test_optimizer.py 33.8 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
from __future__ import print_function

Q
Qiao Longfei 已提交
17 18
import unittest

19 20
import paddle.fluid.framework as framework
import paddle.fluid.optimizer as optimizer
M
mapingshuo 已提交
21
import paddle.compat as cpt
22
from paddle.fluid.backward import append_backward
Q
Qiao Longfei 已提交
23 24 25 26


class TestOptimizer(unittest.TestCase):
    def test_sgd_optimizer(self):
Q
qiaolongfei 已提交
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55
        def check_sgd_optimizer(optimizer_attr):
            init_program = framework.Program()
            program = framework.Program()
            block = program.global_block()
            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})
            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})
56 57
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
Q
Qiao Longfei 已提交
58

Q
qiaolongfei 已提交
59 60 61 62
        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
63

64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96
class TestOptimizerBackwardApplygrad(unittest.TestCase):
    def test_sgd_optimizer(self):
        def check_sgd_optimizer(optimizer_attr):
            init_program = framework.Program()
            program = framework.Program()
            block = program.global_block()
            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})
            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})
97 98
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
99 100 101 102 103 104

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


105 106 107 108 109 110 111 112
class TestMomentumOptimizer(unittest.TestCase):
    class MockMomentum(optimizer.MomentumOptimizer):
        def get_accumulators(self):
            return self._accumulators

        def get_velocity_str(self):
            return self._velocity_acc_str

113
    def test_vanilla_momentum_optimizer(self):
Q
Qiao Longfei 已提交
114
        init_program = framework.Program()
115 116 117
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
118 119 120 121 122
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
123 124 125 126 127 128 129 130 131 132
        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
133 134 135
        learning_rate = 0.01
        momentum_optimizer = self.MockMomentum(
            learning_rate=learning_rate, momentum=0.2)
136 137 138 139
        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
140
        params_grads = append_backward(mean_out)
141 142
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
143 144
        with framework.program_guard(program, init_program):
            opts = momentum_optimizer.apply_gradients(params_grads)
145
        self.assertEqual(len(opts), 2)
Y
Yancey1989 已提交
146
        sgd_op = opts[-1]
147
        self.assertEqual([op.type for op in opts], ["scale", "momentum"])
148
        self.assertFalse(sgd_op.attr('use_nesterov'))
149 150 151 152 153 154 155 156 157

        # 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
158 159 160 161 162 163 164 165
        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 2)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
        self.assertEqual(init_ops[1].type, "fill_constant")
        self.assertAlmostEqual(init_ops[1].attr('value'), 0.0)

166
    def test_nesterov_momentum_optimizer(self):
Q
Qiao Longfei 已提交
167
        init_program = framework.Program()
168 169 170
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
171 172 173 174 175
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
176 177 178 179 180 181 182 183 184 185
        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})
186 187 188 189
        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
190
        learning_rate = 0.01
191
        momentum_optimizer = self.MockMomentum(
Q
Qiao Longfei 已提交
192
            learning_rate=learning_rate, momentum=0.2, use_nesterov=True)
F
fengjiayi 已提交
193
        params_grads = append_backward(mean_out)
194 195
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
196 197
        with framework.program_guard(program, init_program):
            opts = momentum_optimizer.apply_gradients(params_grads)
198
        self.assertEqual(len(opts), 2)
Y
Yancey1989 已提交
199
        sgd_op = opts[-1]
200
        self.assertEqual([op.type for op in opts], ["scale", "momentum"])
201
        self.assertTrue(sgd_op.attr('use_nesterov'))
202 203 204 205 206 207 208 209 210

        # 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
211 212 213 214 215 216 217 218
        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 2)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
        self.assertEqual(init_ops[1].type, "fill_constant")
        self.assertAlmostEqual(init_ops[1].attr('value'), 0.0)

219

220 221 222 223 224 225 226 227 228
class TestAdagradOptimizer(unittest.TestCase):
    class MockAdagrad(optimizer.AdagradOptimizer):
        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
229
        init_program = framework.Program()
230 231 232
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
233 234 235 236 237
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
238 239 240 241 242 243 244 245 246 247
        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})
248 249 250 251
        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
252 253 254
        learning_rate = 0.01
        adagrad_optimizer = self.MockAdagrad(
            learning_rate=learning_rate, epsilon=1.0e-6)
F
fengjiayi 已提交
255
        params_grads = append_backward(mean_out)
256 257
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adagrad_optimizer.get_accumulators()), 0)
258 259
        with framework.program_guard(program, init_program):
            opts = adagrad_optimizer.apply_gradients(params_grads)
260 261
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "adagrad"])
262

263
        # Check accumulators
264 265 266 267 268 269 270
        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
271 272
        # Check init_program
        init_ops = init_program.global_block().ops
273
        self.assertEqual(len(init_ops), 3)
Q
Qiao Longfei 已提交
274 275 276 277 278
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
        self.assertEqual(init_ops[1].type, "fill_constant")
        self.assertAlmostEqual(init_ops[1].attr('value'), 0.0)

