SGD and LSTM

Cargo.toml

@@ -30,17 +30,20 @@ parquet = ["amadeus-parquet", "amadeus-derive/parquet"]
 postgres = ["amadeus-postgres", "amadeus-derive/postgres"]
 csv = ["amadeus-serde", "amadeus-derive/serde"]
 json = ["amadeus-serde", "amadeus-derive/serde"]
+ml = ["amadeus-ml"]
+ml-openblas = ["amadeus-ml/openblas"]
 bench = ["serde-csv", "once_cell", "arrow-parquet", "rayon"]
 
 [package.metadata.docs.rs]
-features = ["constellation", "aws", "commoncrawl", "parquet", "postgres", "csv", "json"]
+features = ["constellation", "aws", "commoncrawl", "parquet", "postgres", "csv", "json", "ml"]
 
 [dependencies]
 amadeus-core = { version = "=0.4.1", path = "amadeus-core" }
 amadeus-derive = { version = "=0.4.1", path = "amadeus-derive" }
 amadeus-types = { version = "=0.4.1", path = "amadeus-types" }
 amadeus-aws = { version = "=0.4.1", path = "amadeus-aws", optional = true }
 amadeus-commoncrawl = { version = "=0.4.1", path = "amadeus-commoncrawl", optional = true }
+amadeus-ml = { version = "=0.4.1", path = "amadeus-ml", optional = true }
 amadeus-parquet = { version = "=0.4.1", path = "amadeus-parquet", optional = true }
 amadeus-postgres = { version = "=0.4.1", path = "amadeus-postgres", optional = true }
 amadeus-serde = { version = "=0.4.1", path = "amadeus-serde", optional = true }
@@ -49,7 +52,7 @@ async-channel = "1.1"
 bincode = { version = "1.3", optional = true }
 constellation-rs = { version = "0.2.0-alpha.2", default-features = false, optional = true }
 derive-new = "0.5"
-event-listener = "=2.3.1" # https://github.com/stjepang/event-listener/issues/9
+event-listener = "2.3.3"
 futures = "0.3"
 num_cpus = "1.13"
 pin-project = "0.4"

amadeus-ml/Cargo.toml

@@ -0,0 +1,46 @@
+[package]
+name = "amadeus-ml"
+version = "0.4.1"
+license = "Apache-2.0"
+authors = ["Alec Mocatta <alec@mocatta.net>", "Maciej Kula <maciej.kula@gmail.com>"]
+categories = ["data-structures", "algorithms", "science"]
+keywords = ["streaming-algorithm", "probabilistic", "sketch", "data-structure", "hyperloglog"]
+description = """
+A low-overhead, define-by-run autodifferentiation library.
+"""
+repository = "https://github.com/constellation-rs/amadeus"
+homepage = "https://github.com/constellation-rs/amadeus"
+documentation = "https://docs.rs/amadeus"
+readme = "README.md"
+edition = "2018"
+
+[badges]
+azure-devops = { project = "alecmocatta/amadeus", pipeline = "tests", build = "26" }
+maintenance = { status = "actively-developed" }
+
+[features]
+fast-math = []
+openblas = ["ndarray/blas", "blas-src/openblas", "openblas-src/static"]
+
+[dependencies]
+approx = "0.3"
+blas-src = { version = "0.2", optional = true }
+hibitset = "0.6"
+itertools = "0.9"
+ndarray = { version = "0.13", features = ["approx", "serde"] }
+openblas-src = { version = "0.6", default-features = false, optional = true }
+rand = { version = "0.7", features = ["serde1"] }
+rand_distr = "0.2"
+serde = { version = "1", features = ["derive", "rc"] }
+smallvec = { version = "1.4", features = ["serde"] }
+
+[build-dependencies]
+rustversion = "1.0"
+
+[dev-dependencies]
+criterion = "0.3"
+rayon = "1.3"
+
+[[bench]]
+name = "benchmark"
+harness = false

amadeus-ml/NOTICE.txt

@@ -0,0 +1,10 @@
+This product includes software from the wyrm project (MIT)
+https://github.com/maciejkula/wyrm
+
+Copyright 2017 Maciej Kula
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

