distribute_transpiler.py 60.0 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
# 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.
14 15 16 17 18
"""
Steps to transpile trainer:
1. split variable to multiple blocks, aligned by product(dim[1:]) (width).
2. rename splited grad variables to add trainer_id suffix ".trainer_%d".
3. modify trainer program add split_op to each grad variable.
Q
Qiyang Min 已提交
19
4. append send_op to send splited variables to server and
20 21
5. add recv_op to fetch params(splited blocks or origin param) from server.
6. append concat_op to merge splited blocks to update local weights.
22 23 24 25 26 27 28 29

Steps to transpile pserver:
1. create new program for parameter server.
2. create params and grad variables that assigned to current server instance.
3. create a sub-block in the server side program
4. append ops that should run on current server instance.
5. add listen_and_serv op
"""
D
dzhwinter 已提交
30

T
typhoonzero 已提交
31
from __future__ import print_function
32

T
typhoonzero 已提交
33
import math
34
import numpy as np
35

Y
Yancey1989 已提交
36
from ps_dispatcher import RoundRobin, HashName, PSDispatcher
Y
Yancey 已提交
37
from .. import core, framework
T
typhoonzero 已提交
38
from ..framework import Program, default_main_program, \
Q
Qiyang Min 已提交
39
                        default_startup_program, Block, \
T
typhoonzero 已提交
40
                        Variable, Parameter, grad_var_name
41
from details import *
42 43 44

LOOKUP_TABLE_TYPE = "lookup_table"
LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
45
OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
Y
Yancey1989 已提交
46 47 48
RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName(
)
RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC
T
done  
typhoonzero 已提交
49 50


T
typhoonzero 已提交
51 52 53 54 55 56
class VarBlock:
    def __init__(self, varname, offset, size):
        self.varname = varname
        # NOTE: real offset is offset * size
        self.offset = offset
        self.size = size
T
done  
typhoonzero 已提交
57

T
typhoonzero 已提交
58 59
    def __str__(self):
        return "%s:%d:%d" % (self.varname, self.offset, self.size)
T
done  
typhoonzero 已提交
60 61


62 63 64 65
def same_or_split_var(p_name, var_name):
    return p_name == var_name or p_name.startswith(var_name + ".block")


66
def slice_variable(var_list, slice_count, min_block_size=8192):
T
typhoonzero 已提交
67
    """
68 69 70 71 72 73
    We may need to split dense tensor to one or more blocks and put
    them equally onto parameter server. One block is a sub-tensor
    aligned by dim[0] of the tensor.

    We need to have a minimal block size so that the calculations in
    the parameter server side can gain better performance. By default
74
    minimum block size 8K elements (maybe 16bit or 32bit or 64bit).
75 76 77

    Args:
        var_list (list): List of variables.
78 79
        slice_count (int): Numel of count that variables will be sliced, which
            could be the pserver services' count.
80 81
        min_block_size (int): Minimum splitted block size.
    Returns:
82
        blocks (list[(varname, block_id, current_block_size)]): A list
83
            of VarBlocks. Each VarBlock specifies a shard of the var.
T
typhoonzero 已提交
84 85 86
    """
    blocks = []
    for var in var_list:
87
        split_count = slice_count
T
typhoonzero 已提交
88 89 90 91
        var_numel = reduce(lambda x, y: x * y, var.shape)
        max_pserver_count = int(math.floor(var_numel / float(min_block_size)))
        if max_pserver_count == 0:
            max_pserver_count = 1
92
        if max_pserver_count < slice_count:
T
typhoonzero 已提交
93 94 95 96 97 98 99 100 101
            split_count = max_pserver_count
        block_size = int(math.ceil(var_numel / float(split_count)))

        if len(var.shape) >= 2:
            # align by dim1(width)
            dim1 = reduce(lambda x, y: x * y, var.shape[1:])
            remains = block_size % dim1
            if remains != 0:
                block_size += dim1 - remains
102
        # update split_count after aligning
T
typhoonzero 已提交
103 104 105 106 107 108 109 110 111
        split_count = int(math.ceil(var_numel / float(block_size)))
        for block_id in xrange(split_count):
            curr_block_size = min(block_size, var_numel - (
                (block_id) * block_size))
            block = VarBlock(var.name, block_id, curr_block_size)
            blocks.append(str(block))
    return blocks


Y
gen rst  
yi.wu 已提交
112
class DistributeTranspiler(object):
Y
yi.wu 已提交
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146
    """
    **DistributeTranspiler**

    Convert the fluid program to distributed data-parallelism programs.

    The main_program will be transformed to use a remote parameter server
    to do parameter optimization. And the optimization graph will be put
    into a parameter server program.

