hal-ml 0.2.0

HAL: a machine learning library that is able to run on Nvidia, OpenCL or CPU BLAS based compute backends. It currently provides stackable classical neural networks, RNN's and soon to be LSTM's. A differentiation of this package is that we are looking to implement RTRL (instead of just BPTT) for the recurrent layers in order to provide a solid framework for online learning. We will also (in the future) be implementing various layers such as unitary RNN's, NTM's and Adaptive Computation time based LSTM's. HAL also comes with the ability to plot and do many basic math operations on arrays.
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

use af;
use af::{Array, Dim4, DType};
use itertools::Zip;
use std::collections::HashMap;
use std::default::Default;

use params::ParamManager;
use initializations;
use optimizer;
use optimizer::Optimizer;

#[allow(non_snake_case)]
pub struct SGD {
  pub name: String,
  pub learning_rate: f32,
  pub momemtum: f32,
  pub decay: f32,
  pub nesterov: bool,
  pub clip_grad: f32,
  pub iter: u64,
  velocity: Vec<Array>,
}

impl Default for SGD {
  fn default() -> SGD {
    SGD {
      name: "SGD".to_string(),
      learning_rate: 1e-3,
      momemtum: 0.0,
      decay: 0.0,
      nesterov: false,
      clip_grad: 0.0,
      iter: 0,
      velocity: Vec::new(),
    }
  }
}

impl Optimizer for SGD {
  fn new(params: &HashMap<&str, &str>) -> SGD {
    SGD{
      name: "SGD".to_string(),
      learning_rate: params.get("learning_rate").unwrap().parse::<f32>().unwrap(),
      momemtum: params.get("momemtum").unwrap().parse::<f32>().unwrap(),
      decay: params.get("decay").unwrap().parse::<f32>().unwrap(),
      nesterov: params.get("nesterov").unwrap().parse::<bool>().unwrap(),
      clip_grad: params.get("clip_grad").unwrap().parse::<f32>().unwrap(),
      iter: 0,
      velocity: Vec::new(),
    }
  }

  fn setup(&mut self, dims: Vec<Dim4>) {  //, w_dims: Vec<Dim4>, b_dims: Vec<Dim4>){
    if self.velocity.len() == 0 {
      for dim in dims {
        self.velocity.push(initializations::zeros::<f32>(dim));
      }
    }
  }

  fn update(&mut self, parameter_manager: &mut ParamManager, batch_size: u64)
  {
    self.iter += 1;
    let lr = self.learning_rate * (1.0 / (1.0 + self.decay * (self.iter as f32)));
    let alpha = lr / batch_size as f32;
    let mut running_type = DType::F32;

    // all arrays are returned as [W0, b0, .. WN, bN, ..] (note this is per layer)
    // deltas are returned in the same way
    let num_params = self.velocity.len();
    for (arr, delta, velocity, ind) in Zip::new((parameter_manager.get_all_arrays().iter()   // weights + biases
                                                 , parameter_manager.get_all_deltas().iter() // deltas of above
                                                 , self.velocity.iter_mut()                  // velocity of above
                                                 , 0..num_params))                           // current index
    {
      running_type = arr.get_type();
      let grad_update = match self.clip_grad > 0.0 {
        false => delta.clone(),
        true  => optimizer::clip_grads(&delta, self.clip_grad),
      };

      // v   = momemtum * v + learning_rate * d_w (or d_b)
      // p   = p - v
      *velocity = af::add(&af::mul(&self.momemtum, velocity, false),
                          &af::mul(&alpha, &grad_update, false), false);
      assert!(velocity.dims().get() == arr.dims().get());
      parameter_manager.set_array_from_index(af::sub(arr, velocity, false), ind);
    }

    // zero out the deltas
    parameter_manager.zero_all_deltas(running_type);
  }

  fn info(&self){
    println!("optimizer_name: {}", self.name);
    println!("learning_rate:  {}", self.learning_rate);
    println!("momemtum:       {}", self.momemtum);
    println!("decay:          {}", self.decay);
    println!("nesterov:       {}", self.nesterov);
    println!("clip_grad:      {}", self.clip_grad);
    println!("iter:           {}", self.iter);
  }
}