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
414
415
416
417
418
419
420
421
422
423
424
425
426
427
//! Provides the generics and interfaces for the specific Solvers.
//!
//! See [Solvers][solvers]
//! [solvers]: ../solvers/index.html

pub mod confusion_matrix;
pub mod regression_evaluator;

pub use self::confusion_matrix::ConfusionMatrix;
pub use self::regression_evaluator::{RegressionEvaluator, RegressionLoss};
use crate::co::prelude::*;
use crate::layer::*;
use crate::layers::SequentialConfig;
use crate::solvers::*;
use std::marker::PhantomData;

use crate::util::{ArcLock, LayerOps, SolverOps};
use std::rc::Rc;

#[derive(Debug)]
/// Solver that optimizes a [Layer][1] with a given objective.
/// [1]: ../layer/index.html
pub struct Solver<SolverB, B>
where
    SolverB: IBackend + SolverOps<f32>,
    B: IBackend + LayerOps<f32>,
{
    net: Layer<B>,
    objective: Layer<SolverB>,
    /// The implementation of the Solver
    pub worker: Box<dyn ISolver<SolverB, B>>,

    config: SolverConfig,

    /// The current iteration / number of times weights have been updated
    iter: usize,

    solver_backend: PhantomData<SolverB>,
}

impl<SolverB, B> Solver<SolverB, B>
where
    SolverB: IBackend + SolverOps<f32> + 'static,
    B: IBackend + LayerOps<f32> + 'static,
{
    /// Create Solver from [SolverConfig][1]
    /// [1]: ./struct.SolverConfig.html
    ///
    /// This is the **preferred method** to create a Solver for training a neural network.
    pub fn from_config(net_backend: Rc<B>, obj_backend: Rc<SolverB>, config: &SolverConfig) -> Solver<SolverB, B> {
        let network = Layer::from_config(net_backend, &config.network);
        let mut worker = config.solver.with_config(obj_backend.clone(), &config);
        worker.init(&network);

        Solver {
            worker: worker,
            net: network,
            objective: Layer::from_config(obj_backend, &config.objective),
            iter: 0,

            config: config.clone(),
            solver_backend: PhantomData::<SolverB>,
        }
    }
}

impl<SolverB, B> Solver<SolverB, B>
where
    SolverB: IBackend + SolverOps<f32> + 'static,
    B: IBackend + LayerOps<f32> + 'static,
{
    fn init(&mut self, backend: Rc<B>) {
        info!("Initializing solver from configuration");

        let mut config = self.config.clone();
        self.init_net(backend, &mut config);
    }

    /// Initialize the training net
    fn init_net(&mut self, backend: Rc<B>, param: &mut SolverConfig) {
        self.net = Layer::from_config(backend, &param.network);
    }

    /// Train the network with one minibatch
    pub fn train_minibatch(
        &mut self,
        mb_data: ArcLock<SharedTensor<f32>>,
        mb_target: ArcLock<SharedTensor<f32>>,
    ) -> ArcLock<SharedTensor<f32>> {
        // forward through network and classifier
        let network_out = self.net.forward(&[mb_data])[0].clone();
        let _ = self.objective.forward(&[network_out.clone(), mb_target]);

        // forward through network and classifier
        let classifier_gradient = self.objective.backward(&[]);
        self.net.backward(&classifier_gradient[0..1]);

        self.worker.compute_update(&self.config, &mut self.net, self.iter);
        self.net.update_weights(self.worker.backend());
        self.iter += 1;

        network_out
    }

    /// Returns the network trained by the solver.
    ///
    /// This is the recommended method to get a usable trained network.
    pub fn network(&self) -> &Layer<B> {
        &self.net
    }

    /// Returns the network trained by the solver.
    ///
    /// This is the recommended method to get a trained network,
    /// if you want to alter the network. Keep in mind that altering the network
    /// might render the solver unusable and continuing training the network with it will yield
    /// unexpected results.
    pub fn mut_network(&mut self) -> &mut Layer<B> {
        &mut self.net
    }
}

/// Implementation of a specific Solver.
///
/// See [Solvers][1]
/// [1]: ../solvers/index.html
pub trait ISolver<SolverB, B>
where
    B: IBackend + LayerOps<f32>,
    SolverB: IBackend + SolverOps<f32>,
{
    /// Initialize the solver, setting up any network related data.
    fn init(&mut self, net: &Layer<B>) {}

    /// Update the weights of the net with part of the gradient.
    ///
    /// The [second phase of backpropagation learning][1].
    /// Calculates the gradient update that should be applied to the network,
    /// and then applies that gradient to the network, changing its weights.
    ///
    /// [1]: https://en.wikipedia.org/wiki/Backpropagation#Phase_2:_Weight_update
    ///
    /// Used by [step][2] to optimize the network.
    ///
    /// [2]: ./struct.Solver.html#method.step
    fn compute_update(&mut self, param: &SolverConfig, network: &mut Layer<B>, iter: usize);

