tensorflow/tensorflow/python/training/optimizer.py at master · codeplaysoftware/tensorflow · GitHub

Repository navigation

History

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

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

479

480

481

482

483

484

485

486

487

488

489

490

491

492

493

494

495

496

497

498

499

500

501

502

503

504

505

506

507

508

509

510

511

512

513

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

562

563

564

565

566

567

568

569

570

571

572

573

574

575

576

577

578

579

580

581

582

583

584

585

586

587

588

589

590

591

592

593

594

595

596

597

598

599

600

601

602

603

604

605

606

607

608

609

610

611

612

613

614

615

616

617

618

619

620

621

622

623

624

625

626

627

628

629

630

631

632

633

634

635

636

637

638

639

640

641

642

643

644

645

646

647

648

649

650

651

652

653

654

655

656

657

658

659

660

661

662

663

664

665

666

667

668

669

670

671

672

673

674

675

676

677

678

679

680

681

682

683

684

685

686

687

688

689

690

691

692

693

694

695

696

697

698

699

700

701

702

703

704

705

706

707

708

709

710

711

712

713

714

715

716

717

718

719

720

721

722

723

724

725

726

727

728

729

730

731

732

733

734

735

736

737

738

739

740

741

742

743

744

745

746

747

748

749

750

751

752

753

754

755

756

757

758

759

760

761

762

763

764

765

766

767

768

769

770

771

772

773

774

775

776

777

778

779

780

781

782

783

784

785

786

787

788

789

790

791

792

793

794

795

796

797

798

799

800

801

802

803

804

805

806

807

808

809

810

811

812

813

814

815

816

817

818

819

820

821

822

823

824

825

826

827

828

829

830

831

832

833

834

835

836

837

838

839

840

841

842

843

844

845

846

847

848

849

850

851

852

853

854

855

856

857

858

859

860

861

862

863

864

865

866

867

868

869

870

871

872

873

874

875

876

877

878

879

880

881

882

883

884

885

886

887

888

889

890

891

892

893

894

895

896

897

898

899

900

901

902

903

904

905

906

907

908

909

910

911

912

913

914

915

916

917

918

919

920

921

922

923

924

925

926

927

928

929

930

931

932

933

934

935

936

937

938

939

940

941

942

943

944

945

946

947

948

949

950

951

952

953

954

955

956

957

958

959

960

961

962

963

964

965

966

967

968

969

970

971

972

973

974

975

976

977

978

979

980

981

982

983

984

985

986

987

988

989

990

991

992

993

994

995

996

997

998

999

1000

# Licensed under the Apache License, Version 2.0 (the "License");

# you may not use this file except in compliance with the License.

# You may obtain a copy of the License at

# http://www.apache.org/licenses/LICENSE-2.0

# Unless required by applicable law or agreed to in writing, software

# distributed under the License is distributed on an "AS IS" BASIS,

# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

# See the License for the specific language governing permissions and

# limitations under the License.

# ==============================================================================

"""Base class for optimizers."""

# pylint: disable=g-bad-name

from __future__ import absolute_import

from __future__ import division

from __future__ import print_function

import abc

from tensorflow.python.eager import backprop

from tensorflow.python.eager import context

from tensorflow.python.framework import dtypes

from tensorflow.python.framework import ops

from tensorflow.python.ops import array_ops

from tensorflow.python.ops import control_flow_ops

from tensorflow.python.ops import gradients

from tensorflow.python.ops import math_ops

from tensorflow.python.ops import resource_variable_ops

from tensorflow.python.ops import state_ops

from tensorflow.python.ops import variable_scope

from tensorflow.python.ops import variables

from tensorflow.python.training import distribute as distribute_lib

from tensorflow.python.training import slot_creator

from tensorflow.python.training.checkpointable import base as checkpointable

from tensorflow.python.util import nest

from tensorflow.python.util.tf_export import tf_export

def get_filtered_grad_fn(grad_fn):

# `distributed_context.join()` requires that its arguments are parallel

# across threads, and in particular that `grads_and_vars` has the same

# variables in the same order.

# When computing gradients in eager mode with multiple threads, you

# can get extra variables with a gradient of `None`. This happens when

# those variables are accessed in another thread during the gradient

# computation. To get a consistent set of variables, we filter out

# those with `None` gradients.

def filtered_grad_fn(x=None):

return [(g, v) for g, v in grad_fn(x) if g is not None]

return filtered_grad_fn

def _deduplicate_indexed_slices(values, indices):

"""Sums `values` associated with any non-unique `indices`.

Args:

values: A `Tensor` with rank >= 1.

indices: A one-dimensional integer `Tensor`, indexing into the first

dimension of `values` (as in an IndexedSlices object).

