mlx_rs/nn/
value_and_grad.rs

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
use crate::module::{update_parameters, ModuleParameters};
use crate::transforms::keyed_value_and_grad;
use crate::{error::Exception, Array};

use crate::module::FlattenedModuleParam;

fn trainable_params(model: &impl ModuleParameters) -> FlattenedModuleParam {
    model
        .trainable_parameters()
        .flatten()
        .into_iter()
        .map(|(k, v)| (k, v.clone()))
        .collect()
}

/// Helper trait for [`value_and_grad`]
pub trait IntoModuleValueAndGrad<'a, M, Args, Val, Err>
where
    M: ModuleParameters + 'a,
    Args: Clone,
{
    /// Computes the valud and gradient of the passed function `f(model, args)` with regard to the
    /// model's trainable parameters.
    fn into_module_value_and_grad(
        self,
    ) -> impl FnMut(&mut M, Args) -> Result<(Val, FlattenedModuleParam), Exception> + 'a;
}

impl<'a, F, M, Args> IntoModuleValueAndGrad<'a, M, Args, Vec<Array>, ()> for F
where
    M: ModuleParameters + 'a,
    F: FnMut(&mut M, Args) -> Vec<Array> + 'a,
    Args: Clone,
{
    fn into_module_value_and_grad(
        mut self,
    ) -> impl FnMut(&mut M, Args) -> Result<(Vec<Array>, FlattenedModuleParam), Exception> + 'a
    {
        move |model, arrays| {
            let trainable_parameters = trainable_params(model);
            let inner = |parameters: FlattenedModuleParam, arrays: Args| -> Vec<Array> {
                let flattened_parameters = parameters.into_iter();
                update_parameters(model, flattened_parameters);

                self(model, arrays)
            };
            let mut vg = keyed_value_and_grad(inner);

            let (v, g) = vg(trainable_parameters, arrays)?;
            Ok((v, g))
        }
    }
}

impl<'a, F, M, Args> IntoModuleValueAndGrad<'a, M, Args, Vec<Array>, Exception> for F
where
    M: ModuleParameters + 'a,
    F: FnMut(&mut M, Args) -> Result<Vec<Array>, Exception> + 'a,
    Args: Clone,
{
    fn into_module_value_and_grad(
        mut self,
    ) -> impl FnMut(&mut M, Args) -> Result<(Vec<Array>, FlattenedModuleParam), Exception> + 'a
    {
        move |model, arrays| {
            let trainable_parameters = trainable_params(model);
            let inner =
                |parameters: FlattenedModuleParam, arrays: Args| -> Result<Vec<Array>, Exception> {
                    let flattened_parameters = parameters.into_iter().map(|(k, v)| (k, v.clone()));
                    update_parameters(model, flattened_parameters);

                    self(model, arrays)
                };
            let mut vg = keyed_value_and_grad(inner);

            let (v, g) = vg(trainable_parameters, arrays)?;
            Ok((v, g))
        }
    }
}

impl<'a, F, M, Args> IntoModuleValueAndGrad<'a, M, Args, Array, ()> for F
where
    M: ModuleParameters + 'a,
    F: FnMut(&mut M, Args) -> Array + 'a,
    Args: Clone,
{
    fn into_module_value_and_grad(
        mut self,
    ) -> impl FnMut(&mut M, Args) -> Result<(Array, FlattenedModuleParam), Exception> + 'a {
        move |model, arrays| {
            let trainable_parameters = trainable_params(model);
            let inner = |parameters: FlattenedModuleParam, arrays: Args| -> Vec<Array> {
                let flattened_parameters = parameters.into_iter().map(|(k, v)| (k, v.clone()));
                update_parameters(model, flattened_parameters);

                vec![self(model, arrays)]
            };
            let mut vg = keyed_value_and_grad(inner);

            let (v, g) = vg(trainable_parameters, arrays)?;
            let v = v.into_iter().next().expect("Expected a single value");
            Ok((v, g))
        }
    }
}

impl<'a, F, M, Args> IntoModuleValueAndGrad<'a, M, Args, Array, Exception> for F
where
    M: ModuleParameters + 'a,
    F: FnMut(&mut M, Args) -> Result<Array, Exception> + 'a,
    Args: Clone,
{
    fn into_module_value_and_grad(
        mut self,
    ) -> impl FnMut(&mut M, Args) -> Result<(Array, FlattenedModuleParam), Exception> + 'a {
        move |model, arrays| {
            let trainable_parameters = trainable_params(model);
            let inner =
                |parameters: FlattenedModuleParam, arrays: Args| -> Result<Vec<Array>, Exception> {
                    let flattened_parameters = parameters.into_iter().map(|(k, v)| (k, v.clone()));
                    update_parameters(model, flattened_parameters);

                    self(model, arrays).map(|v| vec![v])
                };
            let mut vg = keyed_value_and_grad(inner);

            let (v, g) = vg(trainable_parameters, arrays)?;
            let v = v.into_iter().next().expect("Expected a single value");
            Ok((v, g))
        }
    }
}