    Examples:
        .. code-block:: python

           # Define your model before these codes.
           port = os.getenv("PADDLE_PSERVER_PORT", "6174")
           pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "")
           eplist = []
           for ip in pserver_ips.split(","):
                eplist.append(':'.join([ip, port]))
           pserver_endpoints = ",".join(eplist)
           trainers = int(os.getenv("PADDLE_TRAINERS"))
           current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port
           trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
           role = os.getenv("PADDLE_TRAINING_ROLE")

           t = distribute_transpiler.DistributeTranspiler()
           t.transpile(
                trainer_id, pservers=pserver_endpoints, trainers=trainers)
           if role == "PSERVER":
                pserver_program = t.get_pserver_program(current_endpoint)
                pserver_startup_program = t.get_startup_program(current_endpoint,
                                                                pserver_program)
           elif role == "TRAINER":
                trainer_program = t.get_trainer_program()
    """
Y
Yancey1989 已提交
147

148 149 150 151 152
    def transpile(self,
                  trainer_id,
                  program=None,
                  pservers="127.0.0.1:6174",
                  trainers=1,
153
                  slice_var_up=True,
154 155 156
                  split_method=RoundRobin,
                  sync_mode=True):
        """
Y
yi.wu 已提交
157 158 159 160 161 162 163 164 165 166 167 168 169 170
        Run the transpiler.

        Args:
            trainer_id (int): id for current trainer worker, if you have
                n workers, the id may range from 0 ~ n-1
            program (Program|None): program to transpile,
                default is fluid.default_main_program().
            pservers (str): comma separated ip:port string for the pserver
                list.
            trainers (int): number of trainers in the distributed job.
            slice_var_up (bool): Do Tensor slice for pservers, default is True.
            split_method (PSDispatcher): RoundRobin or HashName can be used
                try to choose the best method to balance loads for pservers.
            sync_mode (bool): Do sync training or not, default is True.
171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186
        """
        assert (split_method.__bases__[0] == PSDispatcher)
        if program is None:
            program = default_main_program()
        self.origin_program = program
        self.trainer_num = trainers
        self.sync_mode = sync_mode
        self.trainer_id = trainer_id
        pserver_endpoints = pservers.split(",")
        self.pserver_endpoints = pserver_endpoints
        self.optimize_ops, self.params_grads = self._get_optimize_pass()

        ps_dispatcher = split_method(self.pserver_endpoints)
        self.has_distributed_lookup_table = self._has_distributed_lookup_table()

        # split and create vars, then put splited vars in dicts for later use.
187
        self._init_splited_vars(slice_var_up)
188

Y
Yancey1989 已提交
189 190
        # step 3.1: insert send op to send gradient vars to parameter servers
        ps_dispatcher.reset()
Y
update  
Yancey1989 已提交
191
        send_vars = []
192 193 194 195 196 197

        # in general cases, the number of pservers is times of 2, and this
        # will lead to uneven distribution among weights and bias:
        #       fc_w@GRAD_trainer_0, fc_w@GRAD_trainer_1 --> pserver1
        #       fc_b@GRAD_trainer_0, fc_b@GRAD_trainer_1 --> pserver2
        # shuffle the map will avoid the uneven distribution above
198
        grad_var_mapping_items = self.grad_var_mapping.items()
199
        if not slice_var_up:
200 201 202
            np.random.shuffle(grad_var_mapping_items)

        for orig_varname, splited_vars in grad_var_mapping_items:
Y
update  
Yancey1989 已提交
203
            eplist = ps_dispatcher.dispatch(splited_vars)
204

205
            if not slice_var_up:
206 207
                assert (len(splited_vars) == 1)

Y
Yancey1989 已提交
208 209 210 211 212 213 214 215 216
            if len(splited_vars) == 1:
                orig_varname = splited_vars[0].name
                index = find_op_by_output_arg(program.global_block(),
                                              orig_varname)
            elif len(splited_vars) > 1:
                orig_var = program.global_block().vars[orig_varname]
                index = find_op_by_output_arg(program.global_block(),
                                              orig_varname)
                self._insert_split_op(program, orig_var, index, splited_vars)
Y
update  
Yancey1989 已提交
217
                index += 1
Y
Yancey1989 已提交
218 219 220 221
            else:
                AssertionError("Can not insert the send op by original "
                               "variable name :", orig_varname)

Y
Yancey1989 已提交
222
            program.global_block().insert_op(
Y
update  
Yancey1989 已提交
223
                index=index + 1,
224
                type="send",
Y
update  
Yancey1989 已提交
225
                inputs={"X": splited_vars},
Y
Yancey1989 已提交
226 227 228 229 230
                outputs={},
                attrs={
                    "epmap": eplist,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })
Y
update  
Yancey1989 已提交
231 232
            for _, var in enumerate(splited_vars):
                send_vars.append(var)
Y
Yancey1989 已提交
233 234 235 236 237

        if self.sync_mode:
            program.global_block().append_op(
                type="send_barrier",
                inputs={},
Y
Yancey1989 已提交
238
                outputs={},
Y
Yancey1989 已提交
239 240
                attrs={
                    "endpoints": pserver_endpoints,
Y
Yancey1989 已提交
241 242
                    "sync_mode": self.sync_mode,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
Y
Yancey1989 已提交
243
                })
Y
Yancey1989 已提交
244 245 246

