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
|
// This should reach 91.5% accuracy.
#[cfg(feature = "mkl")]
extern crate intel_mkl_src;
use clap::{Parser, ValueEnum};
use candle::{DType, Device, Result, Shape, Tensor, Var, D};
use candle_nn::{loss, ops, Init, Linear};
use std::sync::{Arc, Mutex};
const IMAGE_DIM: usize = 784;
const LABELS: usize = 10;
struct TensorData {
tensors: std::collections::HashMap<String, Var>,
pub dtype: DType,
pub device: Device,
}
// A variant of candle_nn::VarBuilder for initializing variables before training.
#[derive(Clone)]
struct VarStore {
data: Arc<Mutex<TensorData>>,
path: Vec<String>,
}
impl VarStore {
fn new(dtype: DType, device: Device) -> Self {
let data = TensorData {
tensors: std::collections::HashMap::new(),
dtype,
device,
};
Self {
data: Arc::new(Mutex::new(data)),
path: vec![],
}
}
fn pp(&self, s: &str) -> Self {
let mut path = self.path.clone();
path.push(s.to_string());
Self {
data: self.data.clone(),
path,
}
}
fn get<S: Into<Shape>>(&self, shape: S, tensor_name: &str, init: Init) -> Result<Tensor> {
let shape = shape.into();
let path = if self.path.is_empty() {
tensor_name.to_string()
} else {
[&self.path.join("."), tensor_name].join(".")
};
let mut tensor_data = self.data.lock().unwrap();
if let Some(tensor) = tensor_data.tensors.get(&path) {
let tensor_shape = tensor.shape();
if &shape != tensor_shape {
candle::bail!("shape mismatch on {path}: {shape:?} <> {tensor_shape:?}")
}
return Ok(tensor.as_tensor().clone());
}
let var = init.var(shape, tensor_data.dtype, &tensor_data.device)?;
let tensor = var.as_tensor().clone();
tensor_data.tensors.insert(path, var);
Ok(tensor)
}
fn all_vars(&self) -> Vec<Var> {
let tensor_data = self.data.lock().unwrap();
#[allow(clippy::map_clone)]
tensor_data
.tensors
.values()
.map(|c| c.clone())
.collect::<Vec<_>>()
}
}
fn linear_z(in_dim: usize, out_dim: usize, vs: VarStore) -> Result<Linear> {
let ws = vs.get((out_dim, in_dim), "weight", candle_nn::init::ZERO)?;
let bs = vs.get(out_dim, "bias", candle_nn::init::ZERO)?;
Ok(Linear::new(ws, Some(bs)))
}
fn linear(in_dim: usize, out_dim: usize, vs: VarStore) -> Result<Linear> {
let init_ws = candle_nn::init::DEFAULT_KAIMING_NORMAL;
let ws = vs.get((out_dim, in_dim), "weight", init_ws)?;
let bound = 1. / (in_dim as f64).sqrt();
let init_bs = Init::Uniform {
lo: -bound,
up: bound,
};
let bs = vs.get(out_dim, "bias", init_bs)?;
Ok(Linear::new(ws, Some(bs)))
}
trait Model: Sized {
fn new(vs: VarStore) -> Result<Self>;
fn forward(&self, xs: &Tensor) -> Result<Tensor>;
}
struct LinearModel {
linear: Linear,
}
impl Model for LinearModel {
fn new(vs: VarStore) -> Result<Self> {
let linear = linear_z(IMAGE_DIM, LABELS, vs)?;
Ok(Self { linear })
}
fn forward(&self, xs: &Tensor) -> Result<Tensor> {
self.linear.forward(xs)
}
}
struct Mlp {
ln1: Linear,
ln2: Linear,
}
impl Model for Mlp {
fn new(vs: VarStore) -> Result<Self> {
let ln1 = linear(IMAGE_DIM, 100, vs.pp("ln1"))?;
let ln2 = linear(100, LABELS, vs.pp("ln2"))?;
Ok(Self { ln1, ln2 })
}
fn forward(&self, xs: &Tensor) -> Result<Tensor> {
let xs = self.ln1.forward(xs)?;
let xs = xs.relu()?;
self.ln2.forward(&xs)
}
}
fn training_loop<M: Model>(
m: candle_nn::vision::Dataset,
learning_rate: f64,
) -> anyhow::Result<()> {
let dev = candle::Device::cuda_if_available(0)?;
let train_labels = m.train_labels;
let train_images = m.train_images;
let train_labels = train_labels.to_dtype(DType::U32)?.unsqueeze(1)?;
let vs = VarStore::new(DType::F32, dev);
let model = M::new(vs.clone())?;
let all_vars = vs.all_vars();
let all_vars = all_vars.iter().collect::<Vec<_>>();
let sgd = candle_nn::SGD::new(&all_vars, learning_rate);
let test_images = m.test_images;
let test_labels = m.test_labels.to_dtype(DType::U32)?;
for epoch in 1..200 {
let logits = model.forward(&train_images)?;
let log_sm = ops::log_softmax(&logits, D::Minus1)?;
let loss = loss::nll(&log_sm, &train_labels)?;
sgd.backward_step(&loss)?;
let test_logits = model.forward(&test_images)?;
let sum_ok = test_logits
.argmax(D::Minus1)?
.eq(&test_labels)?
.to_dtype(DType::F32)?
.sum_all()?
.to_scalar::<f32>()?;
let test_accuracy = sum_ok / test_labels.dims1()? as f32;
println!(
"{epoch:4} train loss: {:8.5} test acc: {:5.2}%",
loss.to_scalar::<f32>()?,
100. * test_accuracy
);
}
Ok(())
}
#[derive(ValueEnum, Clone)]
enum WhichModel {
Linear,
Mlp,
}
#[derive(Parser)]
struct Args {
#[clap(value_enum, default_value_t = WhichModel::Linear)]
model: WhichModel,
#[arg(long)]
learning_rate: Option<f64>,
}
pub fn main() -> anyhow::Result<()> {
let args = Args::parse();
// Load the dataset
let m = candle_nn::vision::mnist::load_dir("data")?;
println!("train-images: {:?}", m.train_images.shape());
println!("train-labels: {:?}", m.train_labels.shape());
println!("test-images: {:?}", m.test_images.shape());
println!("test-labels: {:?}", m.test_labels.shape());
match args.model {
WhichModel::Linear => training_loop::<LinearModel>(m, args.learning_rate.unwrap_or(1.)),
WhichModel::Mlp => training_loop::<Mlp>(m, args.learning_rate.unwrap_or(0.01)),
}
}
|
