test_dist_transpiler.py 33.5 KB
Newer Older
Y
Yancey 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15 16
from __future__ import print_function

T
tangwei12 已提交
17 18
import math

19
import unittest
20
import paddle.fluid as fluid
Y
Yancey 已提交
21
from paddle.fluid.transpiler.distribute_transpiler import delete_ops
W
Wu Yi 已提交
22
import traceback
G
gongweibao 已提交
23
import collections
M
minqiyang 已提交
24
import six
25

Y
Yancey 已提交
26

W
Wu Yi 已提交
27
class TranspilerTest(unittest.TestCase):
Y
Yancey 已提交
28
    def setUp(self):
W
Wu Yi 已提交
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
        self.trainer_id = 0
        self.trainers = 2
        self.pservers = 2
        # NOTE: we do not actually bind this port
        self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175"
        self.pserver1_ep = "127.0.0.1:6174"
        self.pserver2_ep = "127.0.0.1:6175"
        self.sync_mode = True
        self.transpiler = None

    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
        sgd_optimizer.minimize(avg_cost)

    def get_main_program(self):
        main = fluid.Program()
54
        main.random_seed = 1
W
Wu Yi 已提交
55 56 57 58 59
        with fluid.program_guard(main):
            self.net_conf()
        self.origin_prog = main.clone()
        return main

G
gongweibao 已提交
60 61 62 63 64
    def get_trainer(self, config=None):
        src = fluid.default_startup_program().clone()

        t = self._transpiler_instance(config)

W
Wu Yi 已提交
65
        trainer_main = t.get_trainer_program(wait_port=False)
G
gongweibao 已提交
66 67 68 69 70 71
        trainer_startup = fluid.default_startup_program()

        assert (src.num_blocks == 1)
        assert (trainer_startup.num_blocks == src.num_blocks)

        return trainer_main, trainer_startup
W
Wu Yi 已提交
72

Q
qiaolongfei 已提交
73 74
    def get_pserver(self, ep, config=None, sync_mode=True):
        t = self._transpiler_instance(config, sync_mode)
W
Wu Yi 已提交
75 76 77 78
        pserver = t.get_pserver_program(ep)
        startup = t.get_startup_program(ep, pserver)
        return pserver, startup

Q
qiaolongfei 已提交
79
    def _transpiler_instance(self, config=None, sync_mode=True):
W
Wu Yi 已提交
80 81
        if not self.transpiler:
            main = self.get_main_program()
G
gongweibao 已提交
82
            self.transpiler = fluid.DistributeTranspiler(config=config)
W
Wu Yi 已提交
83 84 85 86
            self.transpiler.transpile(
                self.trainer_id,
                program=main,
                pservers=self.pserver_eps,
Q
qiaolongfei 已提交
87 88
                trainers=self.trainers,
                sync_mode=sync_mode)
G
gongweibao 已提交
89

W
Wu Yi 已提交
90
        return self.transpiler
Y
Yancey 已提交
91

Q
qiaolongfei 已提交
92 93
    def transpiler_test_impl(self):
        pass
W
Wu Yi 已提交
94

Y
Yancey 已提交
95
    def test_transpiler(self):
Q
qiaolongfei 已提交
96 97
        main = fluid.Program()
        startup = fluid.Program()
T
tangwei12 已提交
98 99 100
        with fluid.unique_name.guard():
            with fluid.program_guard(main, startup):
                self.transpiler_test_impl()
Q
qiaolongfei 已提交
101 102 103 104


class TestBasicModel(TranspilerTest):
    def transpiler_test_impl(self):
W
Wu Yi 已提交
105 106 107
        pserver, startup = self.get_pserver(self.pserver1_ep)
        pserver2, startup2 = self.get_pserver(self.pserver2_ep)

G
gongweibao 已提交
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
        trainer, trainer_startup = self.get_trainer()