Returns:

A tuple of (`summed_values`, `unique_indices`) where `unique_indices` is a

de-duplicated version of `indices` and `summed_values` contains the sum of

`values` slices associated with each unique index.

"""

unique_indices, new_index_positions = array_ops.unique(indices)

summed_values = math_ops.unsorted_segment_sum(

values, new_index_positions,

array_ops.shape(unique_indices)[0])

return (summed_values, unique_indices)

def _var_key(var):

if context.executing_eagerly():

return var._unique_id # pylint: disable=protected-access

return (var.op.graph, var.op.name)

class _OptimizableVariable(object):

"""Interface for abstracting over variables in the optimizers."""

@abc.abstractmethod

def target(self):

"""Returns the optimization target for this variable."""

raise NotImplementedError("Calling an abstract method.")

@abc.abstractmethod

def update_op(self, optimizer, g):

"""Returns the update ops for updating the variable."""

raise NotImplementedError("Calling an abstract method.")

class _RefVariableProcessor(_OptimizableVariable):

"""Processor for Variable."""

def __init__(self, v):

self._v = v

def __str__(self):

return "<_RefVariableProcessor(%s)>" % self._v

def target(self):

return self._v._ref() # pylint: disable=protected-access

def update_op(self, optimizer, g):

if isinstance(g, ops.Tensor):

update_op = optimizer._apply_dense(g, self._v) # pylint: disable=protected-access

if self._v.constraint is not None:

with ops.control_dependencies([update_op]):

return self._v.assign(self._v.constraint(self._v))

else:

return update_op

else:

assert isinstance(g, ops.IndexedSlices), ("Gradient ", g, " is neither a "

"tensor nor IndexedSlices.")

if self._v.constraint is not None:

raise RuntimeError(

"Cannot use a constraint function on a sparse variable.")

# pylint: disable=protected-access

return optimizer._apply_sparse_duplicate_indices(g, self._v)

class _DenseReadResourceVariableProcessor(_OptimizableVariable):

"""Processor for dense ResourceVariables."""

def __init__(self, v):

self._v = v

def target(self):

return self._v

def update_op(self, optimizer, g):

# pylint: disable=protected-access

update_op = optimizer._resource_apply_dense(g, self._v.op.inputs[0])

if self._v.constraint is not None:

with ops.control_dependencies([update_op]):

return self._v.assign(self._v.constraint(self._v))

else:

return update_op

class _DenseResourceVariableProcessor(_OptimizableVariable):

"""Processor for dense ResourceVariables."""

def __init__(self, v):

self._v = v

def target(self):

return self._v

def update_op(self, optimizer, g):

# pylint: disable=protected-access

if isinstance(g, ops.IndexedSlices):

if self._v.constraint is not None:

raise RuntimeError(

"Cannot use a constraint function on a sparse variable.")

return optimizer._resource_apply_sparse_duplicate_indices(

g.values, self._v, g.indices)

update_op = optimizer._resource_apply_dense(g, self._v)

if self._v.constraint is not None:

with ops.control_dependencies([update_op]):

return self._v.assign(self._v.constraint(self._v))

else:

return update_op

class _TensorProcessor(_OptimizableVariable):

"""Processor for ordinary Tensors.

Even though a Tensor can't really be updated, sometimes it is useful to

compute the gradients with respect to a Tensor using the optimizer. Updating

the Tensor is, of course, unsupported.

"""

def __init__(self, v):

self._v = v

def target(self):

return self._v

def update_op(self, optimizer, g):

raise NotImplementedError("Trying to update a Tensor ", self._v)

def _get_processor(v):

"""The processor of v."""

if context.executing_eagerly():

if isinstance(v, ops.Tensor):

return _TensorProcessor(v)

else:

return _DenseResourceVariableProcessor(v)

if isinstance(

v, resource_variable_ops.ResourceVariable) and not v._in_graph_mode: # pylint: disable=protected-access

# True if and only if `v` was initialized eagerly.

return _DenseResourceVariableProcessor(v)

if v.op.type == "VarHandleOp":

return _DenseResourceVariableProcessor(v)

if isinstance(v, variables.Variable):

return _RefVariableProcessor(v)

if isinstance(v, ops.Tensor):

return _TensorProcessor(v)

raise NotImplementedError("Trying to optimize unsupported type ", v)

@tf_export("train.Optimizer")

class Optimizer(

# Optimizers inherit from CheckpointableBase rather than Checkpointable

# since they do most of their dependency management themselves (slot

# variables are special-cased, and non-slot variables are keyed to graphs).

checkpointable.CheckpointableBase):

"""Base class for optimizers.

