Class: Torch::Optim::ASGD

Inherits:
Optimizer show all
Defined in:
lib/torch/optim/asgd.rb

Instance Attribute Summary

Attributes inherited from Optimizer

#param_groups

Instance Method Summary collapse

Methods inherited from Optimizer

#add_param_group, #load_state_dict, #state_dict, #zero_grad

Constructor Details

#initialize(params, lr: 1e-2, lambd: 1e-4, alpha: 0.75, t0: 1e6, weight_decay: 0) ⇒ ASGD

Returns a new instance of ASGD.

Raises:

  • (ArgumentError) 


5
6
7
8
9
10
11
 
# File 'lib/torch/optim/asgd.rb', line 5

def initialize(params, lr: 1e-2, lambd: 1e-4, alpha: 0.75, t0: 1e6, weight_decay: 0)
  raise ArgumentError, "Invalid learning rate: #{lr}" if lr < 0
  raise ArgumentError, "Invalid weight_decay value: #{weight_decay}" if weight_decay < 0

  defaults = {lr: lr, lambd: lambd, alpha: alpha, t0: t0, weight_decay: weight_decay}
  super(params, defaults)
end

Instance Method Details

#step(closure = nil) ⇒ Object 



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
 
# File 'lib/torch/optim/asgd.rb', line 13

def step(closure = nil)
  loss = nil
  if closure
    loss = closure.call
  end

  @param_groups.each do |group|
    group[:params].each do |p|
      next unless p.grad
      grad = p.grad.data
      if grad.sparse?
        raise Error, "ASGD does not support sparse gradients"
      end
      state = @state[p]

      # State initialization
      if state.size == 0
        state[:step] = 0
        state[:eta] = group[:lr]
        state[:mu] = 1
        state[:ax] = Torch.zeros_like(p.data)
      end

      state[:step] += 1

      if group[:weight_decay] != 0
        grad = grad.add(p.data, alpha: group[:weight_decay])
      end

      # decay term
      p.data.mul!(1 - group[:lambd] * state[:eta])

      # update parameter
      p.data.add!(grad, alpha: -state[:eta])

      # averaging
      if state[:mu] != 1
        state[:ax].add!(p.data.sub(state[:ax]).mul(state[:mu]))
      else
        state[:ax].copy!(p.data)
      end

      # update eta and mu
      state[:eta] = (group[:lr] / ((1 + group[:lambd] * group[:lr] * state[:step]) ** group[:alpha]))
      state[:mu] = 1 / [1, state[:step] - group[:t0]].max
    end
  end

  loss
end