optimizer.py 19.7 KB
Newer Older
1
from collections import defaultdict
Q
Qiao Longfei 已提交
2

Q
Qiao Longfei 已提交
3 4 5 6 7 8
import paddle.v2.fluid.framework as framework
from paddle.v2.fluid.framework import unique_name, Program
from paddle.v2.fluid.backward import append_backward_ops
from paddle.v2.fluid.initializer import ConstantInitializer
from paddle.v2.fluid.regularizer import append_regularization_ops
from paddle.v2.fluid.layer_helper import LayerHelper
9

10
__all__ = [
11
    'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
12
    'AdamaxOptimizer', 'DecayedAdagradOptimizer'
13
]
Q
Qiao Longfei 已提交
14 15 16 17 18 19


class Optimizer(object):
    """Optimizer Base class.

    Define the common interface of an optimizer.
20 21
    User should not use this class directly,
    but need to use one of it's implementation.
Q
Qiao Longfei 已提交
22 23
    """

24 25
    def __init__(self, global_step=None):
        self._global_step = global_step
26 27 28 29 30
        # Dictionary of accumulators. Some optimizer subclasses need to
        # allocate and manage extra variables associated with the parameters
        # to train. These variables are called accumulators.
        # {accum_name : { paramter_name : accumulator_for_parameter, ...}, ...}
        self._accumulators = defaultdict(lambda: dict())
Q
Qiao Longfei 已提交
31
        self.helper = None
Q
Qiao Longfei 已提交
32 33 34 35 36 37

    def _append_optimize_op(self, block, param_and_grad):
        """ append optimize operator to block and return all the added optimize_op
        """
        raise NotImplementedError()

38 39 40 41 42 43 44 45 46 47 48 49 50 51 52
    def _create_param_lr(self, param_and_grad):
        # create learning rate variable for every parameter
        param = param_and_grad[0]
        param_lr = param.optimize_attr['learning_rate']
        param_lr_shape = [1]
        param_lr_var = self.helper.create_global_variable(
            name=unique_name("learning_rate"),
            dtype='float32',
            shape=param_lr_shape,
            lod_level=1,
            persistable=True)
        param_lr = param_lr * self._learning_rate
        self.helper.set_variable_initializer(
            var=param_lr_var, initializer=ConstantInitializer(param_lr))
        return param_lr_var
53 54 55 56 57 58 59

    def _create_accumulators(self, block, parameters):
        """Create all accumulators needed by the parameters

        Args:
            block: the block in which the loss variable is present
            parameters: list of parameter variables for the optimizer
Q
Qiao Longfei 已提交
60
        """
61 62
        pass

63 64 65 66 67 68 69 70 71 72 73 74 75
    def _finish_update(self, block):
        """Finish any custom updates needed
           before completing an optimization step

        Args:
            block: the block in which the loss variable is present
            parameters: list of parameter variables for the optimizer

        Returns:
            list of finish ops or None
        """
        pass

Q
Qiao Longfei 已提交
76
    def _add_accumulator(self, name, param, dtype=None, fill_value=0.0):
77 78 79 80 81 82 83 84 85 86 87
        """Utility function to add an accumulator for a parameter

        Args:
            block: the block in which the loss variable is present
            name: name of the accumulator
            param: parameter variable for which accumulator is to be added
            dtype: data type of the accumulator variable
            fill_value: value to initialize the accumulator variable
        """
        if (name in self._accumulators and
                param.name in self._accumulators[name]):
88
            raise Exception("Accumulator {} already exists for parameter {}".
89
                            format(name, param.name))
Q
Qiao Longfei 已提交
90 91 92 93 94

        assert isinstance(self.helper, LayerHelper)
        var = self.helper.create_global_variable(
            name=unique_name(name),
            persistable=True,
F
fengjiayi 已提交
95
            dtype=dtype or param.dtype,
Q
Qiao Longfei 已提交
96 97 98 99 100
            type=param.type,
            shape=param.shape)
        self.helper.set_variable_initializer(
            var, initializer=ConstantInitializer(value=float(fill_value)))
        self._accumulators[name][param.name] = var
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117

    def _get_accumulator(self, name, param):
        """Utility function to fetch an accumulator for a parameter

        Args:
            name: name of the accumulator
            param: parameter variable for which accumulator is to be fetched

        Returns:
            accumulator variable for the parameter
        """
        if (name not in self._accumulators or
                param.name not in self._accumulators[name]):
            raise Exception("Accumulator {} does not exist for parameter {}".
                            format(name, param.name))
        return self._accumulators[name][param.name]