        # step 3.2: insert recv op to receive parameters from parameter server
        recv_vars = []
Y
update  
Yancey1989 已提交
247
        for _, var in enumerate(send_vars):
248
            recv_vars.append(self.grad_param_mapping[var])
Y
update  
Yancey1989 已提交
249
        ps_dispatcher.reset()
Y
Yancey1989 已提交
250 251
        eplist = ps_dispatcher.dispatch(recv_vars)

T
typhoonzero 已提交
252
        for i, ep in enumerate(eplist):
Y
Yancey1989 已提交
253 254
            self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
            self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])
255

Y
Yancey1989 已提交
256
        # step4: Concat the parameters splits together after recv.
257
        for varname, splited_var in self.param_var_mapping.iteritems():
Y
Yancey1989 已提交
258 259 260 261 262 263 264 265
            eps = []
            for var in splited_var:
                index = [v.name for v in recv_vars].index(var.name)
                eps.append(eplist[index])

            program.global_block().append_op(
                type="recv",
                inputs={},
Y
Yancey1989 已提交
266 267 268 269 270
                outputs={"Out": splited_var},
                attrs={
                    "epmap": eps,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })
T
typhoonzero 已提交
271

T
typhoonzero 已提交
272
        program.global_block().append_op(
Y
Yancey1989 已提交
273 274
            type="fetch_barrier",
            inputs={},
Y
Yancey1989 已提交
275
            outputs={},
Q
qiaolongfei 已提交
276 277
            attrs={
                "endpoints": pserver_endpoints,
Y
Yancey1989 已提交
278
                RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
Q
qiaolongfei 已提交
279
            })
Y
Yancey1989 已提交
280

281
        for varname, splited_var in self.param_var_mapping.iteritems():
T
typhoonzero 已提交
282 283
            if len(splited_var) <= 1:
                continue
T
typhoonzero 已提交
284
            orig_param = program.global_block().vars[varname]
T
typhoonzero 已提交
285
            program.global_block().append_op(
T
typhoonzero 已提交
286
                type="concat",
T
typhoonzero 已提交
287
                inputs={"X": splited_var},
T
typhoonzero 已提交
288
                outputs={"Out": [orig_param]},
T
typhoonzero 已提交
289
                attrs={"axis": 0})
T
typhoonzero 已提交
290

291
        if self.has_distributed_lookup_table:
Q
update  
qiaolongfei 已提交
292 293
            self._replace_lookup_table_op_with_prefetch(program,
                                                        pserver_endpoints)
Y
Yancey1989 已提交
294
            self._split_table_grad_and_add_send_vars(program, pserver_endpoints)
295

T
typhoonzero 已提交
296
    def get_trainer_program(self):
Y
yi.wu 已提交
297 298 299 300 301 302
        """
        Get transpiled trainer side program.

        Returns:
            Program: trainer side program.
        """
T
typhoonzero 已提交
303
        # remove optimize ops and add a send op to main_program
304
        delete_ops(self.origin_program.global_block(), self.optimize_ops)
305
        # FIXME(typhoonzero): serialize once will fix error occurs when clone.
306 307
        self.origin_program.__str__()
        return self.origin_program
T
typhoonzero 已提交
308 309 310

    def get_pserver_program(self, endpoint):
        """
Y
yi.wu 已提交
311 312 313 314 315 316 317
        Get parameter server side program.
        
        Args:
            endpoint (str): current parameter server endpoint.
        
        Returns:
            Program: the program for current parameter server to run.
T
typhoonzero 已提交
318
        """
Y
yi.wu 已提交
319 320 321 322 323
        # TODO(panyx0718): Revisit this assumption. what if #blocks > #pservers.
        # NOTE: assume blocks of the same variable is not distributed
        # on the same pserver, only change param/grad varnames for
        # trainers to fetch.

T
typhoonzero 已提交
324 325
        # step1
        pserver_program = Program()
326
        # step2: Create vars to receive vars at parameter servers.
T
typhoonzero 已提交
327 328 329 330 331 332 333 334
        recv_inputs = []
        for v in self.param_grad_ep_mapping[endpoint]["params"]:
            self._clone_var(pserver_program.global_block(), v)
        for v in self.param_grad_ep_mapping[endpoint]["grads"]:
            # create vars for each trainer in global scope, so
            # we don't need to create them when grad arrives.
            # change client side var name to origin name by
            # removing ".trainer_%d" suffix
T
update  
typhoonzero 已提交
335 336 337 338 339
            suff_idx = v.name.find(".trainer_")
            if suff_idx >= 0:
                orig_var_name = v.name[:suff_idx]
            else:
                orig_var_name = v.name
T
typhoonzero 已提交
340 341 342 343 344 345 346 347 348
            # NOTE: single_trainer_var must be created for multi-trainer
            # case to merge grads from multiple trainers
            single_trainer_var = \
                pserver_program.global_block().create_var(
                    name=orig_var_name,
                    persistable=True,
                    type=v.type,
                    dtype=v.dtype,
                    shape=v.shape)
349
            if self.sync_mode and self.trainer_num > 1:
350
                for trainer_id in xrange(self.trainer_num):
T
typhoonzero 已提交
351 352 353 354 355 356 357 358 359
                    var = pserver_program.global_block().create_var(
                        name="%s.trainer_%d" % (orig_var_name, trainer_id),
                        persistable=False,
                        type=v.type,
                        dtype=v.dtype,
                        shape=v.shape)
                    recv_inputs.append(var)
            else:
                recv_inputs.append(single_trainer_var)
360

