File size: 16,526 Bytes
9c6594c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
# mypy: allow-untyped-defs
from typing import Any, cast, Optional, Union

import torch
from torch import Tensor

from .optimizer import (
    _capturable_doc,
    _default_to_fused_or_foreach,
    _differentiable_doc,
    _disable_dynamo_if_unsupported,
    _foreach_doc,
    _get_capturable_supported_devices,
    _get_scalar_dtype,
    _maximize_doc,
    _params_doc,
    _use_grad_for_differentiable,
    _view_as_real,
    Optimizer,
    ParamsT,
)


__all__ = ["Adadelta", "adadelta"]


class Adadelta(Optimizer):
    def __init__(
        self,
        params: ParamsT,
        lr: Union[float, Tensor] = 1.0,
        rho: float = 0.9,
        eps: float = 1e-6,
        weight_decay: float = 0,
        foreach: Optional[bool] = None,
        *,
        capturable: bool = False,
        maximize: bool = False,
        differentiable: bool = False,
    ):
        if isinstance(lr, Tensor) and lr.numel() != 1:
            raise ValueError("Tensor lr must be 1-element")
        if not 0.0 <= lr:
            raise ValueError(f"Invalid learning rate: {lr}")
        if not 0.0 <= rho <= 1.0:
            raise ValueError(f"Invalid rho value: {rho}")
        if not 0.0 <= eps:
            raise ValueError(f"Invalid epsilon value: {eps}")
        if not 0.0 <= weight_decay:
            raise ValueError(f"Invalid weight_decay value: {weight_decay}")

        defaults = dict(
            lr=lr,
            rho=rho,
            eps=eps,
            weight_decay=weight_decay,
            maximize=maximize,
            capturable=capturable,
            foreach=foreach,
            differentiable=differentiable,
        )
        super().__init__(params, defaults)

    def __setstate__(self, state):
        super().__setstate__(state)
        for group in self.param_groups:
            group.setdefault("foreach", None)
            group.setdefault("maximize", False)
            group.setdefault("differentiable", False)
            group.setdefault("capturable", False)
            for p in group["params"]:
                p_state = self.state.get(p, [])
                if len(p_state) != 0 and not torch.is_tensor(p_state["step"]):
                    step_val = float(p_state["step"])
                    p_state["step"] = (
                        torch.tensor(
                            step_val, dtype=_get_scalar_dtype(), device=p.device
                        )
                        if group["capturable"]
                        else torch.tensor(step_val, dtype=_get_scalar_dtype())
                    )

    def _init_group(
        self,
        group: dict[str, Any],
        params_with_grad: list[Tensor],
        grads: list[Tensor],
        square_avgs: list[Tensor],
        acc_deltas: list[Tensor],
        state_steps: list[Tensor],
    ):
        has_complex = False
        p: Tensor
        for p in group["params"]:
            if p.grad is None:
                continue
            has_complex |= torch.is_complex(p)
            params_with_grad.append(p)
            if p.grad.is_sparse:
                raise RuntimeError("Adadelta does not support sparse gradients")
            grads.append(p.grad)

            state = self.state[p]

            # Lazy state initialization
            if len(state) == 0:
                state["step"] = (
                    torch.zeros((), dtype=_get_scalar_dtype(), device=p.device)
                    if group["capturable"]
                    else torch.zeros((), dtype=_get_scalar_dtype())
                )

                state["square_avg"] = torch.zeros_like(
                    p, memory_format=torch.preserve_format
                )
                state["acc_delta"] = torch.zeros_like(
                    p, memory_format=torch.preserve_format
                )

            square_avgs.append(state["square_avg"])
            acc_deltas.append(state["acc_delta"])
            state_steps.append(state["step"])

        return has_complex

    @_use_grad_for_differentiable
    def step(self, closure=None):
        """Perform a single optimization step.

        Args:
            closure (Callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        self._cuda_graph_capture_health_check()

        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            params_with_grad: list[Tensor] = []
            grads: list[Tensor] = []
            square_avgs: list[Tensor] = []
            acc_deltas: list[Tensor] = []
            state_steps: list[Tensor] = []
            (
                lr,
                rho,
                eps,
                weight_decay,
                foreach,
                maximize,
                differentiable,
                capturable,
            ) = (
                group["lr"],
                group["rho"],
                group["eps"],
                group["weight_decay"],
                group["foreach"],
                group["maximize"],
                group["differentiable"],
                group["capturable"],
            )

            has_complex = self._init_group(
                group, params_with_grad, grads, square_avgs, acc_deltas, state_steps
            )

            adadelta(
                params_with_grad,
                grads,
                square_avgs,
                acc_deltas,
                state_steps,
                lr=lr,
                rho=rho,
                eps=eps,
                weight_decay=weight_decay,
                foreach=foreach,
                maximize=maximize,
                differentiable=differentiable,
                capturable=capturable,
                has_complex=has_complex,
            )

        return loss


Adadelta.__doc__ = (
    r"""Implements Adadelta algorithm.