This class defines the API to add Ops to train a model. You never use this

class directly, but instead instantiate one of its subclasses such as

`GradientDescentOptimizer`, `AdagradOptimizer`, or `MomentumOptimizer`.

### Usage

```python

# Create an optimizer with the desired parameters.

opt = GradientDescentOptimizer(learning_rate=0.1)

# Add Ops to the graph to minimize a cost by updating a list of variables.

# "cost" is a Tensor, and the list of variables contains tf.Variable

# objects.

opt_op = opt.minimize(cost, var_list=<list of variables>)

```

In the training program you will just have to run the returned Op.

```python

# Execute opt_op to do one step of training:

opt_op.run()

```

### Processing gradients before applying them.

Calling `minimize()` takes care of both computing the gradients and

applying them to the variables. If you want to process the gradients

before applying them you can instead use the optimizer in three steps:

1. Compute the gradients with `compute_gradients()`.

2. Process the gradients as you wish.

3. Apply the processed gradients with `apply_gradients()`.

Example:

```python

# Create an optimizer.

opt = GradientDescentOptimizer(learning_rate=0.1)

# Compute the gradients for a list of variables.

grads_and_vars = opt.compute_gradients(loss, <list of variables>)

# grads_and_vars is a list of tuples (gradient, variable). Do whatever you

# need to the 'gradient' part, for example cap them, etc.

capped_grads_and_vars = [(MyCapper(gv[0]), gv[1]) for gv in grads_and_vars]

# Ask the optimizer to apply the capped gradients.

opt.apply_gradients(capped_grads_and_vars)

```

### Gating Gradients

Both `minimize()` and `compute_gradients()` accept a `gate_gradients`

argument that controls the degree of parallelism during the application of

the gradients.

The possible values are: `GATE_NONE`, `GATE_OP`, and `GATE_GRAPH`.

`GATE_NONE`: Compute and apply gradients in parallel. This provides

the maximum parallelism in execution, at the cost of some non-reproducibility

in the results. For example the two gradients of `matmul` depend on the input

values: With `GATE_NONE` one of the gradients could be applied to one of the

inputs _before_ the other gradient is computed resulting in non-reproducible

results.

`GATE_OP`: For each Op, make sure all gradients are computed before

they are used. This prevents race conditions for Ops that generate gradients

for multiple inputs where the gradients depend on the inputs.

`GATE_GRAPH`: Make sure all gradients for all variables are computed

before any one of them is used. This provides the least parallelism but can

be useful if you want to process all gradients before applying any of them.

### Slots

Some optimizer subclasses, such as `MomentumOptimizer` and `AdagradOptimizer`

allocate and manage additional variables associated with the variables to

train. These are called Slots. Slots have names and you can ask the

optimizer for the names of the slots that it uses. Once you have a slot name

you can ask the optimizer for the variable it created to hold the slot value.

This can be useful if you want to log debug a training algorithm, report stats

about the slots, etc.

"""

# Values for gate_gradients.

GATE_NONE = 0

GATE_OP = 1

GATE_GRAPH = 2

def __init__(self, use_locking, name):

"""Create a new Optimizer.

This must be called by the constructors of subclasses.

Args:

use_locking: Bool. If True apply use locks to prevent concurrent updates

to variables.

for the optimizer.

Raises:

ValueError: If name is malformed.

"""

if not name:

raise ValueError("Must specify the optimizer name")

self._use_locking = use_locking

self._name = name

# Dictionary of slots.

# {slot_name :

# {_var_key(variable_to_train): slot_for_the_variable, ... },

# ... }

self._slots = {}

self._non_slot_dict = {}

# For implementing Checkpointable. Stores information about how to restore

# slot variables which have not yet been created

# (checkpointable._CheckpointPosition objects).

# {slot_name :

# {_var_key(variable_to_train): [checkpoint_position, ... ], ... },

# ... }

self._deferred_slot_restorations = {}

# TODO(isaprykin): When using a DistributionStrategy, and when an

# optimizer is created in each tower, it might be dangerous to

# rely on some Optimer methods. When such methods are called on a

# per-tower optimizer, an exception needs to be thrown. We do

# allow creation per-tower optimizers however, because the

# compute_gradients()->apply_gradients() sequence is safe.

def get_name(self):

return self._name

def minimize(self, loss, global_step=None, var_list=None,

gate_gradients=GATE_OP, aggregation_method=None,

colocate_gradients_with_ops=False, name=None,

grad_loss=None):

"""Add operations to minimize `loss` by updating `var_list`.