        # splited var blocks should be in startup program
        self.assertTrue("fc_w.block0" in trainer_startup.global_block().vars)
        self.assertTrue("fc_w.block1" in trainer_startup.global_block().vars)
        self.assertTrue("fc_w" in trainer_startup.global_block().vars)
        self.assertTrue("fc_b" in trainer_startup.global_block().vars)
        self.assertTrue("fc_w@GRAD" not in trainer_startup.global_block().vars)
        self.assertTrue("fc_b@GRAD" not in trainer_startup.global_block().vars)

        src = [op.type for op in trainer_startup.global_block().ops]
        dst = ['fill_constant', 'fill_constant', 'uniform_random', 'recv', 'recv', \
               'fetch_barrier', 'concat']

        self.assertEqual(src, dst)
W
Wu Yi 已提交
123 124 125 126 127 128 129

        self.assertEqual([op.type for op in trainer.global_block().ops], [
            'mul', 'elementwise_add', 'elementwise_sub', 'square', 'mean',
            'fill_constant', 'mean_grad', 'square_grad', 'elementwise_sub_grad',
            'elementwise_add_grad', 'send', 'mul_grad', 'split_byref', 'send',
            'send_barrier', 'recv', 'recv', 'fetch_barrier', 'concat'
        ])
Y
Yancey 已提交
130 131 132 133 134

        self.assertEqual(len(pserver.blocks), 3)
        # block0: listen_and_serv
        self.assertEqual([op.type for op in pserver.blocks[0].ops],
                         ["listen_and_serv"])
W
Wu Yi 已提交
135
        # block1~2: optimize pass
Y
Yancey 已提交
136 137 138
        self.assertEqual([op.type for op in pserver.blocks[1].ops],
                         ["sum", "scale", "sgd"])
        # confirm startup program
W
Wu Yi 已提交
139 140
        self.assertEqual([op.type for op in startup.global_block().ops],
                         ["fill_constant", "fill_constant", "uniform_random"])
Y
Yancey1989 已提交
141
        # the variable #fc_w will be split into two blocks
Y
Yancey 已提交
142 143
        fc_w_var = startup.global_block().var("fc_w.block1")
        self.assertEqual(fc_w_var.shape, (500, 1000))
W
Wu Yi 已提交
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164
        # all parameters should be optimized on pserver

        pserver_params = []
        for prog in [pserver, pserver2]:
            for blk in prog.blocks:
                for op in blk.ops:
                    if "Param" in op.input_names:
                        param_name = op.input("Param")[0]
                        is_block_idx = param_name.find(".block")
                        if is_block_idx != -1:
                            origin_param_name = param_name[:is_block_idx]
                        else:
                            origin_param_name = param_name
                        pserver_params.append(origin_param_name)
        trainer_params = []
        for op in self.origin_prog.global_block().ops:
            if "Param" in op.input_names:
                trainer_params.append(op.input("Param")[0])
        self.assertEqual(set(pserver_params), set(trainer_params))


G
gongweibao 已提交
165
class TestBasicModelWithLargeBlockSize(TranspilerTest):
Q
qiaolongfei 已提交
166
    def transpiler_test_impl(self):
G
gongweibao 已提交
167 168 169 170 171 172
        config = fluid.DistributeTranspilerConfig()
        config.min_block_size = 1048576

        pserver, startup = self.get_pserver(self.pserver1_ep, config)
        pserver2, startup2 = self.get_pserver(self.pserver2_ep, config)

G
gongweibao 已提交
173
        trainer, _ = self.get_trainer(config)
G
gongweibao 已提交
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190

        self.assertEqual([op.type for op in trainer.global_block().ops], [
            'mul', 'elementwise_add', 'elementwise_sub', 'square', 'mean',
            'fill_constant', 'mean_grad', 'square_grad', 'elementwise_sub_grad',
            'elementwise_add_grad', 'send', 'mul_grad', 'send', 'send_barrier',
            'recv', 'recv', 'fetch_barrier'
        ])

