test_optimizer.py 58.8 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# 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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# 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.

15 16
import os
import tempfile
Q
Qiao Longfei 已提交
17 18
import unittest

19
import paddle.fluid as fluid
20 21
import paddle.fluid.framework as framework
import paddle.fluid.optimizer as optimizer
22
import paddle.fluid.core as core
M
mapingshuo 已提交
23
import paddle.compat as cpt
24
import numpy as np
25
from paddle.fluid.backward import append_backward
L
Leo Chen 已提交
26
from paddle.fluid.framework import Program, program_guard, convert_np_dtype_to_dtype_
C
chentianyu03 已提交
27
from paddle.fluid.framework import _test_eager_guard
L
Leo Chen 已提交
28
import paddle
29 30
from paddle.io import Dataset
import numpy
31

Q
Qiao Longfei 已提交
32 33

class TestOptimizer(unittest.TestCase):
34

Q
Qiao Longfei 已提交
35
    def test_sgd_optimizer(self):
36

Q
qiaolongfei 已提交
37 38 39 40
        def check_sgd_optimizer(optimizer_attr):
            init_program = framework.Program()
            program = framework.Program()
            block = program.global_block()
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
            mul_x = block.create_parameter(dtype="float32",
                                           shape=[5, 10],
                                           lod_level=0,
                                           name="mul.x",
                                           optimize_attr=optimizer_attr)
            mul_y = block.create_var(dtype="float32",
                                     shape=[10, 8],
                                     lod_level=0,
                                     name="mul.y")
            mul_out = block.create_var(dtype="float32",
                                       shape=[5, 8],
                                       lod_level=0,
                                       name="mul.out")
            mean_out = block.create_var(dtype="float32",
                                        shape=[1],
                                        lod_level=0,
                                        name="mean.out")
            block.append_op(type="mul",
                            inputs={
                                "X": mul_x,
                                "Y": mul_y
                            },
                            outputs={"Out": mul_out},
                            attrs={"x_num_col_dims": 1})
            block.append_op(type="mean",
                            inputs={"X": mul_out},
                            outputs={"Out": mean_out})
Q
qiaolongfei 已提交
68 69 70 71 72
            sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01)
            opts, _ = sgd_optimizer.minimize(mean_out, init_program)
            return opts

        opts = check_sgd_optimizer({'learning_rate': 1.1})
73 74
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
Q
Qiao Longfei 已提交
75

Q
qiaolongfei 已提交
76 77 78 79
        opts = check_sgd_optimizer({'learning_rate': 1.0})
        self.assertEqual(len(opts), 1)
        self.assertEqual([op.type for op in opts], ["sgd"])

Q
Qiao Longfei 已提交
80

81
class TestOptimizerBackwardApplygrad(unittest.TestCase):
82

83
    def test_sgd_optimizer(self):
84

85 86 87 88
        def check_sgd_optimizer(optimizer_attr):
            init_program = framework.Program()
            program = framework.Program()
            block = program.global_block()
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
            mul_x = block.create_parameter(dtype="float32",
                                           shape=[5, 10],
                                           lod_level=0,
                                           name="mul.x",
                                           optimize_attr=optimizer_attr)
            mul_y = block.create_var(dtype="float32",
                                     shape=[10, 8],
                                     lod_level=0,
                                     name="mul.y")
            mul_out = block.create_var(dtype="float32",
                                       shape=[5, 8],
                                       lod_level=0,
                                       name="mul.out")
            mean_out = block.create_var(dtype="float32",
                                        shape=[1],
                                        lod_level=0,
                                        name="mean.out")
            block.append_op(type="mul",
                            inputs={
                                "X": mul_x,
                                "Y": mul_y
                            },
                            outputs={"Out": mul_out},
                            attrs={"x_num_col_dims": 1})
            block.append_op(type="mean",
                            inputs={"X": mul_out},
                            outputs={"Out": mean_out})
116 117 118 119 120 121 122
            sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01)
            with framework.program_guard(program, init_program):
                p_g = sgd_optimizer.backward(mean_out)
                opts = sgd_optimizer.apply_gradients(p_g)
            return opts

        opts = check_sgd_optimizer({'learning_rate': 1.1})
123 124
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
125 126 127 128 129 130

        opts = check_sgd_optimizer({'learning_rate': 1.0})
        self.assertEqual(len(opts), 1)
        self.assertEqual([op.type for op in opts], ["sgd"])


131
class TestMomentumOptimizer(unittest.TestCase):
132

133
    class MockMomentum(optimizer.MomentumOptimizer):
134

135 136 137 138 139 140
        def get_accumulators(self):
            return self._accumulators

        def get_velocity_str(self):
            return self._velocity_acc_str

141
    def test_vanilla_momentum_optimizer(self):
Q
Qiao Longfei 已提交
142
        init_program = framework.Program()
143 144
        program = framework.Program()
        block = program.global_block()
145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x",
                                       optimize_attr={'learning_rate': 1.1})
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
Q
Qiao Longfei 已提交
165
        learning_rate = 0.01
166 167 168 169 170 171 172 173 174
        momentum_optimizer = self.MockMomentum(learning_rate=learning_rate,
                                               momentum=0.2)
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mean",
                        inputs={"X": mul_out},
                        outputs={"Out": mean_out})
F
fengjiayi 已提交
175
        params_grads = append_backward(mean_out)
176 177
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
178 179
        with framework.program_guard(program, init_program):
            opts = momentum_optimizer.apply_gradients(params_grads)
180
        self.assertEqual(len(opts), 2)
Y
Yancey1989 已提交
181
        sgd_op = opts[-1]
182
        self.assertEqual([op.type for op in opts], ["scale", "momentum"])
183
        self.assertFalse(sgd_op.attr('use_nesterov'))
184 185 186 187 188 189 190 191 192

