test_optimizer.py 44.0 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
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
26
from paddle.fluid.framework import Program, program_guard
Q
Qiao Longfei 已提交
27 28 29 30


class TestOptimizer(unittest.TestCase):
    def test_sgd_optimizer(self):
Q
qiaolongfei 已提交
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
        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})
60 61
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
Q
Qiao Longfei 已提交
62

Q
qiaolongfei 已提交
63 64 65 66
        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
67

68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
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})
101 102
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
103 104 105 106 107 108

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


109 110 111 112 113 114 115 116
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

117
    def test_vanilla_momentum_optimizer(self):
Q
Qiao Longfei 已提交
118
        init_program = framework.Program()
119 120 121
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
122 123 124 125 126
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
127 128 129 130 131 132 133 134 135 136
        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
137 138 139
        learning_rate = 0.01
        momentum_optimizer = self.MockMomentum(
            learning_rate=learning_rate, momentum=0.2)
140 141 142 143
        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
144
        params_grads = append_backward(mean_out)
145 146
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
147 148
        with framework.program_guard(program, init_program):
            opts = momentum_optimizer.apply_gradients(params_grads)
149
        self.assertEqual(len(opts), 2)
Y
Yancey1989 已提交
150
        sgd_op = opts[-1]
151
        self.assertEqual([op.type for op in opts], ["scale", "momentum"])
152
        self.assertFalse(sgd_op.attr('use_nesterov'))
153 154 155 156 157 158 159 160 161

        # 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
162 163 164 165 166 167 168 169
        # 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)

170
    def test_nesterov_momentum_optimizer(self):
Q
Qiao Longfei 已提交
171
        init_program = framework.Program()
172 173 174
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
175 176 177 178 179
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
180 181 182 183 184 185 186 187 188 189
        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})
190 191 192 193
        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
194
        learning_rate = 0.01
195
        momentum_optimizer = self.MockMomentum(
Q
Qiao Longfei 已提交
196
            learning_rate=learning_rate, momentum=0.2, use_nesterov=True)
F
fengjiayi 已提交
197
        params_grads = append_backward(mean_out)
198 199
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
200 201
        with framework.program_guard(program, init_program):
            opts = momentum_optimizer.apply_gradients(params_grads)
202
        self.assertEqual(len(opts), 2)
Y
Yancey1989 已提交
203
        sgd_op = opts[-1]
204
        self.assertEqual([op.type for op in opts], ["scale", "momentum"])
205
        self.assertTrue(sgd_op.attr('use_nesterov'))
206 207 208 209 210 211 212 213 214

        # 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
215 216 217 218 219 220 221 222
        # 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)

223

224 225 226 227 228 229 230 231 232
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
233
        init_program = framework.Program()
234 235 236
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
237 238 239 240 241
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
242 243 244 245 246 247 248 249 250 251
        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})
252 253 254 255
        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
256 257 258
        learning_rate = 0.01
        adagrad_optimizer = self.MockAdagrad(
            learning_rate=learning_rate, epsilon=1.0e-6)
F
fengjiayi 已提交
259
        params_grads = append_backward(mean_out)
260 261
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adagrad_optimizer.get_accumulators()), 0)
262 263
        with framework.program_guard(program, init_program):
            opts = adagrad_optimizer.apply_gradients(params_grads)
264 265
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "adagrad"])
266

267
        # Check accumulators
268 269 270 271 272 273 274
        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
275 276
        # Check init_program
        init_ops = init_program.global_block().ops
Z
zhongpu 已提交
277
        self.assertEqual(len(init_ops), 2)
Q
Qiao Longfei 已提交
278 279 280 281 282
        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)

283

284 285 286 287 288 289 290 291 292 293 294 295
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
296
        init_program = framework.Program()
297 298 299
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
300 301 302 303 304
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
305 306 307 308 309 310 311 312 313 314
        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})
315 316 317 318
        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
319
        learning_rate = 0.01
320
        adam_optimizer = self.MockAdam(
Q
Qiao Longfei 已提交
321
            learning_rate=learning_rate, beta1=0.9, beta2=0.999)
F
fengjiayi 已提交
322
        params_grads = append_backward(mean_out)
323 324
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adam_optimizer.get_accumulators()), 0)
325 326
        with framework.program_guard(program, init_program):
            opts = adam_optimizer.apply_gradients(params_grads)
A
Aurelius84 已提交
327 328
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "adam"])
329 330 331

        # Check accumulators
        accumulators = adam_optimizer.get_accumulators()
Q
qiaolongfei 已提交
332
        self.assertEqual(len(accumulators), 4)
333 334 335 336 337 338 339 340 341
        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
342 343 344 345 346 347
        # 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)

348

349 350 351 352 353 354 355 356 357 358 359 360
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
361
        init_program = framework.Program()
362 363 364
        program = framework.Program()
        block = program.global_block()
        mul_x = block.create_parameter(
Q
qiaolongfei 已提交
365 366 367 368 369
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
370 371 372 373 374 375 376 377 378 379
        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})
380 381 382 383
        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
384
        learning_rate = 0.01
385
        adamax_optimizer = self.MockAdamax(
Q
Qiao Longfei 已提交
386
            learning_rate=learning_rate, beta1=0.9, beta2=0.999)
F
fengjiayi 已提交
387
        params_grads = append_backward(mean_out)
388 389
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(adamax_optimizer.get_accumulators()), 0)
390 391
        with framework.program_guard(program, init_program):
            opts = adamax_optimizer.apply_gradients(params_grads)
392 393
        self.assertEqual(len(opts), 3)
        self.assertEqual([op.type for op in opts], ["scale", "adamax", "scale"])
394 395 396