    .. math::
       \begin{aligned}
            &\rule{110mm}{0.4pt}                                                                 \\
            &\textbf{input}      : \gamma \text{ (lr)}, \: \theta_0 \text{ (params)},
                \: f(\theta) \text{ (objective)}, \: \rho \text{ (decay)},
                \: \lambda \text{ (weight decay)}                                                \\
            &\textbf{initialize} :  v_0  \leftarrow 0 \: \text{ (square avg)},
                \: u_0 \leftarrow 0 \: \text{ (accumulate variables)}                     \\[-1.ex]
            &\rule{110mm}{0.4pt}                                                                 \\
            &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do}                         \\
            &\hspace{5mm}g_t           \leftarrow   \nabla_{\theta} f_t (\theta_{t-1})           \\
            &\hspace{5mm}if \: \lambda \neq 0                                                    \\
            &\hspace{10mm} g_t \leftarrow g_t + \lambda  \theta_{t-1}                            \\
            &\hspace{5mm} v_t      \leftarrow v_{t-1} \rho + g^2_t (1 - \rho)                    \\
            &\hspace{5mm}\Delta x_t    \leftarrow   \frac{\sqrt{u_{t-1} +
                \epsilon }}{ \sqrt{v_t + \epsilon}  }g_t \hspace{21mm}                           \\
            &\hspace{5mm} u_t  \leftarrow   u_{t-1}  \rho +
                 \Delta x^2_t  (1 - \rho)                                                        \\
            &\hspace{5mm}\theta_t      \leftarrow   \theta_{t-1} - \gamma  \Delta x_t            \\
            &\rule{110mm}{0.4pt}                                                          \\[-1.ex]
            &\bf{return} \:  \theta_t                                                     \\[-1.ex]
            &\rule{110mm}{0.4pt}                                                          \\[-1.ex]
       \end{aligned}

    For further details regarding the algorithm we refer to `ADADELTA: An Adaptive Learning Rate Method`_.
    """
    + rf"""
    Args:
        {_params_doc}
        lr (float, Tensor, optional): coefficient that scale delta before it is applied
            to the parameters (default: 1.0)
        rho (float, optional): coefficient used for computing a running average
            of squared gradients (default: 0.9). A higher value of `rho` will
            result in a slower average, which can be helpful for preventing
            oscillations in the learning process.
        eps (float, optional): term added to the denominator to improve
            numerical stability (default: 1e-6).
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
        {_foreach_doc}
        {_capturable_doc}
        {_maximize_doc}
        {_differentiable_doc}

    .. _ADADELTA\: An Adaptive Learning Rate Method:
        https://arxiv.org/abs/1212.5701

    """
)


def _single_tensor_adadelta(
    params: list[Tensor],
    grads: list[Tensor],
    square_avgs: list[Tensor],
    acc_deltas: list[Tensor],
    state_steps: list[Tensor],
    *,
    lr: float,
    rho: float,
    eps: float,
    weight_decay: float,
    maximize: bool,
    differentiable: bool,
    capturable: bool,
    has_complex: bool,
):
    # If compiling, the compiler will handle cudagraph checks, see note [torch.compile x capturable]
    if not torch.compiler.is_compiling() and capturable:
        capturable_supported_devices = _get_capturable_supported_devices(
            supports_xla=False
        )
        assert all(
            p.device.type == step.device.type
            and p.device.type in capturable_supported_devices
            for p, step in zip(params, state_steps)
        ), f"If capturable=True, params and state_steps must be on supported devices: {capturable_supported_devices}."

    for param, grad, square_avg, acc_delta, step in zip(
        params, grads, square_avgs, acc_deltas, state_steps
    ):
        step += 1
        grad = grad if not maximize else -grad

        if weight_decay != 0:
            grad = grad.add(param, alpha=weight_decay)

        if torch.is_complex(param):
            square_avg = torch.view_as_real(square_avg)
            acc_delta = torch.view_as_real(acc_delta)
            grad = torch.view_as_real(grad)

        square_avg.mul_(rho).addcmul_(grad, grad, value=1 - rho)
        std = square_avg.add(eps).sqrt_()
        delta = acc_delta.add(eps).sqrt_()
        if differentiable:
            delta = delta.clone()
        delta.div_(std).mul_(grad)
        acc_delta.mul_(rho).addcmul_(delta, delta, value=1 - rho)

        if torch.is_complex(param):
            delta = torch.view_as_complex(delta)
        param.add_(delta, alpha=-lr)


def _multi_tensor_adadelta(
    params: list[Tensor],
    grads: list[Tensor],
    square_avgs: list[Tensor],
    acc_deltas: list[Tensor],
    state_steps: list[Tensor],
    *,
    lr: float,
    rho: float,
    eps: float,
    weight_decay: float,
    maximize: bool,
    differentiable: bool,
    capturable: bool,
    has_complex: bool,
):
    assert not differentiable, "_foreach ops don't support autograd"