        # Check accumulators
        accumulators = momentum_optimizer.get_accumulators()
        self.assertEqual(len(accumulators), 1)
        self.assertTrue(momentum_optimizer.get_velocity_str() in accumulators)
        velocity_acc = accumulators[momentum_optimizer.get_velocity_str()]
        self.assertEqual(len(velocity_acc), 1)
        self.assertTrue(mul_x.name in velocity_acc)

Q
Qiao Longfei 已提交
193 194 195 196
        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 2)
        self.assertEqual(init_ops[1].type, "fill_constant")
197 198 199
        self.assertAlmostEqual(init_ops[1].attr('value'), learning_rate)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), 0.0)
Q
Qiao Longfei 已提交
200

201
    def test_nesterov_momentum_optimizer(self):
Q
Qiao Longfei 已提交
202
        init_program = framework.Program()
203 204
        program = framework.Program()
        block = program.global_block()
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
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x",
                                       optimize_attr={'learning_rate': 1.1})
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mean",
                        inputs={"X": mul_out},
                        outputs={"Out": mean_out})
Q
Qiao Longfei 已提交
232
        learning_rate = 0.01
233 234 235
        momentum_optimizer = self.MockMomentum(learning_rate=learning_rate,
                                               momentum=0.2,
                                               use_nesterov=True)
F
fengjiayi 已提交
236
        params_grads = append_backward(mean_out)
237 238
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
239 240
        with framework.program_guard(program, init_program):
            opts = momentum_optimizer.apply_gradients(params_grads)
241
        self.assertEqual(len(opts), 2)
Y
Yancey1989 已提交
242
        sgd_op = opts[-1]
243
        self.assertEqual([op.type for op in opts], ["scale", "momentum"])
244
        self.assertTrue(sgd_op.attr('use_nesterov'))
245 246 247 248 249 250 251 252 253

        # Check accumulators
        accumulators = momentum_optimizer.get_accumulators()
        self.assertEqual(len(accumulators), 1)
        self.assertTrue(momentum_optimizer.get_velocity_str() in accumulators)
        velocity_acc = accumulators[momentum_optimizer.get_velocity_str()]
        self.assertEqual(len(velocity_acc), 1)
        self.assertTrue(mul_x.name in velocity_acc)

Q
Qiao Longfei 已提交
254 255 256 257
        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 2)
        self.assertEqual(init_ops[1].type, "fill_constant")
258 259 260
        self.assertAlmostEqual(init_ops[1].attr('value'), learning_rate)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), 0.0)
Q
Qiao Longfei 已提交
261

262

263
class TestAdagradOptimizer(unittest.TestCase):
264

265
    class MockAdagrad(optimizer.AdagradOptimizer):
266

267 268 269 270 271 272 273
        def get_accumulators(self):
            return self._accumulators

        def get_moment_str(self):
            return self._moment_acc_str

    def test_adagrad_optimizer(self):
Q
Qiao Longfei 已提交
274
        init_program = framework.Program()
275 276
        program = framework.Program()
        block = program.global_block()
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
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x",
                                       optimize_attr={'learning_rate': 1.1})
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mean",
                        inputs={"X": mul_out},
                        outputs={"Out": mean_out})
Q
Qiao Longfei 已提交
304
        learning_rate = 0.01
305 306
        adagrad_optimizer = self.MockAdagrad(learning_rate=learning_rate,
                                             epsilon=1.0e-6)
F
fengjiayi 已提交
307
        params_grads = append_backward(mean_out)
308 309
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adagrad_optimizer.get_accumulators()), 0)
310 311
        with framework.program_guard(program, init_program):
            opts = adagrad_optimizer.apply_gradients(params_grads)
312 313
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "adagrad"])
314

315
        # Check accumulators
316 317 318 319 320 321 322
        accumulators = adagrad_optimizer.get_accumulators()
        self.assertEqual(len(accumulators), 1)
        self.assertTrue(adagrad_optimizer.get_moment_str() in accumulators)
        moment_acc = accumulators[adagrad_optimizer.get_moment_str()]
        self.assertEqual(len(moment_acc), 1)
        self.assertTrue(mul_x.name in moment_acc)

Q
Qiao Longfei 已提交
323 324
        # Check init_program
        init_ops = init_program.global_block().ops
Z
zhongpu 已提交
325
        self.assertEqual(len(init_ops), 2)
Q
Qiao Longfei 已提交
326
        self.assertEqual(init_ops[1].type, "fill_constant")
327 328 329
        self.assertAlmostEqual(init_ops[1].attr('value'), learning_rate)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), 0.0)
Q
Qiao Longfei 已提交
330

331

332
class TestAdamOptimizer(unittest.TestCase):
333

334
    class MockAdam(optimizer.AdamOptimizer):
335

336 337 338 339 340 341 342 343 344 345
        def get_accumulators(self):
            return self._accumulators

        def get_moment1_str(self):
            return self._moment1_acc_str

        def get_moment2_str(self):
            return self._moment2_acc_str

    def test_adam_optimizer(self):
Q
Qiao Longfei 已提交
346
        init_program = framework.Program()
347 348
        program = framework.Program()
        block = program.global_block()
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
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x",
                                       optimize_attr={'learning_rate': 1.1})
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mean",
                        inputs={"X": mul_out},
                        outputs={"Out": mean_out})
Q
Qiao Longfei 已提交
376
        learning_rate = 0.01
377 378 379
        adam_optimizer = self.MockAdam(learning_rate=learning_rate,
                                       beta1=0.9,
                                       beta2=0.999)
F
fengjiayi 已提交
380
        params_grads = append_backward(mean_out)
381 382
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adam_optimizer.get_accumulators()), 0)
383 384
        with framework.program_guard(program, init_program):
            opts = adam_optimizer.apply_gradients(params_grads)
A
Aurelius84 已提交
385 386
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "adam"])
387 388 389