This method simply combines calls `compute_gradients()` and

`apply_gradients()`. If you want to process the gradient before applying

them call `compute_gradients()` and `apply_gradients()` explicitly instead

of using this function.

Args:

loss: A `Tensor` containing the value to minimize.

global_step: Optional `Variable` to increment by one after the

variables have been updated.

var_list: Optional list or tuple of `Variable` objects to update to

minimize `loss`. Defaults to the list of variables collected in

the graph under the key `GraphKeys.TRAINABLE_VARIABLES`.

gate_gradients: How to gate the computation of gradients. Can be

`GATE_NONE`, `GATE_OP`, or `GATE_GRAPH`.

aggregation_method: Specifies the method used to combine gradient terms.

Valid values are defined in the class `AggregationMethod`.

colocate_gradients_with_ops: If True, try colocating gradients with

the corresponding op.

grad_loss: Optional. A `Tensor` holding the gradient computed for `loss`.

Returns:

An Operation that updates the variables in `var_list`. If `global_step`

was not `None`, that operation also increments `global_step`.

Raises:

ValueError: If some of the variables are not `Variable` objects.

@compatibility(eager)

When eager execution is enabled, `loss` should be a Python function that

takes elements of `var_list` as arguments and computes the value to be

minimized. If `var_list` is None, `loss` should take no arguments.

Minimization (and gradient computation) is done with respect to the

elements of `var_list` if not None, else with respect to any trainable

variables created during the execution of the `loss` function.

`gate_gradients`, `aggregation_method`, `colocate_gradients_with_ops` and

`grad_loss` are ignored when eager execution is enabled.

@end_compatibility

"""

grads_and_vars = self.compute_gradients(

loss, var_list=var_list, gate_gradients=gate_gradients,

aggregation_method=aggregation_method,

colocate_gradients_with_ops=colocate_gradients_with_ops,

grad_loss=grad_loss)

vars_with_grad = [v for g, v in grads_and_vars if g is not None]

if not vars_with_grad:

raise ValueError(

"No gradients provided for any variable, check your graph for ops"

" that do not support gradients, between variables %s and loss %s." %

([str(v) for _, v in grads_and_vars], loss))

return self.apply_gradients(grads_and_vars, global_step=global_step,

name=name)

def compute_gradients(self, loss, var_list=None,

gate_gradients=GATE_OP,

aggregation_method=None,

colocate_gradients_with_ops=False,

grad_loss=None):

"""Compute gradients of `loss` for the variables in `var_list`.

This is the first part of `minimize()`. It returns a list

of (gradient, variable) pairs where "gradient" is the gradient

for "variable". Note that "gradient" can be a `Tensor`, an

`IndexedSlices`, or `None` if there is no gradient for the

given variable.

Args:

loss: A Tensor containing the value to minimize or a callable taking

no arguments which returns the value to minimize. When eager execution

is enabled it must be a callable.

var_list: Optional list or tuple of `tf.Variable` to update to minimize

`loss`. Defaults to the list of variables collected in the graph

under the key `GraphKeys.TRAINABLE_VARIABLES`.

gate_gradients: How to gate the computation of gradients. Can be

`GATE_NONE`, `GATE_OP`, or `GATE_GRAPH`.

aggregation_method: Specifies the method used to combine gradient terms.

Valid values are defined in the class `AggregationMethod`.

colocate_gradients_with_ops: If True, try colocating gradients with

the corresponding op.

grad_loss: Optional. A `Tensor` holding the gradient computed for `loss`.

Returns:

A list of (gradient, variable) pairs. Variable is always present, but

gradient can be `None`.

Raises:

TypeError: If `var_list` contains anything else than `Variable` objects.

ValueError: If some arguments are invalid.

RuntimeError: If called with eager execution enabled and `loss` is

not callable.

@compatibility(eager)

When eager execution is enabled, `gate_gradients`, `aggregation_method`,

and `colocate_gradients_with_ops` are ignored.

@end_compatibility

"""

if callable(loss):

with backprop.GradientTape() as tape:

if var_list is not None:

tape.watch(var_list)

loss_value = loss()

# Scale loss if using a "mean" loss reduction and multiple towers.

# Have to be careful to call distribute_lib.get_loss_reduction()

# *after* loss() is evaluated, so we know what loss reduction it uses.

# TODO(josh11b): Test that we handle weight decay in a reasonable way.

if distribute_lib.get_loss_reduction() == "mean":

num_towers = distribute_lib.get_distribution_strategy().num_towers

if num_towers > 1:

loss_value *= (1. / num_towers)

if var_list is None:

var_list = tape.watched_variables()

grads = tape.gradient(loss_value, var_list, grad_loss)

return list(zip(grads, var_list))

# Non-callable/Tensor loss case

if context.executing_eagerly():

raise RuntimeError(

"`loss` passed to Optimizer.compute_gradients should "

"be a function when eager execution is enabled.")