    # If compiling, the compiler will handle cudagraph checks, see note [torch.compile x capturable]
    if not torch.compiler.is_compiling() and capturable:
        capturable_supported_devices = _get_capturable_supported_devices(
            supports_xla=False
        )
        assert all(
            p.device.type == step.device.type
            and p.device.type in capturable_supported_devices
            for p, step in zip(params, state_steps)
        ), f"If capturable=True, params and state_steps must be on supported devices: {capturable_supported_devices}."

    if len(params) == 0:
        return

    grouped_tensors = Optimizer._group_tensors_by_device_and_dtype(
        [params, grads, square_avgs, acc_deltas, state_steps]  # type: ignore[list-item]
    )
    for (
        device_params_,
        device_grads_,
        device_square_avgs_,
        device_acc_deltas_,
        device_state_steps_,
    ), _ in grouped_tensors.values():
        device_params = cast(list[Tensor], device_params_)
        device_grads = cast(list[Tensor], device_grads_)
        device_square_avgs = cast(list[Tensor], device_square_avgs_)
        device_acc_deltas = cast(list[Tensor], device_acc_deltas_)
        device_state_steps = cast(list[Tensor], device_state_steps_)
        if has_complex:
            _view_as_real(
                device_params, device_grads, device_square_avgs, device_acc_deltas
            )

        # Update steps
        # If steps are on CPU, foreach will fall back to the slow path, which is a for-loop calling t.add(1) over
        # and over. 1 will then be wrapped into a Tensor over and over again, which is slower than if we just
        # wrapped it once now. The alpha is required to assure we go to the right overload.
        if not torch.compiler.is_compiling() and device_state_steps[0].is_cpu:
            torch._foreach_add_(
                device_state_steps, torch.tensor(1.0, device="cpu"), alpha=1.0
            )
        else:
            torch._foreach_add_(device_state_steps, 1)

        if maximize:
            device_grads = torch._foreach_neg(device_grads)  # type: ignore[assignment]

        if weight_decay != 0:
            # Re-use the intermediate memory (device_grads) already allocated for maximize
            if maximize:
                torch._foreach_add_(device_grads, device_params, alpha=weight_decay)
            else:
                device_grads = torch._foreach_add(  # type: ignore[assignment]
                    device_grads, device_params, alpha=weight_decay
                )

        torch._foreach_mul_(device_square_avgs, rho)
        torch._foreach_addcmul_(
            device_square_avgs, device_grads, device_grads, value=1 - rho
        )

        std = torch._foreach_add(device_square_avgs, eps)
        torch._foreach_sqrt_(std)

        deltas = torch._foreach_add(device_acc_deltas, eps)
        torch._foreach_sqrt_(deltas)
        torch._foreach_div_(deltas, std)
        torch._foreach_mul_(deltas, device_grads)

        torch._foreach_mul_(device_acc_deltas, rho)
        torch._foreach_addcmul_(device_acc_deltas, deltas, deltas, value=1 - rho)

        # If LR is a tensor, the else branch will internally call item()
        # which will cause silent incorrectness if we are capturing
        if capturable and isinstance(lr, torch.Tensor):
            torch._foreach_mul_(deltas, -lr)
            torch._foreach_add_(device_params, deltas)
        else:
            torch._foreach_add_(device_params, deltas, alpha=-lr)


@_disable_dynamo_if_unsupported(single_tensor_fn=_single_tensor_adadelta)
def adadelta(
    params: list[Tensor],
    grads: list[Tensor],
    square_avgs: list[Tensor],
    acc_deltas: list[Tensor],
    state_steps: list[Tensor],
    # kwonly args with defaults are not supported by functions compiled with torchscript issue #70627
    # setting this as kwarg for now as functional API is compiled by torch/distributed/optim
    capturable: bool = False,
    foreach: Optional[bool] = None,
    differentiable: bool = False,
    has_complex: bool = False,
    *,
    lr: float,
    rho: float,
    eps: float,
    weight_decay: float,
    maximize: bool,
):
    r"""Functional API that performs Adadelta algorithm computation.

    See :class:`~torch.optim.Adadelta` for details.
    """

    # this check is slow during compilation, so we skip it
    # if it's strictly needed we can add this check back in dynamo
    if not torch.compiler.is_compiling() and not all(
        isinstance(t, torch.Tensor) for t in state_steps
    ):
        raise RuntimeError(
            "API has changed, `state_steps` argument must contain a list of singleton tensors"
        )

    # We still respect when the user inputs False for foreach.
    if foreach is None:
        _, foreach = _default_to_fused_or_foreach(
            params, differentiable, use_fused=False
        )

    if foreach and torch.jit.is_scripting():
        raise RuntimeError("torch.jit.script not supported with foreach optimizers")

    if foreach and not torch.jit.is_scripting():
        func = _multi_tensor_adadelta
    else:
        func = _single_tensor_adadelta

    func(
        params,
        grads,
        square_avgs,
        acc_deltas,
        state_steps,
        lr=lr,
        rho=rho,
        eps=eps,
        weight_decay=weight_decay,
        maximize=maximize,
        differentiable=differentiable,
        capturable=capturable,
        has_complex=has_complex,
    )