118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
    def _increment_global_step(self, block):
        """Increment the global step by 1 after every iteration

        Args:
            block: the block in which the loss variable is present

        Returns:
            list with global_step increment op as its only element
        """
        assert isinstance(block, framework.Block)
        assert self._global_step is not None
        # create the increment op
        increment_op = block.append_op(
            type="increment",
            inputs={"X": self._global_step},
            outputs={"Out": self._global_step},
            attrs={"step": 1.0})

        return increment_op

Q
Qiao Longfei 已提交
138 139 140
    def create_optimization_pass(self,
                                 parameters_and_grads,
                                 loss,
141
                                 startup_program=None):
Q
Qiao Longfei 已提交
142 143 144 145 146 147 148
        """Add optimization operators to update gradients to variables.

        Args:
          loss: the target that this optimization is for.
          parameters_and_grads: a list of (variable, gradient) pair to update.

        Returns:
149 150 151 152
          return_op_list: a list of operators that will complete one step of
          optimization. This will include parameter update ops, global step
          update ops and any other custom ops required by subclasses to manage
          their internal state.
153
          :param startup_program: 
Q
Qiao Longfei 已提交
154
        """
155 156 157 158 159
        # This is a default implementation of create_optimization_pass that
        # can be shared by most optimizers. This implementation assumes that
        # the subclass will implement the _append_optimize_op method and the
        #  _initialize_tensors method. The subclass can extend the
        # _create_accumulators method if it needs to create accumulators
160
        # for parameters and extend _finish_update method to add custom ops.
161 162

        # Create any accumulators
Q
Qiao Longfei 已提交
163 164
        program = loss.block.program
        self.helper = LayerHelper(
165 166 167
            self.__class__.__name__,
            main_program=program,
            startup_program=startup_program)
168 169 170
        self._create_accumulators(loss.block,
                                  [p[0] for p in parameters_and_grads])

Q
Qiao Longfei 已提交
171 172
        optimize_ops = []
        for param_and_grad in parameters_and_grads:
Q
Qiao Longfei 已提交
173 174
            if param_and_grad[0].trainable is True and param_and_grad[
                    1] is not None:
Q
Qiao Longfei 已提交
175 176 177
                optimize_op = self._append_optimize_op(loss.block,
                                                       param_and_grad)
                optimize_ops.append(optimize_op)
178

179 180 181 182 183 184 185 186 187 188
        # Returned list of ops can include more ops in addition
        # to optimization ops
        return_ops = optimize_ops

        # Get custom finish ops for subclasses
        # FIXME: Need to fix this once we figure out how to handle dependencies
        finish_ops = self._finish_update(loss.block)
        if finish_ops is not None:
            return_ops += finish_ops

189 190
        if self._global_step is not None:
            return_ops.append(self._increment_global_step(loss.block))
191
        return return_ops
Q
Qiao Longfei 已提交
192

Q
Qiao Longfei 已提交
193 194
    def minimize(self,
                 loss,
195
                 startup_program=None,
Q
Qiao Longfei 已提交
196 197
                 parameter_list=None,
                 no_grad_set=None):
Q
Qiao Longfei 已提交
198 199
        """Add operations to minimize `loss` by updating `parameter_list`.

200
        This method combines interface `append_backward_ops()` and
Q
Qiao Longfei 已提交
201 202
        `create_optimization_pass()` into one.
        """
203 204
        params_grads = append_backward_ops(loss, parameter_list, no_grad_set or
                                           set())
F
fengjiayi 已提交
205
        # Add regularization if any
206
        params_grads = append_regularization_ops(params_grads)
Q
Qiao Longfei 已提交
207
        optimize_ops = self.create_optimization_pass(params_grads, loss,
208
                                                     startup_program)
Q
Qiao Longfei 已提交
209 210 211 212 213 214 215
        return optimize_ops


class SGDOptimizer(Optimizer):
    """ Simple SGD optimizer without any state.
    """

216
    def __init__(self, learning_rate, global_step=None):
Q
Qiao Longfei 已提交
217
        assert learning_rate is not None
218
        super(SGDOptimizer, self).__init__(global_step)
Q
Qiao Longfei 已提交
219 220 221
        self.type = "sgd"
        self._learning_rate = learning_rate