# Scale loss if using a "mean" loss reduction and multiple towers.

if distribute_lib.get_loss_reduction() == "mean":

num_towers = distribute_lib.get_distribution_strategy().num_towers

if num_towers > 1:

loss *= (1. / num_towers)

if gate_gradients not in [Optimizer.GATE_NONE, Optimizer.GATE_OP,

Optimizer.GATE_GRAPH]:

raise ValueError("gate_gradients must be one of: Optimizer.GATE_NONE, "

"Optimizer.GATE_OP, Optimizer.GATE_GRAPH. Not %s" %

gate_gradients)

self._assert_valid_dtypes([loss])

if grad_loss is not None:

self._assert_valid_dtypes([grad_loss])

if var_list is None:

var_list = (

variables.trainable_variables() +

ops.get_collection(ops.GraphKeys.TRAINABLE_RESOURCE_VARIABLES))

else:

var_list = nest.flatten(var_list)

# pylint: disable=protected-access

var_list += ops.get_collection(ops.GraphKeys._STREAMING_MODEL_PORTS)

# pylint: enable=protected-access

processors = [_get_processor(v) for v in var_list]

if not var_list:

raise ValueError("No variables to optimize.")

var_refs = [p.target() for p in processors]

grads = gradients.gradients(

loss, var_refs, grad_ys=grad_loss,

gate_gradients=(gate_gradients == Optimizer.GATE_OP),

aggregation_method=aggregation_method,

colocate_gradients_with_ops=colocate_gradients_with_ops)

if gate_gradients == Optimizer.GATE_GRAPH:

grads = control_flow_ops.tuple(grads)

grads_and_vars = list(zip(grads, var_list))

self._assert_valid_dtypes(

[v for g, v in grads_and_vars

if g is not None and v.dtype != dtypes.resource])

return grads_and_vars

def apply_gradients(self, grads_and_vars, global_step=None, name=None):

"""Apply gradients to variables.

This is the second part of `minimize()`. It returns an `Operation` that

applies gradients.

Args:

grads_and_vars: List of (gradient, variable) pairs as returned by

`compute_gradients()`.

global_step: Optional `Variable` to increment by one after the

variables have been updated.

name passed to the `Optimizer` constructor.

Returns:

An `Operation` that applies the specified gradients. If `global_step`

was not None, that operation also increments `global_step`.

Raises:

TypeError: If `grads_and_vars` is malformed.

ValueError: If none of the variables have gradients.

RuntimeError: If you should use `_distributed_apply()` instead.

"""

# This is a default implementation of apply_gradients() that can be shared

# by most optimizers. It relies on the subclass implementing the following

# methods: _create_slots(), _prepare(), _apply_dense(), and _apply_sparse().

# Handle DistributionStrategy case.

if distribute_lib.get_cross_tower_context():

raise RuntimeError("Use `_distributed_apply()` instead of "

"`apply_gradients()` in a cross-tower context.")

# TODO(isaprykin): Get rid of `has_distribution_strategy()` check by

# always calling _distributed_apply(), using the default distribution

# as needed.

if distribute_lib.has_distribution_strategy():

grads_and_vars = get_filtered_grad_fn(lambda _: grads_and_vars)()

return distribute_lib.get_tower_context().merge_call(

self._distributed_apply, grads_and_vars, global_step, name)

# No DistributionStrategy case.

grads_and_vars = tuple(grads_and_vars) # Make sure repeat iteration works.

if not grads_and_vars:

raise ValueError("No variables provided.")

converted_grads_and_vars = []

for g, v in grads_and_vars:

if g is not None:

try:

# Convert the grad to Tensor or IndexedSlices if necessary.

g = ops.convert_to_tensor_or_indexed_slices(g)

except TypeError:

raise TypeError(

"Gradient must be convertible to a Tensor"

