incin-core 0.1.0

Core tensor operations and static shape checking for incin
Documentation
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
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
use crate::err::{Error, Result};
use crate::nn::module::{Module, ShapeInfo, TrainMode};
use crate::shapes::{DynShape, Shape};
use crate::tensor::base::Tensor;
use crate::tensor::device::Device;
use crate::tensor::grad::RequiresGrad;
use alloc::{
    string::{String, ToString},
    vec::Vec,
};

/// The Rectified Linear Unit (ReLU) activation function: `f(x) = max(0, x)`.
///
/// This is a stateless module with no learnable parameters.
#[derive(Debug, Clone, Default)]
pub struct ReLU;

macro_rules! impl_stateless_shape_info {
    ($($ty:ty),+ $(,)?) => {
        $(impl ShapeInfo for $ty {
            fn shape_info(&self) -> Option<String> {
                None
            }
        })+
    };
}

impl_stateless_shape_info!(ReLU);

/// Stateless - no training-dependent behavior, opts in with the trait's
/// default no-op so it can appear inside a `Sequential` alongside layers
/// that do have one (e.g. `Dropout`).
impl TrainMode for ReLU {}

use crate::exec::catalog::op;
use crate::tensor::backend::Execute;

impl<
    S: Shape + DynShape,
    B: crate::tensor::backend::VariableBackend + Execute<op::Relu>,
    G: RequiresGrad,
> Module<Tensor<S, B, f32, G>> for ReLU
where
    <B as Execute<op::Relu>>::Output: Into<B::Storage<f32>>,
{
    /// The output tensor type produced by this module's forward pass.
    type Output = Tensor<S, B, f32, G>;
    /// The error type returned if the forward pass fails.
    type Error = Error;

    #[inline]
    /// Runs the forward pass of this module on the given input.
    fn forward(
        &self,
        x: Tensor<S, B, f32, G>,
    ) -> core::result::Result<Tensor<S, B, f32, G>, Error> {
        x.relu()
    }
}

/// The Gaussian Error Linear Unit (GELU) activation function.
///
/// GELU is a smooth approximation to ReLU commonly used in transformer architectures.
/// This is a stateless module with no learnable parameters.
#[derive(Debug, Clone, Default)]
pub struct GELU;

impl_stateless_shape_info!(GELU);

/// Stateless - no training-dependent behavior, opts in with the trait's
/// default no-op so it can appear inside a `Sequential` alongside layers
/// that do have one (e.g. `Dropout`).
impl TrainMode for GELU {}

impl<S: Shape + DynShape, B: crate::tensor::backend::VariableBackend + Execute<op::Gelu>>
    Module<Tensor<S, B>> for GELU
where
    <B as Execute<op::Gelu>>::Output: Into<B::Storage<f32>>,
{
    /// The output tensor type produced by this module's forward pass.
    type Output = Tensor<S, B>;
    /// The error type returned if the forward pass fails.
    type Error = Error;

    #[inline]
    /// Runs the forward pass of this module on the given input.
    fn forward(&self, x: Tensor<S, B>) -> core::result::Result<Tensor<S, B>, Error> {
        x.gelu()
    }
}

/// The Swish (SiLU) activation function: `f(x) = x * sigmoid(x)`.
///
/// Swish is a smooth, non-monotonic function that consistently performs better than ReLU
/// in deeper networks. This is a stateless module with no learnable parameters.
#[derive(Debug, Clone, Default)]
pub struct Swish;

impl_stateless_shape_info!(Swish);

/// Stateless - no training-dependent behavior, opts in with the trait's
/// default no-op so it can appear inside a `Sequential` alongside layers
/// that do have one (e.g. `Dropout`).
impl TrainMode for Swish {}

impl<S: Shape + DynShape, B: crate::tensor::backend::VariableBackend + Execute<op::Swish>>
    Module<Tensor<S, B>> for Swish
where
    <B as Execute<op::Swish>>::Output: Into<B::Storage<f32>>,
{
    /// The output tensor type produced by this module's forward pass.
    type Output = Tensor<S, B>;
    /// The error type returned if the forward pass fails.
    type Error = Error;

    #[inline]
    /// Runs the forward pass of this module on the given input.
    fn forward(&self, x: Tensor<S, B>) -> core::result::Result<Tensor<S, B>, Error> {
        x.swish()
    }
}

/// The Mish activation function: `f(x) = x * tanh(softplus(x))`.
///
/// Mish is a smooth, continuous, non-monotonic function that can improve training dynamics.
/// This is a stateless module with no learnable parameters.
#[derive(Debug, Clone, Default)]
pub struct Mish;

impl_stateless_shape_info!(Mish);

/// Stateless - no training-dependent behavior, opts in with the trait's
/// default no-op so it can appear inside a `Sequential` alongside layers
/// that do have one (e.g. `Dropout`).
impl TrainMode for Mish {}

impl<S: Shape + DynShape, B: crate::tensor::backend::VariableBackend + Execute<op::Mish>>
    Module<Tensor<S, B>> for Mish
where
    <B as Execute<op::Mish>>::Output: Into<B::Storage<f32>>,
{
    /// The output tensor type produced by this module's forward pass.
    type Output = Tensor<S, B>;
    /// The error type returned if the forward pass fails.
    type Error = Error;