Q
qiaolongfei 已提交
361
        # step 3
362
        # Create a union-find data structure from optimize ops,
T
typhoonzero 已提交
363 364 365
        # If two ops are connected, we could add these two ops
        # into one set.
        ufind = self._create_ufind(self.optimize_ops)
Q
qiaolongfei 已提交
366
        # step 3.2
T
typhoonzero 已提交
367 368 369 370
        # Iterate through the ops and append optimize op which
        # located on current pserver
        opt_op_on_pserver = []
        for _, op in enumerate(self.optimize_ops):
371 372
            if self._is_optimizer_op(op) and self._is_opt_op_on_pserver(
                    endpoint, op):
T
typhoonzero 已提交
373
                opt_op_on_pserver.append(op)
Q
qiaolongfei 已提交
374
        # step 3.3
T
typhoonzero 已提交
375
        # Iterate through the ops, and if an op and the optimize ops
376
        # which located on current pserver are in one set, then
T
typhoonzero 已提交
377
        # append it into the sub program.
T
typhoonzero 已提交
378 379 380 381 382

        global_ops = []
        # HACK: optimization global ops only used to scale beta1 and beta2
        # replace it with dependency engine.
        for op in self.optimize_ops:
383 384
            if self._is_adam_connected_op(op):
                global_ops.append(op)
T
typhoonzero 已提交
385

386 387
        def __append_optimize_op__(op, block, grad_to_block_id, merged_var):
            if self._is_optimizer_op(op):
Q
qiaolongfei 已提交
388
                self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
389
                                         self.origin_program, merged_var)
T
typhoonzero 已提交
390
            else:
Q
Qiyang Min 已提交
391
                self._append_pserver_non_opt_ops(block, op)
392 393 394 395 396 397

        def __op_have_grad_input__(op):
            for varname in op.input_arg_names:
                if varname.find("@GRAD") >= 0:
                    return varname
            return ""
T
typhoonzero 已提交
398

Y
Yancey1989 已提交
399
        def __clone_lr_op_sub_block__(op, program, lr_block):
Q
Qiyang Min 已提交
400 401 402 403 404 405 406 407
            if not op.has_attr('sub_block'):
                return

            origin_block_desc = op.attr('sub_block')
            origin_block = self.origin_program.block(origin_block_desc.id)
            assert isinstance(origin_block, Block)
            # we put the new sub block to new block to follow the block
            # hierarchy of the original blocks
Y
Yancey1989 已提交
408
            new_sub_block = program.create_block(lr_block.idx)
Q
Qiyang Min 已提交
409 410 411 412 413 414

            # clone vars
            for var in origin_block.vars:
                new_sub_block.clone_variable(var)

            # clone ops
Y
Yancey1989 已提交
415 416
            for origin_op in origin_block.ops:
                cloned_op = self._clone_lr_op(program, new_sub_block, origin_op)
Q
Qiyang Min 已提交
417
                # clone sub_block of op
Y
Yancey1989 已提交
418
                __clone_lr_op_sub_block__(cloned_op, program, new_sub_block)
Q
Qiyang Min 已提交
419 420 421 422

            # reset the block of op
            op.set_attr('sub_block', new_sub_block)

423
        # append lr decay ops to the child block if exists
424
        lr_ops = self._get_lr_ops()
425 426
        # record optimize blocks and we can run them on pserver parallel
        optimize_blocks = []
427
        if len(lr_ops) > 0:
Q
qiaolongfei 已提交
428 429
            lr_decay_block = pserver_program.create_block(
                pserver_program.num_blocks - 1)
430
            optimize_blocks.append(lr_decay_block)
431
            for _, op in enumerate(lr_ops):
Y
Yancey1989 已提交
432
                cloned_op = self._append_pserver_non_opt_ops(lr_decay_block, op)
Q
Qiyang Min 已提交
433
                # append sub blocks to pserver_program in lr_decay_op
Y
Yancey1989 已提交
434 435
                __clone_lr_op_sub_block__(cloned_op, pserver_program,
                                          lr_decay_block)
436