amadeus-ml/benches/benchmark.rs

@@ -0,0 +1,241 @@
+#![allow(clippy::many_single_char_names)]
+
+use criterion::{criterion_group, criterion_main, Criterion};
+use rayon::prelude::*;
+use std::sync::Arc;
+
+use amadeus_ml::{
+	nn::{lstm, xavier_normal}, optim::{Optimizer, SGD}, DataInput, HogwildParameter, ParameterNode
+};
+
+fn bench_node_reuse(c: &mut Criterion) {
+	c.bench_function("node_reuse", |b| {
+		let dim = 128;
+
+		let x = ParameterNode::new(xavier_normal(1, dim));
+		let y = ParameterNode::new(xavier_normal(dim, 10));
+		let v = x.dot(&y);
+		let z = v.clone() + v.clone() + v.clone() + v;
+
+		b.iter(|| {
+			z.forward();
+			z.zero_gradient();
+		})
+	});
+}
+
+fn bench_matrix_multiply(c: &mut Criterion) {
+	c.bench_function("bench_matrix_multiply", |b| {
+		let dim = 64;
+		let num_epochs = 20;
+
+		let x_data = Arc::new(HogwildParameter::new(xavier_normal(1, dim)));
+		let y_data = Arc::new(HogwildParameter::new(xavier_normal(dim, 10)));
+
+		b.iter(|| {
+			(0..rayon::current_num_threads())
+				.into_par_iter()
+				.for_each(|_| {
+					let x = ParameterNode::shared(x_data.clone());
+					let y = ParameterNode::shared(y_data.clone());
+
+					let v = x.dot(&y);
+
+					for _ in 0..num_epochs {
+						v.forward();
+						v.zero_gradient();
+					}
+				});
+		})
+	});
+}
+
+// fn bench_sofmax_exp_sum(b: &mut Criterion) {
+// 	c.bench_function("bench_softmax_exp_sum", |b| {
+// 		let x = vec![1.0; 32];
+// 		let max = 1.0;
+
+// 		b.iter(|| x.iter().map(|&x| amadeus_ml::exp(x - max)).sum::<f32>().ln())
+// 	})
+// }
+
+// #[bench]
+// fn bench_sofmax_exp_sum_unrolled(b: &mut Criterion) {
+// 	let x = vec![1.0; 32];
+// 	let max = 1.0;
+
+// 	b.iter(|| softmax_exp_sum(&x, max).ln())
+// }
+
+// fn bench_exp(b: &mut Criterion) {
+// 	let x: Vec<f32> = vec![1.0; 32];
+
+// 	let mut v = 0.0;
+
+// 	b.iter(|| x.iter().for_each(|&y| v += y.exp()));
+// }
+
+// fn bench_fastexp(b: &mut Criterion) {
+// 	let x: Vec<f32> = vec![1.0; 32];
+
+// 	let mut v = 0.0;
+
+// 	b.iter(|| x.iter().for_each(|&y| v += fastexp(y)));
+// }
+
+// fn bench_dot(b: &mut Criterion) {
+// 	let xs = vec![0.0; 256];
+// 	let ys = vec![0.0; 256];
+
+// 	b.iter(|| dot(&xs[..], &ys[..]));
+// }
+
+// fn bench_unrolled_dot(b: &mut Criterion) {
+// 	let xs = vec![0.0; 256];
+// 	let ys = vec![0.0; 256];
+
+// 	b.iter(|| unrolled_dot(&xs[..], &ys[..]));
+// }
+
+// fn bench_simd_dot(b: &mut Criterion) {
+// 	let xs = vec![0.0; 256];
+// 	let ys = vec![0.0; 256];
+
+// 	b.iter(|| simd_dot(&xs[..], &ys[..]));
+// }
+
+// fn bench_array_scaled_assign(b: &mut Criterion) {
+// 	let mut xs = random_matrix(256, 1);
+// 	let ys = random_matrix(256, 1);
+
+// 	b.iter(|| array_scaled_assign(&mut xs, &ys, 3.5));
+// }
+
+// fn bench_slice_scaled_assign(b: &mut Criterion) {
+// 	let mut xs = random_matrix(256, 1);
+// 	let ys = random_matrix(256, 1);
+
+// 	b.iter(|| scaled_assign(&mut xs, &ys, 3.5));
+// }
+
+// fn bench_array_assign(b: &mut Criterion) {
+// 	let mut xs = random_matrix(256, 1);
+// 	let ys = random_matrix(256, 1);
+
+// 	b.iter(|| array_assign(&mut xs, &ys));
+// }
+
+// fn bench_slice_assign(b: &mut Criterion) {