    #[inline]
    /// Runs the forward pass of this module on the given input.
    fn forward(&self, x: Tensor<S, B>) -> core::result::Result<Tensor<S, B>, Error> {
        x.mish()
    }
}

/// The Exponential Linear Unit (ELU) activation function.
///
/// ELU approaches a negative constant as the input gets smaller.
/// This implementation hardcodes alpha to 1.0.
/// This is a stateless module with no learnable parameters.
#[derive(Debug, Clone, Default)]
#[allow(clippy::upper_case_acronyms)]
pub struct ELU;

impl_stateless_shape_info!(ELU);

/// Stateless - no training-dependent behavior, opts in with the trait's
/// default no-op so it can appear inside a `Sequential` alongside layers
/// that do have one (e.g. `Dropout`).
impl TrainMode for ELU {}

impl<S: Shape + DynShape, B: crate::tensor::backend::VariableBackend + Execute<op::Elu>>
    Module<Tensor<S, B>> for ELU
where
    <B as Execute<op::Elu>>::Output: Into<B::Storage<f32>>,
{
    /// The output tensor type produced by this module's forward pass.
    type Output = Tensor<S, B>;
    /// The error type returned if the forward pass fails.
    type Error = Error;

    #[inline]
    /// Runs the forward pass of this module on the given input.
    fn forward(&self, x: Tensor<S, B>) -> core::result::Result<Tensor<S, B>, Error> {
        x.elu()
    }
}

/// The Softmax activation function, applied along a specified axis.
///
/// Converts a vector of raw logits into a probability distribution that sums to 1.
///
/// ## Parameters
/// * `dim` - The axis along which the softmax normalization is applied.
#[derive(Debug, Clone)]
pub struct Softmax {
    /// The axis along which softmax is applied.
    pub dim: usize,
}

impl ShapeInfo for Softmax {
    fn shape_info(&self) -> Option<String> {
        None
    }
}

impl Softmax {
    /// Creates a new instance with default (statically inferred) shape arguments.
    pub fn new(dim: usize) -> Self {
        Self { dim }
    }
}

/// Stateless - no training-dependent behavior, opts in with the trait's
/// default no-op so it can appear inside a `Sequential` alongside layers
/// that do have one (e.g. `Dropout`).
impl TrainMode for Softmax {}

impl<
    S: Shape + DynShape + crate::shapes::RuntimeRankProjection,
    B: crate::tensor::backend::VariableBackend
        + crate::tensor::backend::Execute<crate::exec::catalog::op::Softmax>,
> Module<Tensor<S, B>> for Softmax
where
    <B as crate::tensor::backend::Execute<crate::exec::catalog::op::Softmax>>::Output:
        Into<B::Storage<f32>>,
{
    /// The output tensor type produced by this module's forward pass.
    type Output = Tensor<S, B>;
    /// The error type returned if the forward pass fails.
    type Error = Error;

    #[inline]
    /// Runs the forward pass of this module on the given input.
    fn forward(&self, x: Tensor<S, B>) -> core::result::Result<Tensor<S, B>, Error> {
        x.softmax(self.dim as isize)
    }
}

/// The Sigmoid activation function: `f(x) = 1 / (1 + exp(-x))`.
///
/// Squashes each element into the range `(0, 1)`. This is a stateless module.
#[derive(Debug, Clone, Default)]
pub struct Sigmoid;

impl_stateless_shape_info!(Sigmoid);

/// Stateless - no training-dependent behavior, opts in with the trait's
/// default no-op so it can appear inside a `Sequential` alongside layers
/// that do have one (e.g. `Dropout`).
impl TrainMode for Sigmoid {}

impl<S: Shape + DynShape, B: crate::tensor::backend::VariableBackend + Execute<op::Sigmoid>>
    Module<Tensor<S, B>> for Sigmoid
where
    <B as Execute<op::Sigmoid>>::Output: Into<B::Storage<f32>>,
{
    /// The output tensor type produced by this module's forward pass.
    type Output = Tensor<S, B>;
    /// The error type returned if the forward pass fails.
    type Error = Error;

    #[inline]
    /// Runs the forward pass of this module on the given input.
    fn forward(&self, x: Tensor<S, B>) -> core::result::Result<Tensor<S, B>, Error> {
        x.sigmoid()
    }
}

/// The Hyperbolic Tangent (Tanh) activation function: `f(x) = tanh(x)`.
///
/// Squashes each element into the range `(-1, 1)`. This is a stateless module.
#[derive(Debug, Clone, Default)]
pub struct Tanh;

impl_stateless_shape_info!(Tanh);