279

280 281 282 283 284 285 286 287 288 289 290 291
class TestAdamOptimizer(unittest.TestCase):
    class MockAdam(optimizer.AdamOptimizer):
        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
292
        init_program = framework.Program()
293 294 295
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
296 297 298 299 300
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
301 302 303 304 305 306 307 308 309 310
        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})
311 312 313 314
        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
315
        learning_rate = 0.01
316
        adam_optimizer = self.MockAdam(
Q
Qiao Longfei 已提交
317
            learning_rate=learning_rate, beta1=0.9, beta2=0.999)
F
fengjiayi 已提交
318
        params_grads = append_backward(mean_out)
319 320
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adam_optimizer.get_accumulators()), 0)
321 322
        with framework.program_guard(program, init_program):
            opts = adam_optimizer.apply_gradients(params_grads)
323 324 325
        self.assertEqual(len(opts), 4)
        self.assertEqual([op.type for op in opts],
                         ["scale", "adam", "scale", "scale"])
326 327 328

        # Check accumulators
        accumulators = adam_optimizer.get_accumulators()
Q
qiaolongfei 已提交
329
        self.assertEqual(len(accumulators), 4)
330 331 332 333 334 335 336 337 338
        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
339 340 341 342 343 344
        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 5)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)

345

346 347 348 349 350 351 352 353 354 355 356 357
class TestAdamaxOptimizer(unittest.TestCase):
    class MockAdamax(optimizer.AdamaxOptimizer):
        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
358
        init_program = framework.Program()
359 360 361
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
362 363 364 365 366
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
367 368 369 370 371 372 373 374 375 376
        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})
377 378 379 380
        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
381
        learning_rate = 0.01
382
        adamax_optimizer = self.MockAdamax(
Q
Qiao Longfei 已提交
383
            learning_rate=learning_rate, beta1=0.9, beta2=0.999)
F
fengjiayi 已提交
384
        params_grads = append_backward(mean_out)
385 386
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adamax_optimizer.get_accumulators()), 0)
387 388
        with framework.program_guard(program, init_program):
            opts = adamax_optimizer.apply_gradients(params_grads)
389 390
        self.assertEqual(len(opts), 3)
        self.assertEqual([op.type for op in opts], ["scale", "adamax", "scale"])
391 392 393

        # Check accumulators
        accumulators = adamax_optimizer.get_accumulators()
Q
qiaolongfei 已提交
394
        self.assertEqual(len(accumulators), 3)
395 396 397 398 399 400 401 402 403
        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
404 405 406 407 408 409
        # Check init_program
        init_ops = init_program.global_block().ops
        self.assertEqual(len(init_ops), 4)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)

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
class TestDpsgdOptimizer(unittest.TestCase):
    def test_dpsgd_optimizer(self):
        def check_dpsgd_optimizer(optimizer_attr):
            init_program = framework.Program()
            program = framework.Program()
            block = program.global_block()
            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)
            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"])


452 453 454 455 456 457 458 459 460 461 462 463 464
class TestDecayedAdagradOptimizer(unittest.TestCase):
    class MockDecayedAdagrad(optimizer.DecayedAdagradOptimizer):
        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()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
465 466 467 468 469
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
470 471 472 473 474 475 476 477 478 479
        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})
480 481 482 483
        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})
484 485 486
        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
487
        params_grads = append_backward(mean_out)
488 489
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(decayed_adagrad_optimizer.get_accumulators()), 0)
490 491
        with framework.program_guard(program, init_program):
            opts = decayed_adagrad_optimizer.apply_gradients(params_grads)
492 493
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "decayed_adagrad"])
494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512

        # 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[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
        self.assertEqual(init_ops[1].type, "fill_constant")
        self.assertAlmostEqual(init_ops[1].attr('value'), 0.0)


Q
qiaolongfei 已提交
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
class TestFtrlOptimizer(unittest.TestCase):
    class MockFtrl(optimizer.FtrlOptimizer):
        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()
        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})
        learning_rate = 0.01
        ftrl_optimizer = self.MockFtrl(
            learning_rate=learning_rate, l1=0.0, l2=0.0, lr_power=-0.5)
        params_grads = append_backward(mean_out)
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(ftrl_optimizer.get_accumulators()), 0)
554 555
        with framework.program_guard(program, init_program):
            opts = ftrl_optimizer.apply_gradients(params_grads)
556 557
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "ftrl"])
Q
qiaolongfei 已提交
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577

        # 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)
        self.assertEqual(init_ops[0].type, "fill_constant")
        self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)


M
mapingshuo 已提交
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
class TestLookaheadOptimizer(unittest.TestCase):
    def test_lookahead_optimizer(self):
        init_program = framework.Program()
        program = framework.Program()
        block = program.global_block()
        init_block = init_program.global_block()
        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})

        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)
612 613
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
M
mapingshuo 已提交
614 615


M
mapingshuo 已提交
616
class TestRecomputeOptimizer(unittest.TestCase):
617
    def net(self, return_input=False):
M
mapingshuo 已提交
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
        program = framework.Program()
        block = program.global_block()
        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")
        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})
        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})

655 656
        if return_input == True:
            return mul_x, mul_out, b1_out, b2_out, mean_out
M
mapingshuo 已提交
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
        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)
        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"
        ])

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
    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
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
    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
        params_grads = recompute_optimizer.backward(
            mean_out,
            startup_program=None,
            parameter_list=None,
            no_grad_set=None,
            checkpoints=[b1_out])

        # 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:
            stat_dict = {}
            recompute_optimizer.load(stat_dict)
        except NotImplementedError as e:
            self.assertEqual(
                "load function is not supported by Recompute Optimizer for now",
                cpt.get_exception_message(e))


Q
Qiao Longfei 已提交
803 804
if __name__ == '__main__':
    unittest.main()