222 223
    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
224

Q
Qiao Longfei 已提交
225 226 227 228 229 230
        # create the optimize op
        sgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
231
                "LearningRate": self._create_param_lr(param_and_grad)
Q
Qiao Longfei 已提交
232
            },
233
            outputs={"ParamOut": param_and_grad[0]})
Q
Qiao Longfei 已提交
234 235

        return sgd_op
236 237 238 239 240 241 242


class MomentumOptimizer(Optimizer):
    """Simple Momentum optimizer with velocity state
    """
    _velocity_acc_str = "velocity"

243 244 245 246 247
    def __init__(self,
                 learning_rate,
                 momentum,
                 use_nesterov=False,
                 global_step=None):
248 249
        assert learning_rate is not None
        assert momentum is not None
250
        super(MomentumOptimizer, self).__init__(global_step)
251 252 253
        self.type = "momentum"
        self._learning_rate = learning_rate
        self._momentum = momentum
254
        self._use_nesterov = bool(use_nesterov)
255 256 257 258 259

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        for p in parameters:
Q
Qiao Longfei 已提交
260
            self._add_accumulator(self._velocity_acc_str, p)
261 262 263 264 265 266 267 268 269 270 271 272 273

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        velocity_acc = self._get_accumulator(self._velocity_acc_str,
                                             param_and_grad[0])
        # create the momentum optimize op
        momentum_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Velocity": velocity_acc,
274
                "LearningRate": self._create_param_lr(param_and_grad)
275 276 277 278 279
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "VelocityOut": velocity_acc
            },
280
            attrs={"mu": self._momentum,
281
                   "use_nesterov": self._use_nesterov})
282 283

        return momentum_op
284 285 286 287 288 289 290


class AdagradOptimizer(Optimizer):
    """Simple Adagrad optimizer with moment state
    """
    _moment_acc_str = "moment"

291
    def __init__(self, learning_rate, epsilon=1.0e-6, global_step=None):
292 293
        assert learning_rate is not None
        assert epsilon is not None
294
        super(AdagradOptimizer, self).__init__(global_step)
295 296 297 298 299 300 301 302
        self.type = "adagrad"
        self._learning_rate = learning_rate
        self._epsilon = epsilon

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        for p in parameters:
Q
Qiao Longfei 已提交
303
            self._add_accumulator(self._moment_acc_str, p)
304 305 306 307 308 309 310

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment_acc = self._get_accumulator(self._moment_acc_str,
                                           param_and_grad[0])

311
        # Create the adagrad optimizer op
312 313 314 315 316 317
        adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
318
                "LearningRate": self._create_param_lr(param_and_grad)
319 320 321 322 323 324
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
            attrs={"epsilon": self._epsilon})

        return adagrad_op
325 326 327 328 329 330 331 332 333 334 335 336


class AdamOptimizer(Optimizer):
    """Implements the Adam Optimizer
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
337 338
                 epsilon=1e-8,
                 global_step=None):
339 340 341 342
        assert learning_rate is not None
        assert beta1 is not None
        assert beta2 is not None
        assert epsilon is not None
343
        super(AdamOptimizer, self).__init__(global_step)
344 345 346 347 348 349 350 351 352
        self.type = "adam"
        self._learning_rate = learning_rate
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

Q
Qiao Longfei 已提交
353
        main_block = block.program.global_block()
354 355
        # Create beta1 and beta2 power tensors
        beta_shape = [1]
Q
Qiao Longfei 已提交
356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373
        self._beta1_pow_acc = self.helper.create_global_variable(
            name=unique_name('beta1_pow_acc'),
            dtype='float32',
            shape=beta_shape,
            lod_level=0,
            persistable=True)
        self.helper.set_variable_initializer(
            self._beta1_pow_acc, initializer=ConstantInitializer(self._beta1))

        self._beta2_pow_acc = self.helper.create_global_variable(
            name=unique_name('beta2_pow_acc'),
            dtype='float32',
            shape=beta_shape,
            lod_level=0,
            persistable=True)

        self.helper.set_variable_initializer(
            self._beta2_pow_acc, initializer=ConstantInitializer(self._beta2))
374 375 376

        # Create accumulator tensors for first and second moments
        for p in parameters:
Q
Qiao Longfei 已提交
377 378
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
379 380 381 382 383 384 385 386