        self.assertEqual(len(pserver.blocks), 2)
        # block0: listen_and_serv
        self.assertEqual([op.type for op in pserver.blocks[0].ops],
                         ["listen_and_serv"])
        # block1~2: optimize pass
        self.assertEqual([op.type for op in pserver.blocks[1].ops],
                         ["sum", "scale", "sgd"])
        # confirm startup program
        self.assertEqual([op.type for op in startup.global_block().ops],
Q
qiaolongfei 已提交
191
                         ["fill_constant", "fill_constant"])
G
gongweibao 已提交
192 193
        # the variable #fc_w will be split into two blocks
        fc_w_var = startup2.global_block().var("fc_w")
194
        self.assertEqual(fc_w_var.shape, (1000, 1000))
G
gongweibao 已提交
195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
        # all parameters should be optimized on pserver

        pserver_params = []
        for prog in [pserver, pserver2]:
            for blk in prog.blocks:
                for op in blk.ops:
                    if "Param" in op.input_names:
                        param_name = op.input("Param")[0]
                        is_block_idx = param_name.find(".block")
                        if is_block_idx != -1:
                            origin_param_name = param_name[:is_block_idx]
                        else:
                            origin_param_name = param_name
                        pserver_params.append(origin_param_name)
        trainer_params = []
        for op in self.origin_prog.global_block().ops:
            if "Param" in op.input_names:
                trainer_params.append(op.input("Param")[0])
        self.assertEqual(set(pserver_params), set(trainer_params))


W
Wu Yi 已提交
216 217 218 219
class TestNoSliceVar(TranspilerTest):
    def setUp(self):
        super(TestNoSliceVar, self).setUp()

Q
qiaolongfei 已提交
220
    def transpiler_test_impl(self):
G
gongweibao 已提交
221 222 223 224 225
        config = fluid.DistributeTranspilerConfig()
        config.slice_var_up = False

        _, startup = self.get_pserver(self.pserver1_ep, config)
        _, startup2 = self.get_pserver(self.pserver2_ep, config)
W
Wu Yi 已提交
226

227
        if "fc_w" in startup.global_block().vars:
W
Wu Yi 已提交
228
            fc_w_var = startup.global_block().vars["fc_w"]
229
        elif "fc_w" in startup2.global_block().vars:
W
Wu Yi 已提交
230 231 232
            fc_w_var = startup2.global_block().vars["fc_w"]

        self.assertEqual(fc_w_var.shape, (1000, 1000))
Y
Yancey 已提交
233 234


W
Wu Yi 已提交
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
class TestLRDecay(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        sgd_optimizer = fluid.optimizer.SGD(
            learning_rate=fluid.layers.exponential_decay(
                learning_rate=1.0,
                decay_steps=2100,
                decay_rate=0.1,
                staircase=True))
        sgd_optimizer.minimize(avg_cost)

Q
qiaolongfei 已提交
254
    def transpiler_test_impl(self):
W
Wu Yi 已提交
255
        pserver, startup = self.get_pserver(self.pserver1_ep)
G
gongweibao 已提交
256
        trainer, _ = self.get_trainer()
W
Wu Yi 已提交
257 258 259 260 261 262 263 264 265 266

        self.assertEqual(len(pserver.blocks), 4)
        lr_decay_ops = [op.type for op in pserver.blocks[1].ops]
        self.assertEqual(lr_decay_ops, [
            "increment", "cast", "fill_constant", "elementwise_div", "floor",
            "fill_constant", "elementwise_pow", "fill_constant",
            "elementwise_mul"
        ])


267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285
class TestDecayedAdagrad(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        opt = fluid.optimizer.DecayedAdagrad(learning_rate=0.1)
        opt.minimize(avg_cost)

    def transpiler_test_impl(self):
        pserver, startup = self.get_pserver(self.pserver1_ep)
        trainer, _ = self.get_trainer()


286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304
class TestFtrl(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        opt = fluid.optimizer.Ftrl(learning_rate=0.1)
        opt.minimize(avg_cost)

    def transpiler_test_impl(self):
        pserver, startup = self.get_pserver(self.pserver1_ep)
        trainer, _ = self.get_trainer()


W
Wu Yi 已提交
305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320
class TestLRDecayConditional(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        sgd_optimizer = fluid.optimizer.SGD(
            learning_rate=fluid.layers.piecewise_decay([10000, 20000],
                                                       [1.0, 0.5, 1.0]))
        sgd_optimizer.minimize(avg_cost)