        # Check accumulators
        accumulators = adamax_optimizer.get_accumulators()
Q
qiaolongfei 已提交
397
        self.assertEqual(len(accumulators), 3)
398 399 400 401 402 403 404 405 406
        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
407 408 409 410 411 412
        # 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)

413

414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454
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"])


455 456 457 458 459 460 461 462 463 464 465 466 467
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
468 469 470 471 472
            dtype="float32",
            shape=[5, 10],
            lod_level=0,
            name="mul.x",
            optimize_attr={'learning_rate': 1.1})
473 474 475 476 477 478 479 480 481 482
        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})
483 484 485 486
        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})
487 488 489
        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
490
        params_grads = append_backward(mean_out)
491 492
        self.assertEqual(len(params_grads), 1)
        self.assertEqual(len(decayed_adagrad_optimizer.get_accumulators()), 0)
493 494
        with framework.program_guard(program, init_program):
            opts = decayed_adagrad_optimizer.apply_gradients(params_grads)
495 496
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "decayed_adagrad"])
497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515

        # 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
516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556
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)
557 558
        with framework.program_guard(program, init_program):
            opts = ftrl_optimizer.apply_gradients(params_grads)
559 560
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "ftrl"])
Q
qiaolongfei 已提交
561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580

        # 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
581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614
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)
615 616
        self.assertEqual(len(opts), 2)
        self.assertEqual([op.type for op in opts], ["scale", "sgd"])
M
mapingshuo 已提交
617 618


M
mapingshuo 已提交
619
class TestRecomputeOptimizer(unittest.TestCase):
M
mapingshuo 已提交
620
    def net(self, return_input=False, with_dropout=False):
M
mapingshuo 已提交
621 622 623 624 625 626 627 628
        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")
M
mapingshuo 已提交
629 630 631 632 633 634 635 636
        if with_dropout == True:
            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")
M
mapingshuo 已提交
637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652
        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})
M
mapingshuo 已提交
653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670
        if with_dropout == True:
            block.append_op(
                type='dropout',
                inputs={'X': [mul_out]},
                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})
        else:
            block.append_op(
                type="elementwise_add",
                inputs={"X": mul_out,
                        "Y": b1},
                outputs={"Out": b1_out})
M
mapingshuo 已提交
671 672 673 674 675 676 677 678
        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})

679 680
        if return_input == True:
            return mul_x, mul_out, b1_out, b2_out, mean_out
M
mapingshuo 已提交
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
        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)
711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727
        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
728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768
        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"
        ])

769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804
    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
805 806 807 808 809 810 811 812 813 814
    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,
815
            no_grad_set=None)
M
mapingshuo 已提交
816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841

        # 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))

M
mapingshuo 已提交
842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864
    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"
        ])

865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966
    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'),
                "y": np.random.randint(
                    2, size=(100, 1)).astype('int64')
            }

        def mlp(input_x, input_y):
            drop_res = fluid.layers.dropout(
                input_x, dropout_prob=0.5, name="dropout_with_seed_cpu")
            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):
                input_x = fluid.layers.data(
                    name="x", shape=[3], dtype='float32')
                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):
    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'),
                "y": np.random.randint(
                    2, size=(100, 1)).astype('int64')
            }

        def mlp(input_x, input_y):
            drop_res = fluid.layers.dropout(
                input_x, dropout_prob=0.5, name="dropout_with_seed_gpu")
            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):
                input_x = fluid.layers.data(
                    name="x", shape=[3], dtype='float32')
                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
967

968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044
class TestGradientMergeOptimizer(unittest.TestCase):
    def net(self):
        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")
        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})
        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)

        self.assertEqual(main_program.num_blocks, 4)

        # main block
        self.assertEqual(len(cost.block.ops), 17)
        self.assertEqual([op.type for op in cost.block.ops], [
            'mul', 'elementwise_add', 'mean', 'fill_constant', 'mean_grad',
            'elementwise_add_grad', 'mul_grad', 'increment', 'fill_constant',
            'fill_constant', 'elementwise_mod', 'cast', 'not_equal',
            'logical_not', 'conditional_block', 'conditional_block',
            'conditional_block_grad'
        ])

        # merge block
        self.assertEqual(len(main_program.block(1).ops), 2)
        self.assertEqual([op.type for op in main_program.block(1).ops], [
            'elementwise_add',
            'elementwise_add',
        ])

        # reset block
        self.assertEqual(len(main_program.block(2).ops), 6)
        self.assertEqual([op.type for op in main_program.block(2).ops], [
            'elementwise_add', 'scale', 'elementwise_add', 'scale',
            'fill_constant', 'fill_constant'
        ])

        # optimize block
        self.assertEqual(len(main_program.block(3).ops), 2)
        self.assertEqual([op.type for op in main_program.block(3).ops],
                         ['sgd', 'sgd'])


Q
Qiao Longfei 已提交
1045 1046
if __name__ == '__main__':
    unittest.main()