        # Check accumulators
        accumulators = adam_optimizer.get_accumulators()
Q
qiaolongfei 已提交
390
        self.assertEqual(len(accumulators), 4)
391 392 393 394 395 396 397 398 399
        self.assertTrue(adam_optimizer.get_moment1_str() in accumulators)
        self.assertTrue(adam_optimizer.get_moment2_str() in accumulators)
        moment1_acc = accumulators[adam_optimizer.get_moment1_str()]
        moment2_acc = accumulators[adam_optimizer.get_moment2_str()]
        self.assertEqual(len(moment1_acc), 1)
        self.assertEqual(len(moment2_acc), 1)
        self.assertTrue(mul_x.name in moment1_acc)
        self.assertTrue(mul_x.name in moment2_acc)

Q
Qiao Longfei 已提交
400 401 402
        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 5)
403 404
        self.assertEqual(init_ops[-1].type, "fill_constant")
        self.assertAlmostEqual(init_ops[-1].attr('value'), learning_rate)
Q
Qiao Longfei 已提交
405

406

407
class TestAdamaxOptimizer(unittest.TestCase):
408

409
    class MockAdamax(optimizer.AdamaxOptimizer):
410

411 412 413 414 415 416 417 418 419 420
        def get_accumulators(self):
            return self._accumulators

        def get_moment_str(self):
            return self._moment_acc_str

        def get_inf_norm_str(self):
            return self._inf_norm_acc_str

    def test_adamax_optimizer(self):
Q
Qiao Longfei 已提交
421
        init_program = framework.Program()
422 423
        program = framework.Program()
        block = program.global_block()
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
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x",
                                       optimize_attr={'learning_rate': 1.1})
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mean",
                        inputs={"X": mul_out},
                        outputs={"Out": mean_out})
Q
Qiao Longfei 已提交
451
        learning_rate = 0.01
452 453 454
        adamax_optimizer = self.MockAdamax(learning_rate=learning_rate,
                                           beta1=0.9,
                                           beta2=0.999)
F
fengjiayi 已提交
455
        params_grads = append_backward(mean_out)
456 457
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adamax_optimizer.get_accumulators()), 0)
458 459
        with framework.program_guard(program, init_program):
            opts = adamax_optimizer.apply_gradients(params_grads)
460 461
        self.assertEqual(len(opts), 3)
        self.assertEqual([op.type for op in opts], ["scale", "adamax", "scale"])
462 463 464

        # Check accumulators
        accumulators = adamax_optimizer.get_accumulators()
Q
qiaolongfei 已提交
465
        self.assertEqual(len(accumulators), 3)
466 467 468 469 470 471 472 473 474
        self.assertTrue(adamax_optimizer.get_moment_str() in accumulators)
        self.assertTrue(adamax_optimizer.get_inf_norm_str() in accumulators)
        moment_acc = accumulators[adamax_optimizer.get_moment_str()]
        inf_norm_acc = accumulators[adamax_optimizer.get_inf_norm_str()]
        self.assertEqual(len(moment_acc), 1)
        self.assertEqual(len(inf_norm_acc), 1)
        self.assertTrue(mul_x.name in moment_acc)
        self.assertTrue(mul_x.name in inf_norm_acc)

Q
Qiao Longfei 已提交
475 476 477
        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 4)
478 479
        self.assertEqual(init_ops[-1].type, "fill_constant")
        self.assertAlmostEqual(init_ops[-1].attr('value'), learning_rate)
Q
Qiao Longfei 已提交
480

481

482
class TestDpsgdOptimizer(unittest.TestCase):
483

484
    def test_dpsgd_optimizer(self):
485

486 487 488 489
        def check_dpsgd_optimizer(optimizer_attr):
            init_program = framework.Program()
            program = framework.Program()
            block = program.global_block()
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
            mul_x = block.create_parameter(dtype="float32",
                                           shape=[5, 10],
                                           lod_level=0,
                                           name="mul.x",
                                           optimize_attr=optimizer_attr)
            mul_y = block.create_var(dtype="float32",
                                     shape=[10, 8],
                                     lod_level=0,
                                     name="mul.y")
            mul_out = block.create_var(dtype="float32",
                                       shape=[5, 8],
                                       lod_level=0,
                                       name="mul.out")
            block.append_op(type="mul",
                            inputs={
                                "X": mul_x,
                                "Y": mul_y
                            },
                            outputs={"Out": mul_out},
                            attrs={"x_num_col_dims": 1})
            mean_out = block.create_var(dtype="float32",
                                        shape=[1],
                                        lod_level=0,
                                        name="mean.out")
            block.append_op(type="mean",
                            inputs={"X": mul_out},
                            outputs={"Out": mean_out})
            dpsgd_optimizer = optimizer.DpsgdOptimizer(learning_rate=0.01,
                                                       clip=100.0,
                                                       batch_size=16.0,
                                                       sigma=0.0)
521 522 523 524 525 526 527 528 529 530 531 532 533
            opts, _ = dpsgd_optimizer.minimize(mean_out, init_program)
            return opts

        opts = check_dpsgd_optimizer({
            'learning_rate': 1.1,
            'clip': 100.0,
            'batch_size': 16.0,
            'sigma': 4.0
        })
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "dpsgd"])


