Real-time collaboration for Jupyter Notebooks, Linux Terminals, LaTeX, VS Code, R IDE, and more,
all in one place. Commercial Alternative to JupyterHub.

1
import torch
2
from torch.optim import Optimizer
3
import math
4
import apex
5
import unittest
6

7
from test_fused_optimizer import TestFusedOptimizer
8
from itertools import product
9

10
class Novograd(Optimizer):
11
    """
12
    Implements Novograd algorithm.
13

14
    Args:
15
        params (iterable): iterable of parameters to optimize or dicts defining
16
            parameter groups
17
        lr (float, optional): learning rate (default: 1e-3)
18
        betas (Tuple[float, float], optional): coefficients used for computing
19
            running averages of gradient and its square (default: (0.95, 0))
20
        eps (float, optional): term added to the denominator to improve
21
            numerical stability (default: 1e-8)
22
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
23
        grad_averaging: gradient averaging
24
        amsgrad (boolean, optional): whether to use the AMSGrad variant of this
25
            algorithm from the paper `On the Convergence of Adam and Beyond`_
26
            (default: False)
27
    """
28

29
    def __init__(self, params, lr=1e-3, betas=(0.95, 0), eps=1e-8,
30
                 weight_decay=0, grad_averaging=False, amsgrad=False):
31
        if not 0.0 <= lr:
32
            raise ValueError("Invalid learning rate: {}".format(lr))
33
        if not 0.0 <= eps:
34
            raise ValueError("Invalid epsilon value: {}".format(eps))
35
        if not 0.0 <= betas[0] < 1.0:
36
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
37
        if not 0.0 <= betas[1] < 1.0:
38
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
39
        defaults = dict(lr=lr, betas=betas, eps=eps,
40
                      weight_decay=weight_decay,
41
                      grad_averaging=grad_averaging,
42
                      amsgrad=amsgrad)
43

44
        super(Novograd, self).__init__(params, defaults)
45

46
    def __setstate__(self, state):
47
        super(Novograd, self).__setstate__(state)
48
        for group in self.param_groups:
49
            group.setdefault('amsgrad', False)
50

51
    def step(self, closure=None):
52
        """Performs a single optimization step.
53

54
        Arguments:
55
            closure (callable, optional): A closure that reevaluates the model
56
            and returns the loss.
57
        """
58
        loss = None
59
        if closure is not None:
60
            loss = closure()
61

62
        for group in self.param_groups:
63
            for p in group['params']:
64
                if p.grad is None:
65
                    continue
66
                grad = p.grad.data
67
                if grad.is_sparse:
68
                    raise RuntimeError('Sparse gradients are not supported.')
69
                amsgrad = group['amsgrad']
70

71
                state = self.state[p]
72

73
                # State initialization
74
                if len(state) == 0:
75
                    state['step'] = 0
76
                    # Exponential moving average of gradient values
77
                    state['exp_avg'] = torch.zeros_like(p.data)
78
                    # Exponential moving average of squared gradient values
79
                    state['exp_avg_sq'] = torch.zeros([]).to(state['exp_avg'].device)
80
                    if amsgrad:
81
                        # Maintains max of all exp. moving avg. of sq. grad. values
82
                        state['max_exp_avg_sq'] = torch.zeros([]).to(state['exp_avg'].device)
83

84
                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
85
                if amsgrad:
86
                    max_exp_avg_sq = state['max_exp_avg_sq']
87
                beta1, beta2 = group['betas']
88

89
                state['step'] += 1
90

91
                norm = torch.sum(torch.pow(grad, 2))
92

93
                if exp_avg_sq == 0:
94
                    exp_avg_sq.copy_(norm)
95
                else:
96
                    exp_avg_sq.mul_(beta2).add_(norm, alpha=1 - beta2)
97

98
                if amsgrad:
99
                    # Maintains the maximum of all 2nd moment running avg. till now
100
                    torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
101
                    # Use the max. for normalizing running avg. of gradient
102
                    denom = max_exp_avg_sq.sqrt().add_(group['eps'])
103
                else:
104
                    denom = exp_avg_sq.sqrt().add_(group['eps'])
105

106
                grad.div_(denom)
107
                if group['weight_decay'] != 0:
108
                    grad.add_(p.data, alpha=group['weight_decay'])
109
                if group['grad_averaging']:
110
                    grad.mul_(1 - beta1)
111
                exp_avg.mul_(beta1).add_(grad)
112

113
                p.data.add_(exp_avg, alpha=-group['lr'])
114
        
115
        return loss
116

117

118
class TestFusedNovoGrad(TestFusedOptimizer):
119

120
    def __init__(self, *args, **kwargs):
121
        super(TestFusedNovoGrad, self).__init__(*args, **kwargs)
122

123
        # The options for NovoGrad and FusedNovoGrad are very specific if they
124
        # are expected to behave the same.
125
        self.options = {'lr':1e-3, 'betas':(0.95, 0), 'eps':1e-8,
126
                 'weight_decay':0, 'grad_averaging':False, 'amsgrad':False}
127
        
128
        self.tst_options = {'lr':1e-3, 'betas':(0.95, 0), 'eps':1e-8,
129
                 'weight_decay':0, 'grad_averaging':False, 'amsgrad':False, 
130
                 'bias_correction':False, 'reg_inside_moment':True, 
131
                 'norm_type':2, 'init_zero':False, 'set_grad_none':True}
132

133
        self.ref_optim = Novograd
134
        self.fused_optim = apex.optimizers.FusedNovoGrad
135

136
    def test_float(self):
137
        self.gen_single_type_test(param_type=torch.float)
138

139
    def test_half(self):
140
        self.gen_single_type_test(param_type=torch.float16)
141

142
    @unittest.skipIf(torch.cuda.device_count()<2, "more than 1 GPU required")
143
    def test_multi_device(self):
144
        devices = ("cuda:1", "cuda:0")
145
        for current_dev, tensor_dev in product(devices, devices):
146
            with torch.cuda.device(current_dev):
147
                torch.cuda.synchronize()
148
                self.gen_single_type_test(param_type=torch.float, device=tensor_dev)
149
                
150

151
    def test_multi_params(self):
152
        sizes = [[4096, 1024], [4096], [4096, 2048], [32320, 1024], [1]]
153

154
        tensors = []
155
        for size in sizes:
156
            tensors.append(torch.rand(size, dtype=torch.float, device="cuda"))
157
        ref_param, tst_param, ref_optim, tst_optim = self.gen_param_optim(
158
            tensors, self.options, self.tst_options
159
        )
160

161
        for _ in range(self.iters):
162
            self.gen_grad(ref_param, tst_param)
163
            ref_optim.step()
164
            tst_optim.step()
165
            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param)
166
            self.assertLessEqual(max_abs_diff, self.max_abs_diff)
167
            self.assertLessEqual(max_rel_diff, self.max_rel_diff)
168

169
if __name__ == '__main__':
170
    unittest.main()
171

172