/// Transform the passed function `f(model, args)` to a function that computes the gradients of `f`
/// with regard to the model's trainable parameters and also its value.
pub fn value_and_grad<'a, F, M, Args, Val, Err>(
    f: F,
) -> impl FnMut(&mut M, Args) -> Result<(Val, FlattenedModuleParam), Exception> + 'a
where
    M: ModuleParameters + 'a,
    F: IntoModuleValueAndGrad<'a, M, Args, Val, Err>,
    Args: Clone,
{
    f.into_module_value_and_grad()
}

#[cfg(test)]
mod tests {
    use crate::module::Module;
    use crate::{array, error::Exception, Array};

    use crate::nn::{self, Linear};

    // The unit test below is adapted from `test_compiled_optimizer` in
    // `mlx/python/tests/test_optimizers.py``
    #[test]
    fn test_value_and_grad() {
        let mut model = Linear::new(2, 2).unwrap();
        let x = crate::random::uniform::<_, f32>(1.0, 2.0, &[2, 2], None).unwrap();

        let loss = |model: &mut Linear, x: &Array| -> Vec<Array> {
            vec![model.forward(x).unwrap().sum(None, None).unwrap()]
        };

        let mut vg = nn::value_and_grad(loss);
        let (v, g) = vg(&mut model, &x).unwrap();

        assert_ne!(v[0].sum(None, None).unwrap(), array!(0.0));
        assert_ne!(g["weight"].sum(None, None).unwrap(), array!(0.0));
        assert_ne!(g["bias"].sum(None, None).unwrap(), array!(0.0));
    }

    #[test]
    fn test_value_and_grad_with_unary_output() {
        let mut model = Linear::new(2, 2).unwrap();
        let x = crate::random::uniform::<_, f32>(1.0, 2.0, &[2, 2], None).unwrap();

        let loss = |model: &mut Linear, x: &Array| -> Array {
            model.forward(x).unwrap().sum(None, None).unwrap()
        };

        let mut vg = nn::value_and_grad(loss);
        let (v, g) = vg(&mut model, &x).unwrap();

        assert_ne!(v.sum(None, None).unwrap(), array!(0.0));
        assert_ne!(g["weight"].sum(None, None).unwrap(), array!(0.0));
        assert_ne!(g["bias"].sum(None, None).unwrap(), array!(0.0));
    }

    #[test]
    fn test_fallible_module_value_and_grad() {
        let mut model = Linear::new(2, 2).unwrap();
        let x = crate::random::uniform::<_, f32>(1.0, 2.0, &[2, 2], None).unwrap();

        let loss = |model: &mut Linear, x: &Array| -> Result<Vec<Array>, Exception> {
            Ok(vec![model.forward(x)?.sum(None, None)?])
        };

        let mut vg = nn::value_and_grad(loss);
        let (v, g) = vg(&mut model, &x).unwrap();

        assert_ne!(v[0].sum(None, None).unwrap(), array!(0.0));
        assert_ne!(g["weight"].sum(None, None).unwrap(), array!(0.0));
        assert_ne!(g["bias"].sum(None, None).unwrap(), array!(0.0));
    }

    #[test]
    fn test_value_and_grad_with_two_args() {
        let mut model = Linear::new(2, 2).unwrap();
        let x = crate::random::uniform::<_, f32>(1.0, 2.0, &[2, 2], None).unwrap();
        let y = crate::ops::ones::<f32>(x.shape()).unwrap();

        let loss =
            |model: &mut Linear, (x, y): (&Array, &Array)| -> Result<Vec<Array>, Exception> {
                model
                    .forward(x)?
                    .subtract(y)?
                    .square()?
                    .sum(None, None)
                    .map(|v| vec![v])
            };

        let mut vg = nn::value_and_grad(loss);
        let (v, g) = vg(&mut model, (&x, &y)).unwrap();

        assert_ne!(v[0].sum(None, None).unwrap(), array!(0.0));
        assert_ne!(g["weight"].sum(None, None).unwrap(), array!(0.0));
        assert_ne!(g["bias"].sum(None, None).unwrap(), array!(0.0));
    }

    #[test]
    fn test_value_and_grad_with_error() {
        let mut model = Linear::new(2, 2).unwrap();
        // Use a shape that is not compatible with the model
        let x = crate::random::uniform::<_, f32>(1.0, 2.0, &[3, 3], None).unwrap();

        let loss = |model: &mut Linear, x: &Array| -> Result<Vec<Array>, Exception> {
            Ok(vec![model.forward(x)?.sum(None, None)?])
        };

        let mut vg = nn::value_and_grad(loss);
        let result = vg(&mut model, &x);

        assert!(result.is_err());

        // Check that the error message is not just "mlx_closure returned a non-zero value"
        let err = result.unwrap_err();
        assert!(!err.what().contains("non-zero value"))
    }
}