/// Stateless - no training-dependent behavior, opts in with the trait's
/// default no-op so it can appear inside a `Sequential` alongside layers
/// that do have one (e.g. `Dropout`).
impl TrainMode for Tanh {}

impl<S: Shape + DynShape, B: crate::tensor::backend::VariableBackend + Execute<op::Tanh>>
    Module<Tensor<S, B>> for Tanh
where
    <B as Execute<op::Tanh>>::Output: Into<B::Storage<f32>>,
{
    /// The output tensor type produced by this module's forward pass.
    type Output = Tensor<S, B>;
    /// The error type returned if the forward pass fails.
    type Error = Error;

    #[inline]
    /// Runs the forward pass of this module on the given input.
    fn forward(&self, x: Tensor<S, B>) -> core::result::Result<Tensor<S, B>, Error> {
        x.tanh()
    }
}

macro_rules! impl_stateless_state_visitors {
    ($($t:ty),+ $(,)?) => {$ (
        impl<B: crate::tensor::backend::VariableBackend> crate::nn::VisitState<B> for $t {
            fn visit_state<V: crate::nn::StateVisitor<B>>(
                &self,
                _: &crate::nn::StatePath,
                _: &mut V,
            ) -> crate::err::Result<()> { Ok(()) }
        }

        impl<B: crate::tensor::backend::VariableBackend> crate::nn::VisitStateMut<B> for $t {
            fn visit_state_mut<V: crate::nn::StateMutVisitor<B>>(
                &mut self,
                _: &crate::nn::StatePath,
                _: &mut V,
            ) -> crate::err::Result<()> { Ok(()) }
        }

        impl<B: crate::tensor::backend::VariableBackend> crate::nn::VisitParameters<B> for $t {
            fn visit_parameters<V: crate::nn::ParameterVisitor<B>>(
                &self,
                _: &crate::nn::StatePath,
                _: &mut V,
            ) -> crate::err::Result<()> { Ok(()) }
        }
    )+ };
}

impl_stateless_state_visitors!(ReLU, GELU, Swish, Mish, ELU, Softmax, Sigmoid, Tanh);

impl crate::nn::module::NamedLayers for ReLU {
    /// Returns the layer hierarchy rooted at this module for visualization.
    fn layer_structure(&self, prefix: &str) -> Vec<crate::nn::module::LayerNode> {
        vec![crate::nn::module::LayerNode {
            name: prefix.to_string(),
            type_name: alloc::string::String::from("ReLU"),
            shape_info: "".to_string(),
            children: vec![],
        }]
    }
}

impl crate::nn::module::NamedLayers for GELU {
    /// Returns the layer hierarchy rooted at this module for visualization.
    fn layer_structure(&self, prefix: &str) -> Vec<crate::nn::module::LayerNode> {
        vec![crate::nn::module::LayerNode {
            name: prefix.to_string(),
            type_name: alloc::string::String::from("GELU"),
            shape_info: "".to_string(),
            children: vec![],
        }]
    }
}

impl crate::nn::module::NamedLayers for Swish {
    /// Returns the layer hierarchy rooted at this module for visualization.
    fn layer_structure(&self, prefix: &str) -> Vec<crate::nn::module::LayerNode> {
        vec![crate::nn::module::LayerNode {
            name: prefix.to_string(),
            type_name: alloc::string::String::from("Swish"),
            shape_info: "".to_string(),
            children: vec![],
        }]
    }
}

impl crate::nn::module::NamedLayers for Sigmoid {
    /// Returns the layer hierarchy rooted at this module for visualization.
    fn layer_structure(&self, prefix: &str) -> Vec<crate::nn::module::LayerNode> {
        vec![crate::nn::module::LayerNode {
            name: prefix.to_string(),
            type_name: alloc::string::String::from("Sigmoid"),
            shape_info: "".to_string(),
            children: vec![],
        }]
    }
}

impl crate::nn::module::NamedLayers for Tanh {
    /// Returns the layer hierarchy rooted at this module for visualization.
    fn layer_structure(&self, prefix: &str) -> Vec<crate::nn::module::LayerNode> {
        vec![crate::nn::module::LayerNode {
            name: prefix.to_string(),
            type_name: alloc::string::String::from("Tanh"),
            shape_info: "".to_string(),
            children: vec![],
        }]
    }
}

impl crate::nn::module::NamedLayers for Softmax {
    /// Returns the layer hierarchy rooted at this module for visualization.
    fn layer_structure(&self, prefix: &str) -> Vec<crate::nn::module::LayerNode> {
        vec![crate::nn::module::LayerNode {
            name: prefix.to_string(),
            type_name: alloc::string::String::from("Softmax"),
            shape_info: format!("dim={}", self.dim),
            children: vec![],
        }]
    }
}

macro_rules! impl_unit_to_device {
    ($($t:ty),+) => {
        $(
            impl<B: crate::tensor::backend::VariableBackend, NewD: Device> crate::tensor::transfer::ToDevice<B, NewD> for $t {
                type Output = $t;
                fn to_device(self, _arg: &NewD::Arg) -> Result<Self::Output> {
                    Ok(self)
                }
            }
        )+
    };
}

impl_unit_to_device!(ReLU, GELU, Swish, Mish, ELU, Softmax, Sigmoid, Tanh);