" or IndexedSlices, or None: %s" % g)

if not isinstance(g, (ops.Tensor, ops.IndexedSlices)):

raise TypeError(

"Gradient must be a Tensor, IndexedSlices, or None: %s" % g)

p = _get_processor(v)

converted_grads_and_vars.append((g, v, p))

converted_grads_and_vars = tuple(converted_grads_and_vars)

var_list = [v for g, v, _ in converted_grads_and_vars if g is not None]

if not var_list:

raise ValueError("No gradients provided for any variable: %s." %

([str(v) for _, _, v in converted_grads_and_vars],))

with ops.init_scope():

self._create_slots(var_list)

update_ops = []

with ops.name_scope(name, self._name) as name:

self._prepare()

for grad, var, processor in converted_grads_and_vars:

if grad is None:

continue

# We colocate all ops created in _apply_dense or _apply_sparse

# on the same device as the variable.

# TODO(apassos): figure out how to get the variable name here.

if context.executing_eagerly() or isinstance(

var,

resource_variable_ops.ResourceVariable) and not var._in_graph_mode: # pylint: disable=protected-access

scope_name = ""

else:

scope_name = var.op.name

with ops.name_scope("update_" + scope_name), ops.colocate_with(var):

update_ops.append(processor.update_op(self, grad))

if global_step is None:

apply_updates = self._finish(update_ops, name)

else:

with ops.control_dependencies([self._finish(update_ops, "update")]):

with ops.colocate_with(global_step):

if isinstance(global_step, resource_variable_ops.ResourceVariable):

# TODO(apassos): the implicit read in assign_add is slow; consider

# making it less so.

apply_updates = resource_variable_ops.assign_add_variable_op(

global_step.handle,

ops.convert_to_tensor(1, dtype=global_step.dtype),

name=name)

else:

apply_updates = state_ops.assign_add(global_step, 1, name=name)

if not context.executing_eagerly():

if isinstance(apply_updates, ops.Tensor):

apply_updates = apply_updates.op

train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)

if apply_updates not in train_op:

train_op.append(apply_updates)

return apply_updates

def _distributed_apply(self,

distribution,

grads_and_vars,

global_step=None,

name=None):

"""A version of `apply_gradients` for cross-tower context.

This is a version of `apply_gradients()` for when you are using a

`DistributionStrategy` and are in a cross-tower context. If in a

tower context, use `apply_gradients()` as normal.

Args:

distribution: A `DistributionStrategy` object.

grads_and_vars: List of (gradient, variable) pairs as returned by

`compute_gradients()`, and then aggregated across towers.

global_step: Optional (mirrored) `Variable` to increment by one

after the variables have been updated.

name passed to the `Optimizer` constructor.

Returns:

An `Operation` that applies the specified gradients across all

towers. If `global_step` was not None, that operation also

increments `global_step`.

"""

reduced_grads = distribution.batch_reduce("sum", grads_and_vars)

var_list = [v for _, v in grads_and_vars]

grads_and_vars = zip(reduced_grads, var_list)

# Note that this is called in a cross-tower context.

self._create_slots(var_list)

def update(v, g):

"""Apply gradients to a replica variable."""

assert v is not None

try:

# Convert the grad to Tensor or IndexedSlices if necessary.

g = ops.convert_to_tensor_or_indexed_slices(g)

except TypeError:

raise TypeError("Gradient must be convertible to a Tensor"

" or IndexedSlices, or None: %s" % g)

if not isinstance(g, (ops.Tensor, ops.IndexedSlices)):

raise TypeError(

"Gradient must be a Tensor, IndexedSlices, or None: %s" % g)

p = _get_processor(v)

scope_name = "" if context.executing_eagerly() else v.op.name

# device_policy is set because non-mirrored tensors will be read in

# `update_op`. `_resource_apply_dense`, `lr_t`, `beta1_t` and `beta2_t`

# is an example.

with ops.name_scope("update_" + scope_name):

return p.update_op(self, g)

with ops.name_scope(name, self._name) as name:

self._prepare()

update_ops = [

for grad, var in grads_and_vars

for op in distribution.unwrap(distribution.update(var, update, grad))

]

def finish(self, update_ops):

return self._finish(update_ops, "update")

non_slot_devices = distribution.non_slot_devices(var_list)

finish_updates = distribution.update_non_slot(

non_slot_devices, finish, self, update_ops)

if global_step is None:

apply_updates = distribution.group(finish_updates, name=name)

else:

with ops.control_dependencies(distribution.unwrap(finish_updates)):

apply_updates = distribution.group(distribution.update(

global_step, state_ops.assign_add, 1, name=name))

if not context.executing_eagerly():

if isinstance(apply_updates, ops.Tensor):

apply_updates = apply_updates.op

train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)

if apply_updates not in train_op:

train_op.append(apply_updates)

return apply_updates

def get_slot(self, var, name):

"""Return a slot named `name` created for `var` by the Optimizer.

Some `Optimizer` subclasses use additional variables. For example

`Momentum` and `Adagrad` use variables to accumulate updates. This method

gives access to these `Variable` objects if for some reason you need them.

Use `get_slot_names()` to get the list of slot names created by the

`Optimizer`.

Args:

var: A variable passed to `minimize()` or `apply_gradients()`.