534
class TestDecayedAdagradOptimizer(unittest.TestCase):
535

536
    class MockDecayedAdagrad(optimizer.DecayedAdagradOptimizer):
537

538 539 540 541 542 543 544 545 546 547
        def get_accumulators(self):
            return self._accumulators

        def get_moment_str(self):
            return self._moment_acc_str

    def test_decayed_adagrad_optimizer(self):
        init_program = framework.Program()
        program = framework.Program()
        block = program.global_block()
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
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x",
                                       optimize_attr={'learning_rate': 1.1})
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mean",
                        inputs={"X": mul_out},
                        outputs={"Out": mean_out})
575 576 577
        learning_rate = 0.01
        decayed_adagrad_optimizer = self.MockDecayedAdagrad(
            learning_rate=learning_rate, decay=0.95, epsilon=1.0e-6)
F
fengjiayi 已提交
578
        params_grads = append_backward(mean_out)
579 580
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(decayed_adagrad_optimizer.get_accumulators()), 0)
581 582
        with framework.program_guard(program, init_program):
            opts = decayed_adagrad_optimizer.apply_gradients(params_grads)
583 584
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "decayed_adagrad"])
585 586 587 588 589 590 591 592 593 594 595 596 597 598

        # Check accumulators
        accumulators = decayed_adagrad_optimizer.get_accumulators()
        self.assertEqual(len(accumulators), 1)
        self.assertTrue(
            decayed_adagrad_optimizer.get_moment_str() in accumulators)
        moment_acc = accumulators[decayed_adagrad_optimizer.get_moment_str()]
        self.assertEqual(len(moment_acc), 1)
        self.assertTrue(mul_x.name in moment_acc)

        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 2)
        self.assertEqual(init_ops[1].type, "fill_constant")
599 600 601
        self.assertAlmostEqual(init_ops[1].attr('value'), learning_rate)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), 0.0)
602 603


Q
qiaolongfei 已提交
604
class TestFtrlOptimizer(unittest.TestCase):
605

Q
qiaolongfei 已提交
606
    class MockFtrl(optimizer.FtrlOptimizer):
607

Q
qiaolongfei 已提交
608 609 610 611 612 613 614 615 616 617 618 619 620
        def get_accumulators(self):
            return self._accumulators

        def get_squared_str(self):
            return self._squared_acc_str

        def get_linear_str(self):
            return self._linear_acc_str

    def test_ftrl_optimizer(self):
        init_program = framework.Program()
        program = framework.Program()
        block = program.global_block()
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
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x",
                                       optimize_attr={'learning_rate': 1.1})
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mean",
                        inputs={"X": mul_out},
                        outputs={"Out": mean_out})
Q
qiaolongfei 已提交
648
        learning_rate = 0.01
649 650 651 652
        ftrl_optimizer = self.MockFtrl(learning_rate=learning_rate,
                                       l1=0.0,
                                       l2=0.0,
                                       lr_power=-0.5)
Q
qiaolongfei 已提交
653 654 655
        params_grads = append_backward(mean_out)
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(ftrl_optimizer.get_accumulators()), 0)
656 657
        with framework.program_guard(program, init_program):
            opts = ftrl_optimizer.apply_gradients(params_grads)
658 659
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "ftrl"])
Q
qiaolongfei 已提交
660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675

        # Check accumulators
        accumulators = ftrl_optimizer.get_accumulators()
        self.assertEqual(len(accumulators), 2)
        self.assertTrue(ftrl_optimizer.get_squared_str() in accumulators)
        self.assertTrue(ftrl_optimizer.get_linear_str() in accumulators)
        squared_acc = accumulators[ftrl_optimizer.get_squared_str()]
        linear_acc = accumulators[ftrl_optimizer.get_linear_str()]
        self.assertEqual(len(squared_acc), 1)
        self.assertEqual(len(linear_acc), 1)
        self.assertTrue(mul_x.name in squared_acc)
        self.assertTrue(mul_x.name in linear_acc)

        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 3)
676 677
        self.assertEqual(init_ops[-1].type, "fill_constant")
        self.assertAlmostEqual(init_ops[-1].attr('value'), learning_rate)
Q
qiaolongfei 已提交
678 679


M
mapingshuo 已提交
680
class TestLookaheadOptimizer(unittest.TestCase):
681

M
mapingshuo 已提交
682 683 684 685 686
    def test_lookahead_optimizer(self):
        init_program = framework.Program()
        program = framework.Program()
        block = program.global_block()
        init_block = init_program.global_block()
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
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x",
                                       optimize_attr={'learning_rate': 1.1})
        init_mul_x = init_block.create_parameter(dtype="float32",
                                                 shape=[5, 10],
                                                 lod_level=0,
                                                 name="mul.x")
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")

        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
        block.append_op(type="mean",
                        inputs={"X": mul_out},
                        outputs={"Out": mean_out})
M
mapingshuo 已提交
719 720 721 722 723

        sgd = optimizer.SGD(learning_rate=0.01)
        lookahead = optimizer.LookaheadOptimizer(sgd, alpha=0.5, k=5)
        with framework.program_guard(program, init_program):
            opts, _ = lookahead.minimize(mean_out)
724 725
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
M
mapingshuo 已提交
726 727