Q
qiaolongfei 已提交
321
    def transpiler_test_impl(self):
W
Wu Yi 已提交
322
        pserver, startup = self.get_pserver(self.pserver1_ep)
G
gongweibao 已提交
323
        trainer, _ = self.get_trainer()
W
Wu Yi 已提交
324 325 326 327

        serv_op = pserver.blocks[0].ops[0]
        sub_blocks = []
        optimize_blocks = []
G
gongweibao 已提交
328
        for b in serv_op.all_attrs()["optimize_blocks"]:
W
Wu Yi 已提交
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368
            optimize_blocks.append(b.idx)
        for b in pserver.blocks:
            if b.idx not in optimize_blocks:
                sub_blocks.append(b.idx)

        self.assertEqual(len(pserver.blocks), 7)
        lr_decay_ops = [op.type for op in pserver.blocks[1].ops]
        self.assertEqual(lr_decay_ops, [
            "increment", "cast", "fill_constant", "fill_constant", "less_than",
            "logical_not", "conditional_block", "fill_constant",
            "fill_constant", "less_than", "logical_not", "logical_and",
            "logical_and", "conditional_block", "fill_constant",
            "conditional_block"
        ])
        # test the condition blocks
        for b in sub_blocks:
            if b == 0:
                continue
            block = pserver.blocks[b]
            self.assertEqual([op.type for op in block.ops], ["assign"])


class TestL2Decay(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(
            input=x,
            size=1000,
            act=None,
            param_attr=fluid.ParamAttr(
                name='fc_w',
                regularizer=fluid.regularizer.L2Decay(),
                gradient_clip=fluid.clip.GradientClipByValue(0.1)),
            bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
        sgd_optimizer.minimize(avg_cost)

Q
qiaolongfei 已提交
369
    def transpiler_test_impl(self):
W
Wu Yi 已提交
370
        pserver, startup = self.get_pserver(self.pserver1_ep)
G
gongweibao 已提交
371
        trainer, _ = self.get_trainer()
W
Wu Yi 已提交
372 373 374 375 376 377 378 379 380

        self.assertEqual(len(pserver.blocks), 3)
        self.assertEqual([op.type for op in pserver.blocks[1].ops],
                         ["sum", "scale", "clip", "sgd"])
        self.assertEqual(
            [op.type for op in pserver.blocks[2].ops],
            ["sum", "scale", "clip", "scale", "elementwise_add", "sgd"])
        # TODO(typhoonzero): test clipping and L2Decay ops are removed from trainer

Y
Yancey 已提交
381

T
typhoonzero 已提交
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402
class TestL2DecayWithPiecewise(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        base_lr = 1.0
        bd = [1, 10, 20, 30]
        lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
        sgd_optimizer = fluid.optimizer.Momentum(
            learning_rate=fluid.layers.piecewise_decay(
                boundaries=bd, values=lr),
            momentum=0.9,
            regularization=fluid.regularizer.L2Decay(1e-4))
        sgd_optimizer.minimize(avg_cost)

Q
qiaolongfei 已提交
403
    def transpiler_test_impl(self):
T
typhoonzero 已提交
404
        pserver, startup = self.get_pserver(self.pserver1_ep)
G
gongweibao 已提交
405
        trainer, _ = self.get_trainer()
T
typhoonzero 已提交
406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424

        self.assertEqual(len(pserver.blocks), 9)
        self.assertEqual([op.type for op in pserver.blocks[1].ops], [
            "increment", "cast", "fill_constant", "fill_constant", "less_than",
            "logical_not", "conditional_block", "fill_constant",
            "fill_constant", "less_than", "logical_not", "logical_and",
            "logical_and", "conditional_block", "fill_constant",
            "fill_constant", "less_than", "logical_not", "logical_and",
            "logical_and", "conditional_block", "fill_constant",
            "fill_constant", "less_than", "logical_not", "logical_and",
            "logical_and", "conditional_block", "fill_constant",
            "conditional_block"
        ])
        self.assertEqual(
            [op.type for op in pserver.blocks[7].ops],
            ["sum", "scale", "scale", "elementwise_add", "momentum"])
        self.assertEqual(
            [op.type for op in pserver.blocks[8].ops],
            ["sum", "scale", "scale", "elementwise_add", "momentum"])
Y
Yancey 已提交
425 426