    /// Returns the backend used by the solver.
    fn backend(&self) -> &SolverB;
}

impl<SolverB, B: IBackend + LayerOps<f32>> ::std::fmt::Debug for dyn ISolver<SolverB, B> {
    fn fmt(&self, f: &mut ::std::fmt::Formatter) -> ::std::fmt::Result {
        write!(f, "({})", "ILayer")
    }
}

#[derive(Debug, Clone)]
/// Configuration for a Solver
pub struct SolverConfig {
    /// Name of the solver.
    pub name: String,
    /// The [LayerConfig][1] that is used to initialize the network.
    /// [1]: ../layer/struct.LayerConfig.html
    pub network: LayerConfig,
    /// The [LayerConfig][1] that is used to initialize the objective.
    /// [1]: ../layer/struct.LayerConfig.html
    pub objective: LayerConfig,
    /// The [Solver implementation][1] to be used.
    /// [1]: ../solvers/index.html
    pub solver: SolverKind,
    /// Accumulate gradients over `minibatch_size` instances.
    ///
    /// Default: 1
    pub minibatch_size: usize,
    /// The learning rate policy to be used.
    ///
    /// Default: Fixed
    pub lr_policy: LRPolicy,
    /// The base learning rate.
    ///
    /// Default: 0.01
    pub base_lr: f32,
    /// gamma as used in the calculation of most learning rate policies.
    ///
    /// Default: 0.1
    pub gamma: f32,
    /// The stepsize used in Step and Sigmoid learning policies.
    ///
    /// Default: 10
    pub stepsize: usize,
    /// The threshold for clipping gradients.
    ///
    /// Gradient values will be scaled to their [L2 norm][1] of length `clip_gradients`
    /// if their L2 norm is larger than `clip_gradients`.
    /// If set to `None` gradients will not be clipped.
    ///
    /// [1]: https://en.wikipedia.org/wiki/Norm_(mathematics)#Euclidean_norm
    ///
    /// Default: None
    pub clip_gradients: Option<f32>,
    /// The global [weight decay][1] multiplier for [regularization][2].
    /// [1]: http://www.alglib.net/dataanalysis/improvinggeneralization.php#header3
    /// [2]: https://cs231n.github.io/neural-networks-2/#reg
    ///
    /// Regularization can prevent [overfitting][3].
    ///
    /// If set to `None` no regularization will be performed.
    ///
    /// [3]: https://cs231n.github.io/neural-networks-2/#reg
    pub weight_decay: Option<f32>,
    /// The method of [regularization][1] to use.
    /// [1]: https://cs231n.github.io/neural-networks-2/#reg
    ///
    /// There are different methods for regularization.
    /// The two most common ones are [L1 regularization][1] and [L2 regularization][1].
    ///
    /// See [RegularizationMethod][2] for all implemented methods.
    ///
    /// [2]: ./enum.RegularizationMethod.html
    ///
    /// Currently only L2 regularization is implemented.
    /// See [Issue #23](https://github.com/spearow/juice/issues/23).
    pub regularization_method: Option<RegularizationMethod>,
    /// The [momentum][1] multiplier for [SGD solvers][2].
    /// [1]: https://en.wikipedia.org/wiki/Stochastic_gradient_descent#Momentum
    /// [2]: ../solvers/sgd/index.html
    ///
    /// For more information see [SGD with momentum][3]
    /// [3]: ../solvers/sgd/momentum/index.html
    ///
    /// The value should always be between 0 and 1 and dictates how much of the previous
    /// gradient update will be added to the current one.
    ///
    /// Default: 0
    pub momentum: f32,
}

impl Default for SolverConfig {
    fn default() -> SolverConfig {
        SolverConfig {
            name: "".to_owned(),
            network: LayerConfig::new("default", SequentialConfig::default()),
            objective: LayerConfig::new("default", SequentialConfig::default()),
            solver: SolverKind::SGD(SGDKind::Momentum),

            minibatch_size: 1,

            lr_policy: LRPolicy::Fixed,
            base_lr: 0.01f32,
            gamma: 0.1f32,
            stepsize: 10,