T
typhoonzero 已提交
437
        # append op to the current block
Q
qiaolongfei 已提交
438
        grad_to_block_id = []
Q
qiaolongfei 已提交
439
        pre_block_idx = pserver_program.num_blocks - 1
T
typhoonzero 已提交
440
        for idx, opt_op in enumerate(opt_op_on_pserver):
441
            per_opt_block = pserver_program.create_block(pre_block_idx)
442
            optimize_blocks.append(per_opt_block)
443 444 445 446 447 448 449 450
            # append grad merging ops before clip and weight decay
            for _, op in enumerate(self.optimize_ops):
                # find the origin @GRAD var before clipping
                grad_varname_for_block = __op_have_grad_input__(op)
                if ufind.is_connected(op, opt_op) and grad_varname_for_block:
                    merged_var = self._append_pserver_grad_merge_ops(
                        per_opt_block, grad_varname_for_block, endpoint,
                        grad_to_block_id, self.origin_program)
T
typhoonzero 已提交
451 452
            for _, op in enumerate(self.optimize_ops):
                # optimizer is connected to itself
453
                if ufind.is_connected(op, opt_op) and op not in global_ops:
454 455
                    __append_optimize_op__(op, per_opt_block, grad_to_block_id,
                                           merged_var)
T
typhoonzero 已提交
456 457

        # append global ops
458
        if global_ops:
Q
qiaolongfei 已提交
459 460
            opt_state_block = pserver_program.create_block(
                pserver_program.num_blocks - 1)
461
            optimize_blocks.append(opt_state_block)
Q
qiaolongfei 已提交
462
            for glb_op in global_ops:
X
Xi Chen 已提交
463
                __append_optimize_op__(glb_op, opt_state_block,
464
                                       grad_to_block_id, None)
T
typhoonzero 已提交
465

466
        # process distributed lookup_table
Q
qiaolongfei 已提交
467
        prefetch_var_name_to_block_id = []
468 469
        if self.has_distributed_lookup_table:
            pserver_index = self.pserver_endpoints.index(endpoint)
470
            table_opt_block = self._create_table_optimize_block(
471
                pserver_index, pserver_program, pre_block_idx, grad_to_block_id)
Q
qiaolongfei 已提交
472
            prefetch_var_name_to_block_id = self._create_prefetch_block(
473
                pserver_index, pserver_program, table_opt_block)
T
tangwei12 已提交
474 475
            checkpoint_block_id = self._create_checkpoint_save_block(
                pserver_program, table_opt_block.idx)
476 477 478 479

        # NOTE: if has_distributed_lookup_table is False, then prefetch_block will
        # not be executed, so it's safe to use optimize_block to hold the place
        if self.has_distributed_lookup_table:
Q
qiaolongfei 已提交
480
            assert len(prefetch_var_name_to_block_id) > 0
481
        else:
Q
qiaolongfei 已提交
482
            assert len(prefetch_var_name_to_block_id) == 0
483

484
        attrs = {
485
            "optimize_blocks": optimize_blocks,
486 487 488
            "endpoint": endpoint,
            "Fanin": self.trainer_num,
            "sync_mode": self.sync_mode,
Y
Yancey1989 已提交
489
            "grad_to_block_id": grad_to_block_id,
490 491 492 493
        }
        if len(prefetch_var_name_to_block_id) > 0:
            attrs['prefetch_var_name_to_block_id'] \
                = prefetch_var_name_to_block_id
T
tangwei12 已提交
494
            attrs['checkpint_block_id'] = checkpoint_block_id
495

T
typhoonzero 已提交
496 497 498 499 500
        # step5 append the listen_and_serv op
        pserver_program.global_block().append_op(
            type="listen_and_serv",
            inputs={'X': recv_inputs},
            outputs={},
501
            attrs=attrs)
502

T
typhoonzero 已提交
503 504 505 506 507 508 509 510
        pserver_program.sync_with_cpp()
        return pserver_program

    def get_startup_program(self, endpoint, pserver_program):
        """
        Get startup program for current parameter server.
        Modify operator input variables if there are variables that
        were split to several blocks.
Y
yi.wu 已提交
511 512 513 514 515 516 517 518

        Args:
            endpoint (str): current pserver endpoint.
            pserver_program (Program): call get_pserver_program first and
                pass the result here.
        
        Returns:
            Program: parameter server side startup program.
T
typhoonzero 已提交
519 520
        """
        s_prog = Program()
T
typhoonzero 已提交
521
        orig_s_prog = default_startup_program()
T
typhoonzero 已提交
522 523 524 525 526 527 528 529 530 531 532 533 534
        params = self.param_grad_ep_mapping[endpoint]["params"]

        def _get_splited_name_and_shape(varname):
            for idx, splited_param in enumerate(params):
                pname = splited_param.name
                if same_or_split_var(pname, varname) and varname != pname:
                    return pname, splited_param.shape
            return "", []