Q
Qiao Longfei 已提交
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 TestEmptyPserverOptimizeBlocks(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        # only one parameter
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=False)
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        sgd_optimizer = fluid.optimizer.SGD(learning_rate=1.0)
        sgd_optimizer.minimize(avg_cost)

    def transpiler_test_impl(self):
        config = fluid.DistributeTranspilerConfig()
        config.slice_var_up = False

        pserver, startup = self.get_pserver(ep=self.pserver2_ep, config=config)

        self.assertEqual(len(pserver.blocks), 2)
        self.assertEqual(len(pserver.blocks[1].ops), 0)


452 453
class TestDistLookupTableBase(TranspilerTest):
    def network_with_table(self, is_sparse, is_distributed):
T
tangwei12 已提交
454 455
        self.table_size = 1000
        self.emb_size = 64
T
tangwei12 已提交
456
        self.lookup_table_name = 'shared_w'
T
tangwei12 已提交
457

458
        def emb_pool(ids, table_name, is_distributed):
459 460
            emb = fluid.layers.embedding(
                input=ids,
T
tangwei12 已提交
461
                size=[self.table_size, self.emb_size],
462
                dtype='float32',
463
                param_attr=table_name,
464 465 466 467 468 469 470 471 472
                is_sparse=is_sparse,
                is_distributed=is_distributed)
            pool = fluid.layers.sequence_pool(input=emb, pool_type='average')
            return pool

        title_ids = fluid.layers.data(
            name='title_ids', shape=[1], dtype='int64', lod_level=1)
        brand_ids = fluid.layers.data(
            name='brand_ids', shape=[1], dtype='int64', lod_level=1)
473 474 475 476 477
        profile_ids = fluid.layers.data(
            name='brand_ids', shape=[1], dtype='int64', lod_level=1)
        title_emb = emb_pool(title_ids, self.lookup_table_name, is_distributed)
        brand_emb = emb_pool(brand_ids, self.lookup_table_name, is_distributed)
        profile_emb = emb_pool(profile_ids, "profile_emb", False)
Q
Qiao Longfei 已提交
478 479
        fc0 = fluid.layers.concat(
            input=[title_emb, brand_emb, profile_emb], axis=1)
480 481 482 483 484 485 486 487 488 489 490 491 492
        predict = fluid.layers.fc(input=fc0,
                                  size=2,
                                  act=None,
                                  param_attr=fluid.ParamAttr(name='fc_w'),
                                  bias_attr=fluid.ParamAttr(name='fc_b'))

        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
        cost = fluid.layers.cross_entropy(input=predict, label=label)
        avg_cost = fluid.layers.mean(cost)
        optimizer = fluid.optimizer.Adam(learning_rate=0.003)
        optimizer.minimize(avg_cost)


Q
qiaolongfei 已提交
493 494 495 496 497 498 499
class TestLocalLookupTable(TestDistLookupTableBase):
    def net_conf(self):
        self.network_with_table(is_sparse=True, is_distributed=False)

    def transpiler_test_impl(self):
        pserver1, startup1 = self.get_pserver(self.pserver1_ep)

500
        self.assertEqual(len(pserver1.blocks), 4)
Q
qiaolongfei 已提交
501 502 503 504 505 506 507
        # 0 listen_and_serv
        # 1 optimize for fc_w or fc_b adam
        self.assertEqual([op.type for op in pserver1.blocks[1].ops],
                         ["sum", "scale", "adam", "scale", "scale"])
        # 2 optimize for table adam
        # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
        self.assertEqual([op.type for op in pserver1.blocks[2].ops],
Q
qiaolongfei 已提交
508
                         ["sum", "scale", "adam", "scale", "scale"])
Q
qiaolongfei 已提交
509

