::{ |
| 2 | + collections::{BTreeMap, BTreeSet}, |
| 3 | + fmt::Display, |
| 4 | +}; |
| 5 | + |
| 6 | +use thiserror::Error; |
| 7 | + |
| 8 | +#[derive(Debug, Clone, Copy, PartialEq, PartialOrd, Eq, Ord)] |
| 9 | +pub struct Subscript(char); |
| 10 | + |
| 11 | +#[derive(Debug, Clone)] |
| 12 | +pub enum Subscripts { |
| 13 | + Indexes(Vec<Subscript>), |
| 14 | + Ellipsis { |
| 15 | + start: Vec<Subscript>, |
| 16 | + end: Vec<Subscript>, |
| 17 | + }, |
| 18 | +} |
| 19 | + |
| 20 | +/// Represents an Einstein summation expression following the notation described [https://onnx.ai/onnx/operators/onnx__Einsum.html](here). |
| 21 | +#[derive(Debug, Clone)] |
| 22 | +pub struct Einsum { |
| 23 | + inputs: Vec<Subscripts>, |
| 24 | + output: Option<Subscripts>, |
| 25 | +} |
| 26 | + |
| 27 | +#[derive(Error, Debug)] |
| 28 | +pub enum EinsumError { |
| 29 | + #[error("invalid character encountered: {0}")] |
| 30 | + InvalidCharacter(char), |
| 31 | + |
| 32 | + #[error("the formula has no inputs")] |
| 33 | + MissingInputs, |
| 34 | +} |
| 35 | + |
| 36 | +impl Subscript { |
| 37 | + pub fn from(c: char) -> Subscript { |
| 38 | + assert!(c.is_alphabetic()); |
| 39 | + Subscript(c) |
| 40 | + } |
| 41 | +} |
| 42 | + |
| 43 | +fn count_indices(inputs: &[Subscripts]) -> BTreeMap<Subscript, u32> { |
| 44 | + let mut count = BTreeMap::new(); |
| 45 | + for input in inputs { |
| 46 | + for c in input.subscripts() { |
| 47 | + count.entry(c).and_modify(|n| *n += 1).or_insert(1); |
| 48 | + } |
| 49 | + } |
| 50 | + count |
| 51 | +} |
| 52 | + |
| 53 | +impl Subscripts { |
| 54 | + fn push(&mut self, index: Subscript) { |
| 55 | + match self { |
| 56 | + Subscripts::Indexes(idxs) => idxs.push(index), |
| 57 | + Subscripts::Ellipsis { end, .. } => { |
| 58 | + end.push(index); |
| 59 | + } |
| 60 | + } |
| 61 | + } |
| 62 | + |
| 63 | + fn is_empty(&self) -> bool { |
| 64 | + match self { |
| 65 | + Subscripts::Indexes(idx) => idx.is_empty(), |
| 66 | + Subscripts::Ellipsis { start, end } => start.is_empty() && end.is_empty(), |
| 67 | + } |
| 68 | + } |
| 69 | + |
| 70 | + fn subscripts(&self) -> Vec<Subscript> { |
| 71 | + match &self { |
| 72 | + Subscripts::Indexes(indices) => indices.clone(), |
| 73 | + Subscripts::Ellipsis { start, end } => { |
| 74 | + start.iter().chain(end.iter()).cloned().collect() |
| 75 | + } |
| 76 | + } |
| 77 | + } |
| 78 | +} |
| 79 | + |
| 80 | +impl Display for Subscript { |
| 81 | + fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 82 | + write!(f, "{}", self.0) |
| 83 | + } |
| 84 | +} |
| 85 | + |
| 86 | +impl Display for Subscripts { |
| 87 | + fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 88 | + match self { |
| 89 | + Subscripts::Indexes(idxs) => { |
| 90 | + for i in idxs { |
| 91 | + write!(f, "{}", i)?; |
| 92 | + } |
| 93 | + Ok(()) |
| 94 | + } |
| 95 | + Subscripts::Ellipsis { start, end } => { |
| 96 | + for i in start { |
| 97 | + write!(f, "{}", i)?; |
| 98 | + } |
| 99 | + write!(f, "...")?; |
| 100 | + for i in end { |
| 101 | + write!(f, "{}", i)?; |
| 102 | + } |
| 103 | + Ok(()) |
| 104 | + } |
| 105 | + } |
| 106 | + } |
| 107 | +} |
| 108 | + |
| 109 | +impl Einsum { |
| 110 | + #[allow(dead_code)] |
| 111 | + pub fn from(str: &str) -> Result<Einsum, EinsumError> { |
| 112 | + let mut sum = Einsum { |