        # 1. create vars in pserver program to startup program
        pserver_vars = pserver_program.global_block().vars
        created_var_map = dict()
        for _, var in pserver_vars.iteritems():
T
update  
typhoonzero 已提交
535
            tmpvar = s_prog.global_block().clone_variable(var)
T
typhoonzero 已提交
536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567
            created_var_map[var.name] = tmpvar

        # 2. rename op outputs
        for op in orig_s_prog.global_block().ops:
            new_inputs = dict()
            new_outputs = dict()
            # do not append startup op if var is not on this pserver
            op_on_pserver = False
            for key in op.output_names:
                newname, _ = _get_splited_name_and_shape(op.output(key)[0])
                if newname:
                    op_on_pserver = True
                    new_outputs[key] = created_var_map[newname]
                elif op.output(key)[0] in pserver_vars:
                    op_on_pserver = True
                    new_outputs[key] = pserver_vars[op.output(key)[0]]

            # most startup program ops have no inputs
            new_inputs = self._get_input_map_from_op(pserver_vars, op)

            if op_on_pserver:
                if op.type in [
                        "gaussian_random", "fill_constant", "uniform_random"
                ]:
                    op.attrs["shape"] = new_outputs["Out"].shape
                s_prog.global_block().append_op(
                    type=op.type,
                    inputs=new_inputs,
                    outputs=new_outputs,
                    attrs=op.attrs)
        return s_prog

568 569
    # ====================== private transpiler functions =====================

Y
yi.wu 已提交
570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692
    def _has_distributed_lookup_table(self):
        # process lookup_table_op
        # 1. check all lookup_table_op is distributed
        # 2. check all lookup_table_op share the same table.
        distributed_lookup_table_ops = []
        # support only one distributed_lookup_table now
        self.table_name = None
        for op in self.origin_program.global_block().ops:
            if op.type == LOOKUP_TABLE_TYPE:
                if op.attrs['is_distributed'] is True:
                    if self.table_name is None:
                        self.table_name = op.input("W")[0]
                    if self.table_name != op.input("W")[0]:
                        raise RuntimeError("all distributed lookup_table_ops"
                                           " should have only one table")
                    distributed_lookup_table_ops.append(op)
                else:
                    if self.table_name is not None:
                        assert op.input("W")[0] != self.table_name

        return len(distributed_lookup_table_ops) > 0

    def _update_dist_lookup_table_vars(self, param_list, grad_list,
                                       params_grads):
        # TODO(wuyi): put find a way to put dist lookup table stuff all together.
        # update self.table_param_grad and self.trainer_side_table_grad_list
        program = self.origin_program
        if self.has_distributed_lookup_table:
            param_list = [
                param for param in param_list if param.name != self.table_name
            ]
            grad_list = [
                grad for grad in grad_list
                if grad.name != grad_var_name(self.table_name)
            ]
            self.table_param_grad = [
                param_grad for param_grad in params_grads
                if param_grad[0].name == self.table_name
            ][0]
            table_grad_var = self.table_param_grad[1]
            if self.sync_mode:
                self.trainer_side_table_grad_list = [
                    program.global_block().create_var(
                        name="%s.trainer_%d.pserver_%d" %
                        (table_grad_var.name, self.trainer_id, index),
                        type=table_grad_var.type,
                        shape=table_grad_var.shape,
                        dtype=table_grad_var.dtype)
                    for index in range(len(self.pserver_endpoints))
                ]
            else:
                self.trainer_side_table_grad_list = [
                    program.global_block().create_var(
                        name="%s.pserver_%d" % (table_grad_var.name, index),
                        type=table_grad_var.type,
                        shape=table_grad_var.shape,
                        dtype=table_grad_var.dtype)
                    for index in range(len(self.pserver_endpoints))
                ]
        return param_list, grad_list

    def _init_splited_vars(self, slice_var_up):
        # update these mappings for further transpile:
        # 1. param_var_mapping: param var name -> [splited params vars]
        # 2. grad_var_mapping: grad var name -> [splited grads vars]
        # 3. grad_param_mapping: grad.blockx -> param.blockx
        # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []}

        param_list = []
        grad_list = []
        param_grad_set = set()
        for p, g in self.params_grads:
            # skip parameter marked not trainable
            if type(p) == Parameter and p.trainable == False:
                continue
            if p.name not in param_grad_set:
                param_list.append(p)
                param_grad_set.add(p.name)
            if g.name not in param_grad_set:
                grad_list.append(g)
                param_grad_set.add(g.name)

        param_list, grad_list = self._update_dist_lookup_table_vars(
            param_list, grad_list, self.params_grads)

        if slice_var_up:
            # when we slice var up into blocks, we will slice the var according to
            # pserver services' count. A pserver may have two or more listening ports.
            grad_blocks = slice_variable(grad_list, len(self.pserver_endpoints))
            param_blocks = slice_variable(param_list,
                                          len(self.pserver_endpoints))
        else:
            # when we do NOT slice var up into blocks, we will always slice params
            # grads into one block.
            grad_blocks = slice_variable(grad_list, 1)
            param_blocks = slice_variable(param_list, 1)
        assert (len(grad_blocks) == len(param_blocks))