510 511 512 513 514
        # 3 optimize for table 2 adam
        # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
        self.assertEqual([op.type for op in pserver1.blocks[3].ops],
                         ["sum", "scale", "adam", "scale", "scale"])

G
gongweibao 已提交
515
        trainer, _ = self.get_trainer()
Q
qiaolongfei 已提交
516 517 518
        self.assertEqual(len(trainer.blocks), 1)
        ops = [
            'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
Q
Qiao Longfei 已提交
519 520 521 522 523 524 525 526
            'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
            'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
            'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
            'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
            'split_selected_rows', 'send', 'sequence_pool_grad',
            'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
            'sum', 'split_selected_rows', 'send', 'send_barrier', 'recv',
            'recv', 'recv', 'recv', 'fetch_barrier', 'concat', 'concat'
Q
qiaolongfei 已提交
527 528 529 530
        ]
        self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)


531 532 533 534 535 536 537
class TestDistLookupTable(TestDistLookupTableBase):
    def net_conf(self):
        self.network_with_table(is_sparse=True, is_distributed=True)

    def transpiler_test_impl(self):
        pserver1, startup1 = self.get_pserver(self.pserver1_ep)

538
        self.assertEqual(len(pserver1.blocks), 6)
539 540 541 542
        # 0 listen_and_serv
        # 1 optimize for fc_w or fc_b adam
        self.assertEqual([op.type for op in pserver1.blocks[1].ops],
                         ["sum", "scale", "adam", "scale", "scale"])
543
        # 4 prefetch -> lookup_sparse_table for data0
544
        self.assertEqual([op.type for op in pserver1.blocks[2].ops],
545 546 547
                         ["sum", "scale", "adam", "scale", "scale"])
        # 2 optimize for table sgd
        self.assertEqual([op.type for op in pserver1.blocks[3].ops],
548 549
                         ["sum", "sgd"])
        # 3 prefetch -> lookup_sparse_table for data0
550
        self.assertEqual([op.type for op in pserver1.blocks[4].ops],
551
                         ["lookup_sparse_table"])
552 553
        # 5 save table
        self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
554

555
        trainer, trainer_startup = self.get_trainer()
556 557
        self.assertEqual(len(trainer.blocks), 1)
        ops = [
S
seiriosPlus 已提交
558
            'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
Q
Qiao Longfei 已提交
559 560 561 562 563 564 565 566 567
            'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
            'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
            'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
            'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
            'lookup_table_grad', 'split_selected_rows', 'send',
            'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
            'lookup_table_grad', 'sum', 'split_ids', 'send', 'send_barrier',
            'recv', 'recv', 'recv', 'fetch_barrier', 'concat'
        ]
568
        self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
569 570 571 572
        startup_ops = [
            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
Q
Qiao Longfei 已提交
573 574 575 576
            'fill_constant', 'fill_constant', 'uniform_random',
            'uniform_random', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat',
            'fake_init'
        ]
577 578 579
        self.assertEqual([op.type for op in trainer_startup.blocks[0].ops],
                         startup_ops)

580

Q
qiaolongfei 已提交
581 582 583 584 585 586
class TestAsyncLocalLookupTable(TestDistLookupTableBase):
    def net_conf(self):
        self.network_with_table(is_sparse=True, is_distributed=False)

    def transpiler_test_impl(self):
        config = fluid.DistributeTranspilerConfig()
Q
qiaolongfei 已提交
587
        pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False)
Q
qiaolongfei 已提交
588

589
        self.assertEqual(len(pserver1.blocks), 4)
Q
qiaolongfei 已提交
590 591 592 593 594 595 596 597
        # 0 listen_and_serv
        # 1 optimize for fc_w or fc_b adam
        self.assertEqual([op.type for op in pserver1.blocks[1].ops],
                         ["adam", "scale", "scale"])
        # 2 optimize for table adam
        # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
        self.assertEqual([op.type for op in pserver1.blocks[2].ops],
                         ["adam", "scale", "scale"])
598 599 600 601
        # 3 optimize for table adam
        # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
        self.assertEqual([op.type for op in pserver1.blocks[3].ops],
                         ["adam", "scale", "scale"])
Q
qiaolongfei 已提交
602