Returns:

The `Variable` for the slot if it was created, `None` otherwise.

"""

# pylint: disable=protected-access

named_slots = self._slots.get(name, None)

if not named_slots:

return None

if hasattr(var, "_mirrored_container"):

# NOTE: If this isn't patched, then there is no `handle` in

# `_resource_apply_dense`.

mirrored_container = var._mirrored_container()

assert mirrored_container is not None

if context.executing_eagerly():

key = mirrored_container._unique_id

else:

key = (mirrored_container.graph, mirrored_container._shared_name)

# pylint: enable=protected-access

mirrored_slot = named_slots.get(key, None)

if mirrored_slot is None: return None

return mirrored_slot.get(device=var.device)

return named_slots.get(_var_key(var), None)

def get_slot_names(self):

"""Return a list of the names of slots created by the `Optimizer`.

See `get_slot()`.

Returns:

A list of strings.

"""

return sorted(self._slots.keys())

def variables(self):

"""A list of variables which encode the current state of `Optimizer`.

Includes slot variables and additional global variables created by the

optimizer in the current default graph.

Returns:

A list of variables.

"""

executing_eagerly = context.executing_eagerly()

current_graph = ops.get_default_graph()

def _from_current_graph(variable):

if executing_eagerly:

# No variable.op in eager mode. We don't expect lots of eager graphs,

# but behavior should be consistent with graph mode.

return variable._graph_key == current_graph._graph_key # pylint: disable=protected-access

else:

return variable.op.graph is current_graph

optimizer_variables = [v for v in self._non_slot_variables()

if _from_current_graph(v)]

for _, variable_dict in self._slots.items():

for _, slot_for_variable in variable_dict.items():

if _from_current_graph(slot_for_variable):

optimizer_variables.append(slot_for_variable)

# Sort variables by name so that the return is deterministic.

return sorted(optimizer_variables, key=lambda v: v.name)

def _create_non_slot_variable(self, initial_value, name, colocate_with):

"""Add an extra variable, not associated with a slot."""

# Recommendation: Use OptimizerV2 if your optimizer uses non-slot variables.

eager = context.executing_eagerly()

graph = None if eager else colocate_with.graph

key = (name, graph)

v = self._non_slot_dict.get(key, None)

if v is None:

self._maybe_initialize_checkpointable()

distribution_strategy = distribute_lib.get_distribution_strategy()

with distribution_strategy.colocate_vars_with(colocate_with):

if eager:

restored_initial_value = self._preload_simple_restoration(

name=name, shape=None)

if restored_initial_value is not None:

initial_value = restored_initial_value

v = variable_scope.variable(initial_value, name=name, trainable=False)

# Restore this variable by name if necessary, but don't add a

# Checkpointable dependency. Optimizers return the current graph's

# non-slot variables from _checkpoint_dependencies explicitly rather

# than unconditionally adding dependencies (since there may be multiple

# non-slot variables with the same name in different graphs, trying to

# save all of them would result in errors).

self._handle_deferred_dependencies(name=name, checkpointable=v)

self._non_slot_dict[key] = v

return v

@property

def _checkpoint_dependencies(self):

"""From Checkpointable. Gather graph-specific non-slot variables to save."""

current_graph_non_slot_variables = []

current_graph_key = ops.get_default_graph()._graph_key # pylint: disable=protected-access

for (name, _), variable_object in sorted(self._non_slot_dict.items(),

# Avoid comparing graphs

key=lambda item: item[0][0]):

if variable_object._graph_key == current_graph_key: # pylint: disable=protected-access

current_graph_non_slot_variables.append(

checkpointable.CheckpointableReference(

name=name, ref=variable_object))

return (super(Optimizer, self)._checkpoint_dependencies

+ current_graph_non_slot_variables)

def _lookup_dependency(self, name):

"""From Checkpointable. Find a non-slot variable in the current graph."""

unconditional = super(Optimizer, self)._lookup_dependency(name)

if unconditional is not None:

return unconditional

graph = None if context.executing_eagerly() else ops.get_default_graph()

return self._get_non_slot_variable(name, graph=graph)

def _get_non_slot_variable(self, name, graph=None):

non_slot = self._non_slot_dict.get((name, graph), None)

if hasattr(non_slot, "_mirrored_container"):

# This is a mirrored non-slot. In order to enable code like `_finish`

# to assign to a non-slot, return the current context replica.

return non_slot.get()

else:

return non_slot

def _non_slot_variables(self):

"""Additional variables created by the `Optimizer`.

Returns:

A list or tuple of variables.

"""

return self._non_slot_dict.values()

def _assert_valid_dtypes(self, tensors):

"""Asserts tensors are all valid types (see `_valid_dtypes`).