| 113 | + inputs: vec![], |
| 114 | + output: None, |
| 115 | + }; |
| 116 | + |
| 117 | + // Parse up to arrow |
| 118 | + let mut chars = str.chars(); |
| 119 | + let mut current_subscripts = Subscripts::Indexes(vec![]); |
| 120 | + let mut after_arrow = false; |
| 121 | + while let Some(character) = &chars.next() { |
| 122 | + match character { |
| 123 | + '-' if chars.next() == Some('>') => { |
| 124 | + // Arrow: switch from inputs to outputs |
| 125 | + if !current_subscripts.is_empty() { |
| 126 | + sum.inputs.push(current_subscripts); |
| 127 | + current_subscripts = Subscripts::Indexes(vec![]); |
| 128 | + } |
| 129 | + if sum.inputs.is_empty() { |
| 130 | + return Err(EinsumError::MissingInputs); |
| 131 | + } |
| 132 | + after_arrow = true; |
| 133 | + } |
| 134 | + '.' if chars.next() == Some('.') && chars.next() == Some('.') => { |
| 135 | + // Ellipsis |
| 136 | + current_subscripts = match current_subscripts { |
| 137 | + Subscripts::Indexes(idxs) => Subscripts::Ellipsis { |
| 138 | + start: idxs, |
| 139 | + end: vec![], |
| 140 | + }, |
| 141 | + Subscripts::Ellipsis { .. } => { |
| 142 | + return Err(EinsumError::InvalidCharacter('.')) |
| 143 | + } |
| 144 | + } |
| 145 | + } |
| 146 | + ' ' => {} |
| 147 | + ',' if !after_arrow => { |
| 148 | + // Next input (cannot occur in output) |
| 149 | + sum.inputs.push(current_subscripts); |
| 150 | + current_subscripts = Subscripts::Indexes(vec![]); |
| 151 | + } |
| 152 | + c if c.is_alphabetic() => { |
| 153 | + current_subscripts.push(Subscript::from(*c)); |
| 154 | + } |
| 155 | + _ => return Err(EinsumError::InvalidCharacter(*character)), |
| 156 | + } |
| 157 | + } |
| 158 | + |
| 159 | + // If we still have subscripts, they are either the last input or the output |
| 160 | + if !current_subscripts.is_empty() |
| 161 | + || matches!(current_subscripts, Subscripts::Ellipsis { .. }) && after_arrow |
| 162 | + { |
| 163 | + if after_arrow { |
| 164 | + sum.output = Some(current_subscripts); |
| 165 | + } else { |
| 166 | + sum.inputs.push(current_subscripts); |
| 167 | + } |
| 168 | + } |
| 169 | + |
| 170 | + Ok(sum) |
| 171 | + } |
| 172 | + |
| 173 | + fn output_or_implicit_subscripts(&self) -> Vec<Subscript> { |
| 174 | + match &self.output { |
| 175 | + Some(o) => o.subscripts(), |
| 176 | + None => { |
| 177 | + // In implicit mode output indices are set to the alphabetically sorted sequence of indices |
| 178 | + // appearing exactly once in the equation. |
| 179 | + let counts = count_indices(&self.inputs); |
| 180 | + let mut keys: Vec<Subscript> = counts |
| 181 | + .into_iter() |
| 182 | + .filter_map(|(k, v)| if v == 1 { Some(k) } else { None }) |
| 183 | + .collect(); |
| 184 | + keys.sort(); |
| 185 | + keys |
| 186 | + } |
| 187 | + } |
| 188 | + } |
| 189 | + |
| 190 | + fn contraction_indices(&self) -> Vec<Subscript> { |
| 191 | + let count = count_indices(&self.inputs); |
| 192 | + let mut subscripts: BTreeSet<Subscript> = count |
| 193 | + .into_iter() |
| 194 | + .filter_map(|(key, value)| if value > 1 { Some(key) } else { None }) |
| 195 | + .collect(); |
| 196 | + for c in &self.output_or_implicit_subscripts() { |
| 197 | + subscripts.remove(c); |
| 198 | + } |
| 199 | + subscripts.into_iter().collect() |
| 200 | + } |
| 201 | +} |
| 202 | + |
| 203 | +impl Display for Einsum { |