        # origin_varname -> [splited_var]
        self.param_var_mapping = self._create_vars_from_blocklist(
            self.origin_program, param_blocks)
        self.grad_var_mapping = self._create_vars_from_blocklist(
            self.origin_program,
            grad_blocks,
            add_trainer_suffix=self.trainer_num > 1)
        self.grad_param_mapping = dict()
        for g, p in zip(grad_blocks, param_blocks):
            g_name, g_bid, _ = g.split(":")
            p_name, p_bid, _ = p.split(":")
            self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] =  \
                    self.param_var_mapping[p_name][int(p_bid)]

        # create mapping of endpoint -> split var to create pserver side program
        self.param_grad_ep_mapping = dict()
        [
            self.param_grad_ep_mapping.update({
                ep: {
                    "params": [],
                    "grads": []
                }
            }) for ep in self.pserver_endpoints
        ]

693
    # transpiler function for dis lookup_table
Q
update  
qiaolongfei 已提交
694 695
    def _replace_lookup_table_op_with_prefetch(self, program,
                                               pserver_endpoints):
696
        # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op
Q
qiaolongfei 已提交
697 698 699 700 701 702 703 704 705
        # self.all_prefetch_input_vars =
        #       [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
        #        [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
        self.all_prefetch_input_vars = []

        # self.all_prefetch_input_vars =
        #       [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
        #        [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
        self.all_prefetch_output_vars = []
706 707 708 709 710 711 712 713 714

        continue_search_lookup_table_op = True
        while continue_search_lookup_table_op:
            continue_search_lookup_table_op = False
            all_ops = program.global_block().ops
            for op in all_ops:
                if op.type == LOOKUP_TABLE_TYPE:
                    continue_search_lookup_table_op = True

715
                    lookup_table_op_index = list(all_ops).index(op)
716 717 718
                    ids_name = op.input("Ids")
                    out_name = op.output("Out")

Q
qiaolongfei 已提交
719 720 721 722 723 724 725 726 727 728 729 730 731
                    ids_var = program.global_block().vars[ids_name[0]]
                    prefetch_input_vars = self.create_splited_vars(
                        source_var=ids_var,
                        block=program.global_block(),
                        tag="_prefetch_in_")
                    self.all_prefetch_input_vars.append(prefetch_input_vars)

                    out_var = program.global_block().vars[out_name[0]]
                    prefetch_output_vars = self.create_splited_vars(
                        source_var=out_var,
                        block=program.global_block(),
                        tag="_prefetch_out_")
                    self.all_prefetch_output_vars.append(prefetch_output_vars)
732 733 734

                    # insert split_ids_op
                    program.global_block().insert_op(
735
                        index=lookup_table_op_index,
736 737 738 739 740 741 742
                        type="split_ids",
                        inputs={
                            'Ids': [
                                program.global_block().vars[varname]
                                for varname in ids_name
                            ]
                        },
Q
qiaolongfei 已提交
743
                        outputs={"Out": prefetch_input_vars})
744 745 746

                    # insert prefetch_op
                    program.global_block().insert_op(
747
                        index=lookup_table_op_index + 1,
748
                        type="prefetch",
Q
qiaolongfei 已提交
749 750
                        inputs={'X': prefetch_input_vars},
                        outputs={"Out": prefetch_output_vars},
Y
Yancey1989 已提交
751
                        attrs={
752
                            "epmap": pserver_endpoints,
Y
Yancey1989 已提交
753 754
                            RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                        })
755 756 757

                    # insert concat_op
                    program.global_block().insert_op(
758 759 760 761 762 763 764
                        index=lookup_table_op_index + 2,
                        type="merge_ids",
                        inputs={
                            'Ids': [
                                program.global_block().vars[varname]
                                for varname in ids_name
                            ],
765
                            'X': prefetch_output_vars
766
                        },
767 768 769 770 771
                        outputs={
                            "Out": [
                                program.global_block().vars[varname]
                                for varname in out_name
                            ]
772
                        })
773 774

                    # delete lookup_table_op
775
                    delete_ops(program.global_block(), [op])
776 777 778
                    # break for loop
                    break

Y
Yancey1989 已提交
779
    def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints):
780
        # 2. add split_ids_op and send_op to send gradient to pservers
781 782
        # there should only be one table_name
        all_ops = program.global_block().ops
T
typhoonzero 已提交
783
        table_grad_name = grad_var_name(self.table_name)
784 785 786 787 788 789 790 791 792 793
        for op in all_ops:
            if table_grad_name in op.output_arg_names:
                op_index = list(all_ops).index(op)
                # insert split_ids_op
                program.global_block().insert_op(
                    index=op_index + 1,
                    type="split_ids",
                    inputs={
                        'Ids': [program.global_block().vars[table_grad_name]]
                    },
794
                    outputs={"Out": self.trainer_side_table_grad_list})
795 796
                program.global_block().insert_op(
                    index=op_index + 2,
797
                    type="send",
798
                    inputs={'X': self.trainer_side_table_grad_list},
Y
Yancey1989 已提交
799 800
                    outputs={},
                    attrs={
801
                        "sync_mode": True,
Y
Yancey1989 已提交
802 803 804
                        "epmap": pserver_endpoints,
                        RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                    })
805 806 807 808 809 810
                break