Args:

tensors: Tensors to check.

Raises:

ValueError: If any tensor is not a valid type.

"""

valid_dtypes = self._valid_dtypes()

for t in tensors:

dtype = t.dtype.base_dtype

if dtype not in valid_dtypes:

raise ValueError(

"Invalid type %r for %s, expected: %s." % (

dtype, t.name, [v for v in valid_dtypes]))

# --------------

# Methods to be implemented by subclasses if they want to use the

# inherited implementation of apply_gradients() or compute_gradients().

# --------------

def _valid_dtypes(self):

"""Valid types for loss, variables and gradients.

Subclasses should override to allow other float types.

Returns:

Valid types for loss, variables and gradients.

"""

return set(

[dtypes.float16, dtypes.bfloat16, dtypes.float32, dtypes.float64])

def _create_slots(self, var_list):

"""Create all slots needed by the variables.

Args:

var_list: A list of `Variable` objects.

"""

# No slots needed by default

pass

def _prepare(self):

"""Create all needed tensors before applying gradients.

This is called with the name_scope using the "name" that

users have chosen for the application of gradients.

"""

pass

def _apply_dense(self, grad, var):

"""Add ops to apply dense gradients to `var`.

Args:

grad: A `Tensor`.

var: A `Variable` object.

Returns:

An `Operation`.

"""

raise NotImplementedError()

def _resource_apply_dense(self, grad, handle):

"""Add ops to apply dense gradients to the variable `handle`.

Args:

grad: a `Tensor` representing the gradient.

handle: a `Tensor` of dtype `resource` which points to the variable

to be updated.

Returns:

An `Operation` which updates the value of the variable.

"""

raise NotImplementedError()

def _resource_apply_sparse_duplicate_indices(self, grad, handle, indices):

"""Add ops to apply sparse gradients to `handle`, with repeated indices.

Optimizers which override this method must deal with repeated indices. See

the docstring of `_apply_sparse_duplicate_indices` for details. By default

the correct behavior, to sum non-unique indices and their associated

gradients, is enforced by first pre-processing `grad` and `indices` and

passing them on to `_resource_apply_sparse`. Optimizers which deal correctly

with duplicate indices may instead override this method to avoid the

overhead of summing.

Args:

grad: a `Tensor` representing the gradient for the affected indices.

handle: a `Tensor` of dtype `resource` which points to the variable

to be updated.

indices: a `Tensor` of integral type representing the indices for

which the gradient is nonzero. Indices may be repeated.

Returns:

An `Operation` which updates the value of the variable.

"""

summed_grad, unique_indices = _deduplicate_indexed_slices(

values=grad, indices=indices)

return self._resource_apply_sparse(summed_grad, handle, unique_indices)

def _resource_apply_sparse(self, grad, handle, indices):

"""Add ops to apply sparse gradients to the variable `handle`.

Similar to `_apply_sparse`, the `indices` argument to this method has been

de-duplicated. Optimizers which deal correctly with non-unique indices may

instead override `_resource_apply_sparse_duplicate_indices` to avoid this

overhead.

Args:

grad: a `Tensor` representing the gradient for the affected indices.

handle: a `Tensor` of dtype `resource` which points to the variable

to be updated.

indices: a `Tensor` of integral type representing the indices for

which the gradient is nonzero. Indices are unique.

Returns:

An `Operation` which updates the value of the variable.

"""

raise NotImplementedError()

def _apply_sparse_duplicate_indices(self, grad, var):

"""Add ops to apply sparse gradients to `var`, with repeated sparse indices.

Optimizers which override this method must deal with IndexedSlices objects

such as the following:

IndexedSlicesValue(values=[1, 1], indices=[0, 0], dense_shape=[1])

The correct interpretation is:

IndexedSlicesValue(values=[2], indices=[0], dense_shape=[1])

Many optimizers deal incorrectly with repeated indices when updating based

on sparse gradients (e.g. summing squares rather than squaring the sum, or

applying momentum terms multiple times). Adding first is always the correct

behavior, so this is enforced here by reconstructing the IndexedSlices to

have only unique indices, then calling _apply_sparse.

Optimizers which deal correctly with repeated indices may instead override

this method to avoid the overhead of summing indices.

Args:

grad: `IndexedSlices`.

var: A `Variable` object.

View remainder of file in raw view

tensorflow/tensorflow/python/training/optimizer.py at master · codeplaysoftware/tensorflow · GitHub

Repository navigation

Expand file tree

Breadcrumbs

optimizer.py

Latest commit

History

Breadcrumbs

optimizer.py

File metadata and controls