G
gongweibao 已提交
603
        trainer, _ = self.get_trainer(config)
Q
qiaolongfei 已提交
604 605 606
        self.assertEqual(len(trainer.blocks), 1)
        ops = [
            'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
607 608
            'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
            'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
Q
Qiao Longfei 已提交
609 610 611 612 613 614
            'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
            'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
            'split_selected_rows', 'send', 'sequence_pool_grad',
            'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
            'sum', 'split_selected_rows', 'send', 'recv', 'recv', 'recv',
            'recv', 'concat', 'concat'
Q
qiaolongfei 已提交
615 616 617 618
        ]
        self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)


Q
qiaolongfei 已提交
619 620 621 622 623 624 625
class TestAsyncDistLookupTable(TestDistLookupTableBase):
    def net_conf(self):
        self.network_with_table(is_sparse=True, is_distributed=True)

    def transpiler_test_impl(self):
        config = fluid.DistributeTranspilerConfig()

Q
qiaolongfei 已提交
626
        pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False)
Q
qiaolongfei 已提交
627

628
        self.assertEqual(len(pserver1.blocks), 6)
Q
qiaolongfei 已提交
629 630 631 632
        # 0 listen_and_serv
        # 1 optimize for fc_w or fc_b adam
        self.assertEqual([op.type for op in pserver1.blocks[1].ops],
                         ["adam", "scale", "scale"])
633 634 635 636 637 638 639
        # 2 optimize for table adam
        self.assertEqual([op.type for op in pserver1.blocks[2].ops],
                         ["adam", "scale", "scale"])
        # 3 optimize for table sgd
        self.assertEqual([op.type for op in pserver1.blocks[3].ops], ["sgd"])
        # 4 prefetch -> lookup_sparse_table for data0
        self.assertEqual([op.type for op in pserver1.blocks[4].ops],
Q
qiaolongfei 已提交
640
                         ["lookup_sparse_table"])
641 642
        # 5 save table
        self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
Q
qiaolongfei 已提交
643

G
gongweibao 已提交
644
        trainer, _ = self.get_trainer(config)
Q
qiaolongfei 已提交
645 646
        self.assertEqual(len(trainer.blocks), 1)
        ops = [
S
seiriosPlus 已提交
647
            'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
Q
Qiao Longfei 已提交
648 649 650 651 652 653 654 655 656
            'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
            'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
            'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
            'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
            'lookup_table_grad', 'split_selected_rows', 'send',
            'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
            'lookup_table_grad', 'sum', 'split_ids', 'send', 'recv', 'recv',
            'recv', 'concat'
        ]
Q
qiaolongfei 已提交
657 658 659
        self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)


T
tangwei12 已提交
660
class TestDistLookupTableSliceSize(TestDistLookupTableBase):
T
tangwei12 已提交
661 662 663 664 665
    def net_conf(self):
        self.network_with_table(is_sparse=True, is_distributed=True)

    def transpiler_test_impl(self):
        config = fluid.DistributeTranspilerConfig()
T
tangwei12 已提交
666
        pserver1, _ = self.get_pserver(self.pserver1_ep, config)
T
tangwei12 已提交
667 668 669 670 671 672 673

        self.assertTrue(self.transpiler.has_distributed_lookup_table)
        lookup_table_var = pserver1.global_block().vars[
            self.transpiler.table_name]
        row_size = lookup_table_var.shape[0]
        calc_row_size = int(math.ceil(self.table_size / self.pservers))
        self.assertEqual(row_size, calc_row_size)
T
tangwei12 已提交
674 675


T
tangwei12 已提交
676 677 678 679 680 681 682 683 684 685 686 687 688 689 690
class TestDistArgsInProgram(TestDistLookupTableBase):
    def net_conf(self):
        self.network_with_table(is_sparse=True, is_distributed=True)

    def transpiler_test_impl(self):
        trainer, _ = self.get_trainer()

        self.assertTrue(trainer._is_distributed)
        self.assertTrue(trainer._is_chief)
        self.assertEqual(trainer._distributed_lookup_table,
                         self.lookup_table_name)
        self.assertEqual(trainer._endpoints,
                         [self.pserver1_ep, self.pserver2_ep])