| 204 | + fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 205 | + write!( |
| 206 | + f, |
| 207 | + "{}", |
| 208 | + self.inputs |
| 209 | + .iter() |
| 210 | + .map(|x| x.to_string()) |
| 211 | + .collect::<Vec<String>>() |
| 212 | + .join(",") |
| 213 | + )?; |
| 214 | + |
| 215 | + if let Some(output) = &self.output { |
| 216 | + write!(f, " -> {}", output)?; |
| 217 | + } |
| 218 | + Ok(()) |
| 219 | + } |
| 220 | +} |
| 221 | + |
| 222 | +#[cfg(test)] |
| 223 | +mod tests { |
| 224 | + use super::{count_indices, Einsum, Subscript}; |
| 225 | + |
| 226 | + pub fn compare_after_reserialize(formula: &str, expected: &str) { |
| 227 | + assert_eq!(Einsum::from(formula).unwrap().to_string(), expected); |
| 228 | + } |
| 229 | + |
| 230 | + pub fn expect_fail(formula: &str) { |
| 231 | + assert!(Einsum::from(formula).is_err()) |
| 232 | + } |
| 233 | + |
| 234 | + #[test] |
| 235 | + pub fn test_parse_einsum() { |
| 236 | + compare_after_reserialize("ij,jk->ik", "ij,jk -> ik"); |
| 237 | + compare_after_reserialize(" i j, j k -> i k", "ij,jk -> ik"); |
| 238 | + compare_after_reserialize(" i j-> i k", "ij -> ik"); |
| 239 | + |
| 240 | + compare_after_reserialize("a ...d,x... z->a ...z", "a...d,x...z -> a...z"); |
| 241 | + compare_after_reserialize(" ...d,x... z->a ...", "...d,x...z -> a..."); |
| 242 | + compare_after_reserialize("a...", "a..."); |
| 243 | + compare_after_reserialize("a ...d,x... z->...", "a...d,x...z -> ..."); |
| 244 | + |
| 245 | + expect_fail("ij- >ik"); |
| 246 | + expect_fail("->ik"); |
| 247 | + expect_fail("a ...d,x... z->a . ..z"); |
| 248 | + expect_fail("a...b...c"); |
| 249 | + expect_fail("a....b...c"); |
| 250 | + expect_fail("a..b...c"); |
| 251 | + } |
| 252 | + |
| 253 | + #[test] |
| 254 | + pub fn test_indices() { |
| 255 | + let es = Einsum::from("ij,jk->ik").unwrap(); |
| 256 | + let out = count_indices(&es.inputs); |
| 257 | + assert_eq!(out.len(), 3); |
| 258 | + assert_eq!(out[&Subscript::from('i')], 1); |
| 259 | + assert_eq!(out[&Subscript::from('j')], 2); |
| 260 | + assert_eq!(out[&Subscript::from('k')], 1); |
| 261 | + |
| 262 | + let es = Einsum::from("i...k,k...m->i...m").unwrap(); |
| 263 | + let out = count_indices(&es.inputs); |
| 264 | + println!("{:?}", out); |
| 265 | + assert_eq!(out.len(), 5); |
| 266 | + assert_eq!(out[&Subscript::from('i')], 1); |
| 267 | + assert_eq!(out[&Subscript::from('j')], 1); |
| 268 | + assert_eq!(out[&Subscript::from('k')], 2); |
| 269 | + assert_eq!(out[&Subscript::from('l')], 1); |
| 270 | + assert_eq!(out[&Subscript::from('m')], 1); |
| 271 | + } |
| 272 | + |
| 273 | + #[test] |
| 274 | + pub fn test_analysis() { |
| 275 | + let es = Einsum::from("ij,jk->ik").unwrap(); |
| 276 | + assert_eq!(es.contraction_indices(), vec![Subscript::from('j')]); |
| 277 | + |
| 278 | + let es = Einsum::from("ij,jk").unwrap(); |
| 279 | + assert_eq!( |
| 280 | + es.output_or_implicit_subscripts(), |
| 281 | + vec![Subscript::from('i'), Subscript::from('k')] |
| 282 | + ); |
| 283 | + assert_eq!(es.contraction_indices(), vec![Subscript::from('j')]); |
| 284 | + |
| 285 | + let transpose = Einsum::from("ba").unwrap(); |
| 286 | + assert_eq!( |
| 287 | + transpose.output_or_implicit_subscripts(), |
| 288 | + vec![Subscript::from('a'), Subscript::from('b')] |
| 289 | + ); |
| 290 | + } |
| 291 | +} |
0 commit comments