M
mapingshuo 已提交
728
class TestRecomputeOptimizer(unittest.TestCase):
729

730
    def net(self, return_input=False, with_dropout=False, with_seed=False):
M
mapingshuo 已提交
731 732
        program = framework.Program()
        block = program.global_block()
733 734 735 736 737 738 739 740 741 742 743 744
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x")
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
745 746

        if with_dropout is True:
747 748 749 750 751 752 753 754
            mul_out_drop = block.create_var(dtype="float32",
                                            shape=[5, 8],
                                            lod_level=0,
                                            name="mul.out.dropout")
            mul_out_mask = block.create_var(dtype="uint8",
                                            shape=[5, 8],
                                            lod_level=0,
                                            name="mul.out.mask")
755
            if with_seed is True:
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
                seed_out = block.create_var(dtype="int32",
                                            shape=[1],
                                            name="seed.out")

        b1 = block.create_parameter(dtype="float32",
                                    shape=[5, 8],
                                    lod_level=0,
                                    name="b1")
        b1_out = block.create_var(dtype="float32",
                                  shape=[5, 8],
                                  lod_level=0,
                                  name="b1_out")
        b2 = block.create_parameter(dtype="float32",
                                    shape=[5, 8],
                                    lod_level=0,
                                    name="b2")
        b2_out = block.create_var(dtype="float32",
                                  shape=[5, 8],
                                  lod_level=0,
                                  name="b2_out")
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
787 788 789 790

        if with_dropout is True:
            dropout_inputs = {'X': [mul_out]}
            if with_seed is True:
791 792 793 794 795 796 797
                block.append_op(type='seed',
                                outputs={'Out': seed_out},
                                attrs={
                                    'deterministic': True,
                                    'rng_name': 'rng0',
                                    'force_cpu': True
                                })
798 799
                dropout_inputs = {'X': [mul_out], 'Seed': [seed_out]}

800 801 802 803 804 805 806 807 808 809 810 811 812 813 814
            block.append_op(type='dropout',
                            inputs=dropout_inputs,
                            outputs={
                                'Out': [mul_out_drop],
                                'Mask': [mul_out_mask]
                            },
                            attrs={
                                'dropout_prob': 0.5,
                            })
            block.append_op(type="elementwise_add",
                            inputs={
                                "X": mul_out_drop,
                                "Y": b1
                            },
                            outputs={"Out": b1_out})
M
mapingshuo 已提交
815
        else:
816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831
            block.append_op(type="elementwise_add",
                            inputs={
                                "X": mul_out,
                                "Y": b1
                            },
                            outputs={"Out": b1_out})

        block.append_op(type="elementwise_add",
                        inputs={
                            "X": b1_out,
                            "Y": b2
                        },
                        outputs={"Out": b2_out})
        block.append_op(type="mean",
                        inputs={"X": b2_out},
                        outputs={"Out": mean_out})
M
mapingshuo 已提交
832

833 834
        if return_input == True:
            return mul_x, mul_out, b1_out, b2_out, mean_out
M
mapingshuo 已提交
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
        return mul_out, b1_out, b2_out, mean_out

    def test_no_checkpoint(self):
        mul_out, b1_out, b2_out, mean_out = self.net()
        self.assertEqual(len(mean_out.block.ops), 4)
        self.assertEqual([op.type for op in mean_out.block.ops],
                         ["mul", "elementwise_add", "elementwise_add", "mean"])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 12)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "elementwise_add", "elementwise_add", "mean",
            "fill_constant", "mean_grad", "elementwise_add_grad",
            "elementwise_add_grad", "mul_grad", "sgd", "sgd", "sgd"
        ])

    def test_one_checkpoint(self):
        mul_out, b1_out, b2_out, mean_out = self.net()
        self.assertEqual(len(mean_out.block.ops), 4)
        self.assertEqual([op.type for op in mean_out.block.ops],
                         ["mul", "elementwise_add", "elementwise_add", "mean"])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([b1_out])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 13)
865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "elementwise_add", "elementwise_add", "mean",
            "fill_constant", "mean_grad", "elementwise_add_grad", "mul",
            "elementwise_add_grad", "mul_grad", "sgd", "sgd", "sgd"
        ])

    def test_str_checkpoints(self):
        mul_out, b1_out, b2_out, mean_out = self.net()
        self.assertEqual(len(mean_out.block.ops), 4)
        self.assertEqual([op.type for op in mean_out.block.ops],
                         ["mul", "elementwise_add", "elementwise_add", "mean"])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([b1_out.name])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 13)
M
mapingshuo 已提交
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
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "elementwise_add", "elementwise_add", "mean",
            "fill_constant", "mean_grad", "elementwise_add_grad", "mul",
            "elementwise_add_grad", "mul_grad", "sgd", "sgd", "sgd"
        ])

    def test_multi_checkpoint(self):
        mul_out, b1_out, b2_out, mean_out = self.net()
        self.assertEqual(len(mean_out.block.ops), 4)
        self.assertEqual([op.type for op in mean_out.block.ops],
                         ["mul", "elementwise_add", "elementwise_add", "mean"])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([mul_out, b2_out])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 13)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "elementwise_add", "elementwise_add", "mean",
            "fill_constant", "mean_grad", "elementwise_add",
            "elementwise_add_grad", "elementwise_add_grad", "mul_grad", "sgd",
            "sgd", "sgd"
        ])

    def test_adjacent_checkpoint(self):
        mul_out, b1_out, b2_out, mean_out = self.net()
        self.assertEqual(len(mean_out.block.ops), 4)
        self.assertEqual([op.type for op in mean_out.block.ops],
                         ["mul", "elementwise_add", "elementwise_add", "mean"])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([mul_out, b1_out])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 12)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "elementwise_add", "elementwise_add", "mean",
            "fill_constant", "mean_grad", "elementwise_add_grad",
            "elementwise_add_grad", "mul_grad", "sgd", "sgd", "sgd"
        ])