W
Wu Yi 已提交
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
class TestRMSPropOptimizer(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        optimizer = fluid.optimizer.RMSProp(learning_rate=0.1)
        optimizer.minimize(avg_cost)

    def transpiler_test_impl(self):
        pserver, startup = self.get_pserver(self.pserver1_ep)
        pserver2, startup2 = self.get_pserver(self.pserver2_ep)

        self.assertEqual(len(pserver.blocks), 3)
        # block1~2: optimize pass
        self.assertEqual([op.type for op in pserver.blocks[1].ops],
                         ["sum", "scale", "rmsprop"])
        # the variable #fc_w will be split into two blocks
        fc_w_var = startup.global_block().var("fc_w.block1")
        self.assertEqual(fc_w_var.shape, (500, 1000))
        moment_var = startup.global_block().var("momentum_1")
        self.assertEqual(moment_var.shape, (500, 1000))


T
tangwei12 已提交
720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737
class TestLoadSliceVar(TranspilerTest):
    def net_conf(self):
        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
        y_predict = fluid.layers.fc(input=x,
                                    size=1000,
                                    act=None,
                                    param_attr=fluid.ParamAttr(name='fc_w'),
                                    bias_attr=fluid.ParamAttr(name='fc_b'))
        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
        avg_cost = fluid.layers.mean(cost)
        optimizer = fluid.optimizer.RMSProp(learning_rate=0.1)
        optimizer.minimize(avg_cost)

    def transpiler_test_impl(self):
        pserver, _ = self.get_pserver(self.pserver1_ep)
        pserver2, _ = self.get_pserver(self.pserver2_ep)

T
tangwei12 已提交
738 739
        self.assertTrue(pserver._slice_vars_and_attrs)
        self.assertTrue(pserver2._slice_vars_and_attrs)
T
tangwei12 已提交
740

M
minqiyang 已提交
741
        for idx in six.moves.xrange(len(pserver._slice_vars_and_attrs)):
T
tangwei12 已提交
742 743
            self.assertEqual(pserver._slice_vars_and_attrs[idx][0],
                             pserver2._slice_vars_and_attrs[idx][0])
T
tangwei12 已提交
744

M
minqiyang 已提交
745 746
            total_numel = six.moves.reduce(
                lambda x, y: x * y, pserver._slice_vars_and_attrs[idx][0].shape)
T
tangwei12 已提交
747 748
            self.assertEqual(
                total_numel,
M
minqiyang 已提交
749 750 751 752
                six.moves.reduce(lambda x, y: x * y,
                                 pserver._slice_vars_and_attrs[idx][2].shape) +
                six.moves.reduce(lambda x, y: x * y,
                                 pserver2._slice_vars_and_attrs[idx][2].shape))
T
tangwei12 已提交
753 754


W
Wu Yi 已提交
755 756
class TestNCCL2Transpile(TranspilerTest):
    def test_nccl2_transpile(self):
J
JiabinYang 已提交
757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775
        if fluid.core.is_compiled_with_cuda():  #test nccl2 only with cuda
            main = fluid.Program()
            startup = fluid.Program()
            with fluid.program_guard(main, startup):
                self.net_conf()

            config = fluid.DistributeTranspilerConfig()
            config.mode = "nccl2"
            t = fluid.DistributeTranspiler(config=config)
            t.transpile(
                0,
                trainers="127.0.0.1:6174,127.0.0.1:6175",
                current_endpoint="127.0.0.1:6174",
                startup_program=startup)
            print([op.type for op in startup.global_block().ops])
            self.assertEqual(startup.global_block().ops[-1].type, "gen_nccl_id")
            self.assertIsNotNone(startup.global_block().vars.get("NCCLID"))
        else:
            pass
W
Wu Yi 已提交
776 777


Y
Yancey 已提交
778 779
if __name__ == "__main__":
    unittest.main()