            clip_gradients: None,

            weight_decay: None,
            regularization_method: None,

            momentum: 0f32,
        }
    }
}

impl SolverConfig {
    /// Return the learning rate for a supplied iteration.
    ///
    /// The way the learning rate is calculated depends on the configured [LRPolicy][1].
    ///
    /// [1]: ./enum.LRPolicy.html
    ///
    /// Used by the [Solver][2] to calculate the learning rate for the current iteration.
    /// The calculated learning rate has a different effect on training dependent on what
    /// [type of Solver][3] you are using.
    ///
    /// [2]: ./struct.Solver.html
    /// [3]: ../solvers/index.html
    pub fn get_learning_rate(&self, iter: usize) -> f32 {
        match self.lr_policy() {
            LRPolicy::Fixed => self.base_lr(),
            LRPolicy::Step => {
                let current_step = self.step(iter);
                self.base_lr() * self.gamma().powf(current_step as f32)
            }
            // LRPolicy::Multistep => {
            //     // TODO: the current step can be calculated on-demand
            //     //   if (this->current_step_ < this->param_.stepvalue_size() &&
            //     //         this->iter_ >= this->param_.stepvalue(this->current_step_)) {
            //     //     this->current_step_++;
            //     //     LOG(INFO) << "MultiStep Status: Iteration " <<
            //     //     this->iter_ << ", step = " << this->current_step_;
            //     //   }
            //     //   rate = this->param_.base_lr() *
            //     //       pow(this->param_.gamma(), this->current_step_);
            //     unimplemented!();
            // }
            LRPolicy::Exp => self.base_lr() * self.gamma().powf(iter as f32),
            // LRPolicy::Inv => {
            //     //   rate = this->param_.base_lr() *
            //     //       pow(Dtype(1) + this->param_.gamma() * this->iter_,
            //     //           - this->param_.power());
            //     unimplemented!();
            // }
            // LRPolicy::Poly => {
            //     //   rate = this->param_.base_lr() * pow(Dtype(1.) -
            //     //       (Dtype(this->iter_) / Dtype(this->param_.max_iter())),
            //     //       this->param_.power());
            //     unimplemented!();
            // }
            // LRPolicy::Sigmoid => {
            //     //   rate = this->param_.base_lr() * (Dtype(1.) /
            //     //       (Dtype(1.) + exp(-this->param_.gamma() * (Dtype(this->iter_) -
            //     //         Dtype(this->param_.stepsize())))));
            //     unimplemented!();
            // }
        }
    }

    /// Return current step at iteration `iter`.
    ///
    /// Small helper for learning rate calculation.
    fn step(&self, iter: usize) -> usize {
        iter / self.stepsize()
    }

    /// Return learning rate policy.
    fn lr_policy(&self) -> LRPolicy {
        self.lr_policy
    }

    /// Return the base learning rate.
    fn base_lr(&self) -> f32 {
        self.base_lr
    }

    /// Return the gamma for learning rate calculations.
    fn gamma(&self) -> f32 {
        self.gamma
    }

    /// Return the stepsize for learning rate calculations.
    fn stepsize(&self) -> usize {
        self.stepsize
    }
}

#[derive(Debug, Copy, Clone)]
/// All available types of solvers.
pub enum SolverKind {
    /// Stochastic Gradient Descent.
    /// See [SGDKind][1] for all available SGD solvers.
    /// [1]: ./enum.SGDKind.html
    SGD(SGDKind),
}

impl SolverKind {
    /// Create a Solver of the specified kind with the supplied SolverConfig.
    pub fn with_config<B: IBackend + SolverOps<f32> + 'static, NetB: IBackend + LayerOps<f32> + 'static>(
        &self,
        backend: Rc<B>,
        config: &SolverConfig,
    ) -> Box<dyn ISolver<B, NetB>> {
        match *self {
            SolverKind::SGD(sgd) => sgd.with_config(backend, config),
        }
    }
}

#[derive(Debug, Copy, Clone)]
/// All available types of Stochastic Gradient Descent solvers.
pub enum SGDKind {
    /// Stochastic Gradient Descent with Momentum. See [implementation][1]
    /// [1] ../solvers/
    Momentum,
}

impl SGDKind {
    /// Create a Solver of the specified kind with the supplied SolverConfig.
    pub fn with_config<B: IBackend + SolverOps<f32> + 'static, NetB: IBackend + LayerOps<f32> + 'static>(
        &self,
        backend: Rc<B>,
        config: &SolverConfig,
    ) -> Box<dyn ISolver<B, NetB>> {
        match *self {
            SGDKind::Momentum => Box::new(Momentum::<B>::new(backend)),
        }
    }
}

#[derive(Debug, Copy, Clone)]
/// Learning Rate Policy for a [Solver][1]
/// [1]: ./struct.Solver.html
///
/// The variables mentioned below are defined in the [SolverConfig][2] apart from
/// iter, which is the current iteration of the solver, that is supplied as a parameter
/// for the learning rate calculation.
///
/// [2]: ./struct.SolverConfig.html
pub enum LRPolicy {
    /// always return base_lr
    Fixed,
    /// learning rate decays every `step` iterations.
    /// return base_lr * gamma ^ (floor(iter / step))
    Step,
    // /// similar to step but it allows non uniform steps defined by
    // /// stepvalue
    // Multistep,
    /// return base_lr * gamma ^ iter
    Exp,
    // /// return base_lr * (1 + gamma * iter) ^ (- power)
    // Inv,
    // /// the effective learning rate follows a polynomial decay, to be
    // /// zero by the max_iter.
    // /// return base_lr (1 - iter/max_iter) ^ (power)
    // Poly,
    // /// the effective learning rate follows a sigmod decay
    // /// return base_lr ( 1/(1 + exp(-gamma * (iter - stepsize))))
    // Sigmoid,
}

#[derive(Debug, Copy, Clone)]
/// [Regularization][1] method for a [Solver][2].
/// [1]: https://cs231n.github.io/neural-networks-2/#reg
/// [2]: ./struct.Solver.html
pub enum RegularizationMethod {
    /// L2 regularization
    L2,
}