Class: Torch::Optim::Adadelta

Inherits:
Optimizer show all
Defined in:
lib/torch/optim/adadelta.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: 1.0, rho: 0.9, eps: 1e-6, weight_decay: 0) ⇒ Adadelta

Returns a new instance of Adadelta.

Raises:

  • (ArgumentError) 


5
6
7
8
9
10
11
12
13
 
# File 'lib/torch/optim/adadelta.rb', line 5

def initialize(params, lr: 1.0, rho: 0.9, eps: 1e-6, weight_decay: 0)
  raise ArgumentError, "Invalid learning rate: #{lr}" if lr < 0
  raise ArgumentError, "Invalid rho value: #{rho}" if rho < 0 || rho > 1
  raise ArgumentError, "Invalid epsilon value: #{eps}" if eps < 0
  raise ArgumentError, "Invalid weight_decay value: #{weight_decay}" if weight_decay < 0

  defaults = {lr: lr, rho: rho, eps: eps, weight_decay: weight_decay}
  super(params, defaults)
end

Instance Method Details

#step(closure = nil) ⇒ Object 



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
 
# File 'lib/torch/optim/adadelta.rb', line 15

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, "Adadelta does not support sparse gradients"
      end
      state = @state[p]

      if state.size == 0
        state[:step] = 0
        state[:square_avg] = Torch.zeros_like(p.data)
        state[:acc_delta] = Torch.zeros_like(p.data)
      end

      square_avg, acc_delta = state[:square_avg], state[:acc_delta]
      rho, eps = group[:rho], group[:eps]

      state[:step] += 1

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

      square_avg.mul!(rho).addcmul!(1 - rho, grad, grad)
      std = square_avg.add(eps).sqrt!
      delta = acc_delta.add(eps).sqrt!.div!(std).mul!(grad)
      p.data.add!(-group[:lr], delta)
      acc_delta.mul!(rho).addcmul!(1 - rho, delta, delta)
    end
  end

  loss
end