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment1 = self._get_accumulator(self._moment1_acc_str,
                                        param_and_grad[0])
        moment2 = self._get_accumulator(self._moment2_acc_str,
                                        param_and_grad[0])
387
        # create the adam optimize op
388 389 390 391 392
        adam_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
393
                "LearningRate": self._create_param_lr(param_and_grad),
394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415
                "Moment1": moment1,
                "Moment2": moment2,
                "Beta1Pow": self._beta1_pow_acc,
                "Beta2Pow": self._beta2_pow_acc
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
                "epsilon": self._epsilon
            })

        return adam_op

    def _finish_update(self, block):
        """Update Beta1 and Beta2 Power accumulators
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
416 417
        main_block = block.program.global_block()
        scale_beta1 = main_block.append_op(
418 419 420 421 422
            type="scale",
            inputs={"X": self._beta1_pow_acc},
            outputs={"Out": self._beta1_pow_acc},
            attrs={"scale": self._beta1})

Q
Qiao Longfei 已提交
423
        scale_beta2 = main_block.append_op(
424 425 426 427 428 429
            type="scale",
            inputs={"X": self._beta2_pow_acc},
            outputs={"Out": self._beta2_pow_acc},
            attrs={"scale": self._beta2})

        return [scale_beta1, scale_beta2]
430 431 432 433 434 435 436 437 438 439 440 441


class AdamaxOptimizer(Optimizer):
    """Implements the Adamax Optimizer
    """
    _moment_acc_str = "moment"
    _inf_norm_acc_str = "inf_norm"

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
442 443
                 epsilon=1e-8,
                 global_step=None):
444 445 446 447 448 449 450 451 452 453 454 455 456 457
        assert learning_rate is not None
        assert beta1 is not None
        assert beta2 is not None
        assert epsilon is not None
        super(AdamaxOptimizer, self).__init__()
        self.type = "adamax"
        self._learning_rate = learning_rate
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon

    def _create_accumulators(self, block, parameters):
        # Create beta1 power accumulator tensor
        beta_shape = [1]
Q
Qiao Longfei 已提交
458 459 460 461 462 463 464 465
        self._beta1_pow_acc = self.helper.create_global_variable(
            name=unique_name('beta1_pow_acc'),
            dtype='float32',
            shape=beta_shape,
            lod_level=0,
            persistable=True)
        self.helper.set_variable_initializer(
            self._beta1_pow_acc, initializer=ConstantInitializer(self._beta1))
466 467 468

        # Create accumulator tensors for first moment and infinity norm
        for p in parameters:
Q
Qiao Longfei 已提交
469 470
            self._add_accumulator(self._moment_acc_str, p)
            self._add_accumulator(self._inf_norm_acc_str, p)
471 472 473 474 475 476 477 478 479 480 481 482 483

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment = self._get_accumulator(self._moment_acc_str, param_and_grad[0])
        inf_norm = self._get_accumulator(self._inf_norm_acc_str,
                                         param_and_grad[0])
        # create the adamax optimize op
        adamax_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
484
                "LearningRate": self._create_param_lr(param_and_grad),
485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505
                "Moment": moment,
                "InfNorm": inf_norm,
                "Beta1Pow": self._beta1_pow_acc
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": moment,
                "InfNormOut": inf_norm
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
                "epsilon": self._epsilon
            })

        return adamax_op

    def _finish_update(self, block):
        """Update Beta1 Power accumulator
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
506 507
        main_block = block.program.global_block()
        scale_beta1 = main_block.append_op(
508 509 510 511 512 513
            type="scale",
            inputs={"X": self._beta1_pow_acc},
            outputs={"Out": self._beta1_pow_acc},
            attrs={"scale": self._beta1})

        return [scale_beta1]
514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561


class DecayedAdagradOptimizer(Optimizer):
    """Simple Decayed Adagrad optimizer with moment state
    """
    _moment_acc_str = "moment"

    def __init__(self,
                 learning_rate,
                 decay=0.95,
                 epsilon=1.0e-6,
                 global_step=None):
        assert learning_rate is not None
        assert decay is not None
        assert epsilon is not None

        super(DecayedAdagradOptimizer, self).__init__(global_step)
        self.type = "decayed_adagrad"
        self._learning_rate = learning_rate
        self._decay = decay
        self._epsilon = epsilon

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        for p in parameters:
            self._add_accumulator(self._moment_acc_str, p)

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment_acc = self._get_accumulator(self._moment_acc_str,
                                           param_and_grad[0])

        # Create the decayed adagrad optimizer op
        decayed_adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
                "LearningRate": self._create_param_lr(param_and_grad)
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
            attrs={"epsilon": self._epsilon})

        return decayed_adagrad_op