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
    def test_out_of_order_checkpoint(self):
        mul_out, b1_out, b2_out, mean_out = self.net()
        self.assertEqual(len(mean_out.block.ops), 4)
        self.assertEqual([op.type for op in mean_out.block.ops],
                         ["mul", "elementwise_add", "elementwise_add", "mean"])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([b2_out, mul_out])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 13)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "elementwise_add", "elementwise_add", "mean",
            "fill_constant", "mean_grad", "elementwise_add",
            "elementwise_add_grad", "elementwise_add_grad", "mul_grad", "sgd",
            "sgd", "sgd"
        ])

    def test_input_as_checkpoints(self):
        mul_x, mul_out, b1_out, b2_out, mean_out = self.net(return_input=True)
        self.assertEqual(len(mean_out.block.ops), 4)
        self.assertEqual([op.type for op in mean_out.block.ops],
                         ["mul", "elementwise_add", "elementwise_add", "mean"])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([mul_x, b2_out])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 14)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "elementwise_add", "elementwise_add", "mean",
            "fill_constant", "mean_grad", "mul", "elementwise_add",
            "elementwise_add_grad", "elementwise_add_grad", "mul_grad", "sgd",
            "sgd", "sgd"
        ])

M
mapingshuo 已提交
959 960 961 962 963 964
    def test_apply_gradients(self):
        mul_out, b1_out, b2_out, mean_out = self.net()
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([b1_out])
        # apply backward
965 966 967 968
        params_grads = recompute_optimizer.backward(mean_out,
                                                    startup_program=None,
                                                    parameter_list=None,
                                                    no_grad_set=None)
M
mapingshuo 已提交
969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987

        # apply gradient
        program = mean_out.block.program
        with framework.program_guard(program, None):
            optimize_ops = recompute_optimizer.apply_gradients(params_grads)

        self.assertEqual(len(mean_out.block.ops), 13)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "elementwise_add", "elementwise_add", "mean",
            "fill_constant", "mean_grad", "elementwise_add_grad", "mul",
            "elementwise_add_grad", "mul_grad", "sgd", "sgd", "sgd"
        ])

    def test_load(self):
        mul_out, b1_out, b2_out, mean_out = self.net()
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([b1_out])
        try:
988 989
            state_dict = {}
            recompute_optimizer.load(state_dict)
M
mapingshuo 已提交
990 991 992 993 994
        except NotImplementedError as e:
            self.assertEqual(
                "load function is not supported by Recompute Optimizer for now",
                cpt.get_exception_message(e))

M
mapingshuo 已提交
995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017
    def test_dropout(self):
        """
        If there are dropout layers in the forward nets, we should add a
        seed op
        """
        mul_out, b1_out, b2_out, mean_out = self.net(with_dropout=True)
        self.assertEqual(len(mean_out.block.ops), 5)
        self.assertEqual(
            [op.type for op in mean_out.block.ops],
            ["mul", "dropout", "elementwise_add", "elementwise_add", "mean"])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([b1_out])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 17)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "seed", "dropout", "elementwise_add", "elementwise_add",
            "mean", "fill_constant", "mean_grad", "elementwise_add_grad", "mul",
            "dropout", "elementwise_add_grad", "dropout_grad", "mul_grad",
            "sgd", "sgd", "sgd"
        ])

1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
    def test_dropout_with_determinate_seed(self):
        mul_out, b1_out, b2_out, mean_out = self.net(with_dropout=True,
                                                     with_seed=True)
        self.assertEqual(len(mean_out.block.ops), 6)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "seed", "dropout", "elementwise_add", "elementwise_add",
            "mean"
        ])
        sgd_optimizer = optimizer.SGD(learning_rate=1.0)
        recompute_optimizer = optimizer.RecomputeOptimizer(sgd_optimizer)
        recompute_optimizer._set_checkpoints([b1_out])
        opts, params_grads = recompute_optimizer.minimize(mean_out)

        self.assertEqual(len(mean_out.block.ops), 17)
        self.assertEqual([op.type for op in mean_out.block.ops], [
            "mul", "seed", "dropout", "elementwise_add", "elementwise_add",
            "mean", "fill_constant", "mean_grad", "elementwise_add_grad", "mul",
            "dropout", "elementwise_add_grad", "dropout_grad", "mul_grad",
            "sgd", "sgd", "sgd"
        ])

1039 1040 1041
    def test_dropout_with_seed(self):
        """
        when we recompute a dropout op, make sure that the recomputed one
1042 1043
        is the same as the original var.
        """
1044 1045 1046 1047

        def gen_data():
            return {
                "x": np.random.random(size=(100, 3)).astype('float32'),
1048
                "y": np.random.randint(2, size=(100, 1)).astype('int64')
1049 1050 1051
            }

        def mlp(input_x, input_y):
1052 1053 1054
            drop_res = fluid.layers.dropout(input_x,
                                            dropout_prob=0.5,
                                            name="dropout_with_seed_cpu")
1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066
            prediction = fluid.layers.fc(input=[drop_res],
                                         size=2,
                                         act='softmax')
            cost = fluid.layers.cross_entropy(input=prediction, label=input_y)
            sum_cost = fluid.layers.reduce_mean(cost)
            return drop_res, prediction, sum_cost