+// 	let mut xs = random_matrix(256, 1);
+// 	let ys = random_matrix(256, 1);
+
+// 	b.iter(|| assign(&mut xs, &ys));
+// }
+
+// fn dot_node_specializations_mm(b: &mut Criterion) {
+// 	let x = random_matrix(64, 64);
+// 	let y = random_matrix(64, 64);
+// 	let mut z = random_matrix(64, 64);
+
+// 	b.iter(|| mat_mul(1.0, &x, &y, 0.0, &mut z));
+// }
+
+// fn dot_node_general_vm(b: &mut Criterion) {
+// 	let x = random_matrix(1, 64);
+// 	let y = random_matrix(64, 64);
+// 	let mut z = random_matrix(1, 64);
+
+// 	b.iter(|| general_mat_mul(1.0, &x, &y, 0.0, &mut z));
+// }
+
+// fn dot_node_specializations_vm(b: &mut Criterion) {
+// 	let x = random_matrix(1, 64);
+// 	let y = random_matrix(64, 64);
+// 	let mut z = random_matrix(1, 64);
+
+// 	b.iter(|| mat_mul(1.0, &x, &y, 0.0, &mut z));
+// }
+
+// fn dot_node_specializations_mv(b: &mut Criterion) {
+// 	let x = random_matrix(64, 64);
+// 	let y = random_matrix(64, 1);
+// 	let mut z = random_matrix(64, 1);
+
+// 	b.iter(|| mat_mul(1.0, &x, &y, 0.0, &mut z));
+// }
+
+// fn dot_node_general_mv(b: &mut Criterion) {
+// 	let x = random_matrix(64, 64);
+// 	let y = random_matrix(64, 1);
+// 	let mut z = random_matrix(64, 1);
+
+// 	b.iter(|| general_mat_mul(1.0, &x, &y, 0.0, &mut z));
+// }
+
+fn pi_digits(num: usize) -> Vec<usize> {
+	let pi_str = include_str!("../src/nn/pi.txt");
+	pi_str
+		.chars()
+		.filter_map(|x| x.to_digit(10))
+		.map(|x| x as usize)
+		.take(num)
+		.collect()
+}
+
+fn bench_lstm(c: &mut Criterion) {
+	c.bench_function("bench_lstm", |b| {
+		let sequence_length = 4;
+		let num_digits = 10;
+		let input_dim = 16;
+		let hidden_dim = 32;
+
+		let lstm_params = lstm::Parameters::new(input_dim, hidden_dim, &mut rand::thread_rng());
+		let lstm = lstm_params.build();
+
+		let final_layer = amadeus_ml::ParameterNode::new(xavier_normal(hidden_dim, num_digits));
+		let embeddings = amadeus_ml::ParameterNode::new(xavier_normal(num_digits, input_dim));
+		let y = amadeus_ml::IndexInputNode::new(&[0]);
+
+		let inputs: Vec<_> = (0..sequence_length)
+			.map(|_| amadeus_ml::IndexInputNode::new(&[0]))
+			.collect();
+		let embeddings: Vec<_> = inputs
+			.iter()
+			.map(|input| embeddings.index(&input))
+			.collect();
+
+		let hidden_states = lstm.forward(&embeddings);
+		let hidden = hidden_states.last().unwrap();
+
+		let prediction = hidden.dot(&final_layer);
+		let mut loss = amadeus_ml::nn::losses::sparse_categorical_crossentropy(&prediction, &y);
+		let optimizer = SGD::new();
+
+		let digits = pi_digits(100);
+
+		b.iter(|| {
+			for i in 0..(digits.len() - sequence_length - 1) {
+				let digit_chunk = &digits[i..(i + sequence_length + 1)];
+				if digit_chunk.len() < sequence_length + 1 {
+					break;
+				}
+
+				for (&digit, input) in digit_chunk[..digit_chunk.len() - 1].iter().zip(&inputs) {
+					input.set_value(digit);
+				}
+
+				let target_digit = *digit_chunk.last().unwrap();
+				y.set_value(target_digit);
+
+				loss.forward();
+				loss.backward(1.0);
+
+				optimizer.step(loss.parameters());
+				loss.zero_gradient();
+			}
+		})
+	});
+}
+
+criterion_group!(benches, bench_node_reuse, bench_matrix_multiply, bench_lstm);
+criterion_main!(benches);