xref: /aosp_15_r20/external/pytorch/torch/distributed/optim/functional_rprop.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45) 
1# mypy: allow-untyped-defs
2from typing import Dict, List, Optional, Tuple
3
4import torch
5import torch.optim._functional as F
6from torch import Tensor
7
8
9__all__: List[str] = []
10
11
12# Define a TorchScript compatible Functional Rprop Optimizer
13# where we use these optimizer in a functional way.
14# Instead of using the `param.grad` when updating parameters,
15# we explicitly allow the distributed optimizer pass gradients to
16# the `step` function. In this way, we could separate the gradients
17# and parameters and allow multithreaded trainer to update the
18# parameters without data traces on accumulating to the same .grad.
19# NOTE: This should be only used by distributed optimizer internals
20# and not meant to expose to the user.
21@torch.jit.script
22class _FunctionalRprop:
23    def __init__(
24        self,
25        params: List[Tensor],
26        lr: float = 1e-2,
27        etas: Tuple[float, float] = (0.5, 1.2),
28        step_sizes: Tuple[float, float] = (1e-6, 50),
29        foreach: bool = False,
30        maximize: bool = False,
31        _allow_empty_param_list: bool = False,
32    ):
33        self.defaults = {
34            "lr": lr,
35        }
36        self.etas = etas
37        self.step_sizes = step_sizes
38        self.foreach = foreach
39        self.maximize = maximize
40
41        if len(params) == 0 and not _allow_empty_param_list:
42            raise ValueError("optimizer got an empty parameter list")
43
44        # NOTE: we only have one param_group and don't allow user to add additional
45        # param group as it's not a common use case.
46        self.param_group = {"params": params}
47
48        self.state = torch.jit.annotate(Dict[torch.Tensor, Dict[str, torch.Tensor]], {})
49
50    def step(self, gradients: List[Optional[Tensor]]):
51        params = self.param_group["params"]
52        params_with_grad = []
53        grads = []
54        prevs = []
55        step_sizes = []
56        state_steps = []
57        lr = self.defaults["lr"]
58        etaminus, etaplus = self.etas
59        step_size_min, step_size_max = self.step_sizes
60
61        if len(params) != len(gradients):
62            raise ValueError(
63                "the gradients passed in does not equal to the size of the parameters!"
64                + f"Params length: {len(params)}. "
65                + f"Gradients length: {len(gradients)}"
66            )
67
68        has_complex = False
69        for param, gradient in zip(params, gradients):
70            if gradient is not None:
71                has_complex |= torch.is_complex(param)
72                params_with_grad.append(param)
73                grads.append(gradient)
74                # Lazy state initialization
75                if param not in self.state:
76                    self.state[param] = {}
77                    state = self.state[param]
78                    state["step"] = torch.tensor(0.0)
79                    state["prev"] = torch.zeros_like(
80                        param, memory_format=torch.preserve_format
81                    )
82                    state["step_size"] = torch.full_like(gradient, lr)
83
84                state = self.state[param]
85                prevs.append(state["prev"])
86                step_sizes.append(state["step_size"])
87                state_steps.append(state["step"])
88
89        with torch.no_grad():
90            F.rprop(
91                params_with_grad,
92                grads,
93                prevs,
94                step_sizes,
95                state_steps,
96                step_size_min=step_size_min,
97                step_size_max=step_size_max,
98                etaminus=etaminus,
99                etaplus=etaplus,
100                foreach=self.foreach,
101                maximize=self.maximize,
102                has_complex=has_complex,
103            )
104