        main_program = Program()
        startup_program = Program()
        scope = fluid.Scope()
        with fluid.scope_guard(scope):
            with program_guard(main_program, startup_program):
1067 1068 1069
                input_x = fluid.layers.data(name="x",
                                            shape=[3],
                                            dtype='float32')
1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092
                input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')
                drop_res, prediction, cost = mlp(input_x, input_y)
                sgd = fluid.optimizer.Adam(learning_rate=0.01)
                sgd = fluid.optimizer.RecomputeOptimizer(sgd)
                sgd._set_checkpoints([prediction])
                sgd.minimize(cost)

                place = fluid.CPUPlace()
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                feed_data = gen_data()
                drop_vec = exe.run(feed=feed_data,
                                   program=fluid.default_main_program(),
                                   fetch_list=[
                                       "dropout_with_seed_cpu.tmp_1",
                                       "dropout_with_seed_cpu.tmp_1.subprog_0"
                                   ])
                self.assertEqual(drop_vec[0].tolist(), drop_vec[1].tolist())


@unittest.skipIf(not core.is_compiled_with_cuda(),
                 "core is not compiled with CUDA")
class TestRecomputeOptimizerCUDA(unittest.TestCase):
1093

1094 1095 1096 1097 1098 1099 1100 1101 1102
    def test_dropout_with_seed(self):
        """
        when we recompute a dropout op, make sure that the recomputed one
        is the same as the original var.
        """

        def gen_data():
            return {
                "x": np.random.random(size=(100, 3)).astype('float32'),
1103
                "y": np.random.randint(2, size=(100, 1)).astype('int64')
1104 1105 1106
            }

        def mlp(input_x, input_y):
1107 1108 1109
            drop_res = fluid.layers.dropout(input_x,
                                            dropout_prob=0.5,
                                            name="dropout_with_seed_gpu")
1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121
            prediction = fluid.layers.fc(input=[drop_res],
                                         size=2,
                                         act='softmax')
            cost = fluid.layers.cross_entropy(input=prediction, label=input_y)
            sum_cost = fluid.layers.reduce_mean(cost)
            return drop_res, prediction, sum_cost

        main_program = Program()
        startup_program = Program()
        scope = fluid.Scope()
        with fluid.scope_guard(scope):
            with program_guard(main_program, startup_program):
1122 1123 1124
                input_x = fluid.layers.data(name="x",
                                            shape=[3],
                                            dtype='float32')
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143
                input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')
                drop_res, prediction, cost = mlp(input_x, input_y)
                sgd = fluid.optimizer.Adam(learning_rate=0.01)
                sgd = fluid.optimizer.RecomputeOptimizer(sgd)
                sgd._set_checkpoints([prediction])
                sgd.minimize(cost)

                place = fluid.CUDAPlace(0)
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                feed_data = gen_data()
                drop_vec = exe.run(feed=feed_data,
                                   program=fluid.default_main_program(),
                                   fetch_list=[
                                       "dropout_with_seed_gpu.tmp_1",
                                       "dropout_with_seed_gpu.tmp_1.subprog_0"
                                   ])
                self.assertEqual(drop_vec[0].tolist(), drop_vec[1].tolist())

M
mapingshuo 已提交
1144

1145
class TestGradientMergeOptimizer(unittest.TestCase):
1146

1147 1148 1149
    def net(self):
        program = framework.Program()
        block = program.global_block()
1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189
        mul_x = block.create_parameter(dtype="float32",
                                       shape=[5, 10],
                                       lod_level=0,
                                       name="mul.x")
        mul_y = block.create_var(dtype="float32",
                                 shape=[10, 8],
                                 lod_level=0,
                                 name="mul.y")
        mul_out = block.create_var(dtype="float32",
                                   shape=[5, 8],
                                   lod_level=0,
                                   name="mul.out")
        b1 = block.create_parameter(dtype="float32",
                                    shape=[5, 8],
                                    lod_level=0,
                                    name="b1")
        b1_out = block.create_var(dtype="float32",
                                  shape=[5, 8],
                                  lod_level=0,
                                  name="b1_out")
        mean_out = block.create_var(dtype="float32",
                                    shape=[1],
                                    lod_level=0,
                                    name="mean.out")
        block.append_op(type="mul",
                        inputs={
                            "X": mul_x,
                            "Y": mul_y
                        },
                        outputs={"Out": mul_out},
                        attrs={"x_num_col_dims": 1})
        block.append_op(type="elementwise_add",
                        inputs={
                            "X": mul_out,
                            "Y": b1
                        },
                        outputs={"Out": b1_out})
        block.append_op(type="mean",
                        inputs={"X": b1_out},
                        outputs={"Out": mean_out})
1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205
        return mean_out

    def test_program_desc(self, ):
        cost = self.net()
        main_program = cost.block.program
        init_program = framework.Program()
        self.assertEqual(main_program.num_blocks, 1)
        self.assertEqual(len(cost.block.ops), 3)
        self.assertEqual([op.type for op in cost.block.ops],
                         ["mul", "elementwise_add", "mean"])

        opt = optimizer.SGD(learning_rate=1.0)
        opt = optimizer.GradientMergeOptimizer(opt, k_steps=4)
        with framework.program_guard(main_program, init_program):
            ops, params_grads = opt.minimize(cost)

1206
        self.assertEqual(main_program.num_blocks, 2)
1207 1208

        # main block
1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226
        self.assertEqual(len(cost.block.ops), 13)
        self.assertEqual(
            [op.type for op in cost.block.ops],
            [
                'mul',
                'elementwise_add',
                'mean',
                'fill_constant',
                'mean_grad',
                'elementwise_add_grad',
                'mul_grad',
                'increment',  # step += 1
                'elementwise_mod',  # step %= k_steps
                'equal',  # cond_var == (step == 0)
                'elementwise_add',
                'elementwise_add',
                'conditional_block',
            ])
1227

1228 1229
        # optimize block
        self.assertEqual(len(main_program.block(1).ops), 6)
1230 1231 1232
        self.assertEqual(
            [op.type for op in main_program.block(1).ops],
            ['scale', 'scale', 'sgd', 'sgd', 'fill_constant', 'fill_constant'])
1233 1234


L
Leo Chen 已提交
1235 1236 1237 1238 1239 1240 1241
class TestOptimizerDtype(unittest.TestCase):
    '''
    The dtype of optimizer should be inferred by parameters, and the learning rate
    is cteated with the same dtype.
    '''

    def check_with_dtype(self, dtype):
1242

L
Leo Chen 已提交
1243
        class MyLayer(paddle.nn.Layer):
1244

L
Leo Chen 已提交
1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267
            def __init__(self, dtype):
                super(MyLayer, self).__init__()
                self._w = self.create_parameter([2, 3], dtype=dtype)
                self._b = self.create_parameter([2, 3], dtype=dtype)

            def forward(self, x):
                return x * self._w + self._b

        with paddle.fluid.dygraph.guard():
            model = MyLayer(dtype)
            x = paddle.rand([10, 2, 3], dtype=dtype)
            loss = model(x)
            adam = paddle.optimizer.Adam(parameters=model.parameters())
            loss.backward()
            adam.step()
            self.assertEqual(adam._dtype, convert_np_dtype_to_dtype_(dtype))

    def test_float64(self):
        self.check_with_dtype('float64')

    def test_float32(self):
        self.check_with_dtype('float32')

C
chentianyu03 已提交
1268 1269 1270 1271 1272
    def test_api_eager_dygraph(self):
        with _test_eager_guard():
            self.test_float64()
            self.test_float32()

L
Leo Chen 已提交
1273

1274 1275
class TestMasterWeightSaveForFP16(unittest.TestCase):
    '''
1276
    For Amp-O2, some optimizer(Momentum, Adam ...) will create master weights for parameters to improve the accuracy.
1277 1278 1279
    Master weights will be saved by optimizer::state_dict.
    '''

1280 1281 1282 1283 1284 1285
    def setUp(self):
        self.temp_dir = tempfile.TemporaryDirectory()

    def tearDown(self):
        self.temp_dir.cleanup()

1286 1287 1288 1289 1290
    def check_with_opt_state_dict(self, use_save_load=True):
        paddle.seed(100)
        numpy.random.seed(100)

        class SimpleNet(paddle.nn.Layer):
1291

1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308
            def __init__(self, input_size, output_size):
                super(SimpleNet, self).__init__()
                self.linears = paddle.nn.LayerList([
                    paddle.nn.Linear(input_size, output_size) for i in range(1)
                ])

            def forward(self, x):
                for i, l in enumerate(self.linears):
                    x = self.linears[i](x)
                return x

        input_size = 2  # 设为较大的值
        output_size = 2  # 设为较大的值
        batch_size = 2  # batch_size 为8的倍数
        nums_batch = 10

        class RandomDataset(Dataset):
1309

1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321
            def __init__(self, num_samples):
                self.num_samples = num_samples

            def __getitem__(self, idx):
                data = numpy.random.random([input_size]).astype('float16')
                label = numpy.random.random([output_size]).astype('float16')
                return data, label

            def __len__(self):
                return self.num_samples

        dataset = RandomDataset(nums_batch * batch_size)
1322 1323 1324 1325 1326
        loader = paddle.io.DataLoader(dataset,
                                      batch_size=batch_size,
                                      shuffle=False,
                                      drop_last=True,
                                      num_workers=0)
1327 1328 1329

        mse = paddle.nn.MSELoss()
        model = SimpleNet(input_size, output_size)  # 定义模型
1330 1331 1332
        optimizer = paddle.optimizer.Momentum(learning_rate=0.0001,
                                              parameters=model.parameters(),
                                              multi_precision=True)  # 定义优化器
1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346
        scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
        model = paddle.amp.decorate(models=model, level='O2')

        for i, (data, label) in enumerate(loader):
            with paddle.amp.auto_cast(level='O2'):
                output = model(data)
                loss = mse(output, label)
            scaled = scaler.scale(loss)
            scaled.backward()
            scaler.step(optimizer)
            scaler.update()
            optimizer.clear_grad(set_to_zero=False)

            if use_save_load and i == 5:
1347 1348 1349 1350 1351 1352
                model_path = os.path.join(self.temp_dir.name, "model.pdparams")
                optimizer_path = os.path.join(self.temp_dir.name, "opt.pdopt")
                paddle.save(model.state_dict(), model_path)
                paddle.save(optimizer.state_dict(), optimizer_path)
                model.set_state_dict(paddle.load(model_path))
                optimizer.set_state_dict(paddle.load(optimizer_path))
1353 1354 1355 1356 1357 1358 1359 1360 1361 1362

        return loss.numpy()

    def test_with_state_dict(self):
        if core.is_compiled_with_cuda():
            with fluid.dygraph.guard():
                out_use_state_dict = self.check_with_opt_state_dict(
                    use_save_load=True)
                out_no_state_dict = self.check_with_opt_state_dict(
                    use_save_load=False)
1363
            np.testing.assert_array_equal(out_use_state_dict, out_no_state_dict)
1364 1365


Q
Qiao Longfei 已提交
1366
if __name__ == '__main__':
1367
    paddle.enable_static()
Q
Qiao Longfei 已提交
1368
    unittest.main()