    def _create_prefetch_block(self, pserver_index, pserver_program,
                               optimize_block):
        # STEP: create prefetch block
        table_var = pserver_program.global_block().vars[self.table_name]
Q
qiaolongfei 已提交
811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838
        prefetch_var_name_to_block_id = []
        for index in range(len(self.all_prefetch_input_vars)):
            prefetch_block = pserver_program.create_block(optimize_block.idx)
            trainer_ids = self.all_prefetch_input_vars[index][pserver_index]
            pserver_ids = pserver_program.global_block().create_var(
                name=trainer_ids.name,
                type=trainer_ids.type,
                shape=trainer_ids.shape,
                dtype=trainer_ids.dtype)
            trainer_out = self.all_prefetch_output_vars[index][pserver_index]
            pserver_out = pserver_program.global_block().create_var(
                name=trainer_out.name,
                type=trainer_out.type,
                shape=trainer_out.shape,
                dtype=trainer_out.dtype)
            prefetch_block.append_op(
                type="lookup_sparse_table",
                inputs={'Ids': pserver_ids,
                        "W": table_var},
                outputs={"Out": pserver_out},
                attrs={
                    "is_sparse": True,  # has no effect on lookup_table op
                    "is_distributed": True,
                    "padding_idx": -1
                })
            prefetch_var_name_to_block_id.append(trainer_ids.name + ":" + str(
                prefetch_block.idx))
        return prefetch_var_name_to_block_id
839 840

    def _create_table_optimize_block(self, pserver_index, pserver_program,
841
                                     pre_block_idx, grad_to_block_id):
842 843
        # STEP: create table optimize block
        # create table param and grad var in pserver program
Y
Yancey1989 已提交
844 845 846 847 848 849 850 851
        origin_param_var = self.origin_program.global_block().vars[
            self.table_name]
        param_var = pserver_program.global_block().create_var(
            name=origin_param_var.name,
            shape=origin_param_var.shape,
            dtype=origin_param_var.dtype,
            type=core.VarDesc.VarType.SELECTED_ROWS,
            persistable=True)
852 853 854
        # parameter must be selected rows
        param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS)
        grad_var = pserver_program.global_block().clone_variable(
T
typhoonzero 已提交
855
            self.origin_program.global_block().vars[grad_var_name(
856
                self.table_name)])
857 858 859 860 861 862

        # create table optimize block in pserver program
        table_opt_op = [
            op for op in self.optimize_ops
            if op.input("Param")[0] == self.table_name
        ][0]
Q
qiaolongfei 已提交
863
        table_opt_block = pserver_program.create_block(pre_block_idx)
864 865 866
        # only support sgd now
        assert table_opt_op.type == "sgd"

867 868 869
        if self.sync_mode:
            # create grad vars in pserver program
            table_grad_var = self.table_param_grad[1]
870
            pserver_side_table_grad_list = [
871 872 873 874 875 876 877 878 879
                pserver_program.global_block().create_var(
                    name="%s.trainer_%d.pserver_%d" %
                    (table_grad_var.name, index, pserver_index),
                    type=table_grad_var.type,
                    shape=table_grad_var.shape,
                    dtype=table_grad_var.dtype)
                for index in range(self.trainer_num)
            ]

880
            # append sum op for pserver_side_table_grad_list
881 882
            table_opt_block.append_op(
                type="sum",
883
                inputs={"X": pserver_side_table_grad_list},
884 885
                outputs={"Out": [grad_var]},
                attrs={"use_mkldnn": False})
886 887
        else:
            # in async_mode, for table gradient, it also need to be splited to each parameter server
888
            origin_grad_name = grad_var.name
889 890
            splited_grad_name = self.trainer_side_table_grad_list[
                pserver_index].name
891 892
            if not splited_grad_name.startswith(origin_grad_name):
                raise ValueError("origin_grad_var: " + splited_grad_name +
893
                                 " grad_var:" + grad_var.name)
894 895
            grad_var = pserver_program.global_block().rename_var(
                origin_grad_name, splited_grad_name)
896 897 898 899 900 901 902 903 904 905 906 907 908 909 910

        lr_var = pserver_program.global_block().vars[table_opt_op.input(
            "LearningRate")[0]]
        inputs = {
            "Param": [param_var],
            "Grad": [grad_var],
            "LearningRate": [lr_var]
        }
        outputs = {"ParamOut": [param_var]}
        table_opt_block.append_op(
            type=table_opt_op.type,
            inputs=inputs,
            outputs=outputs,
            attrs=table_opt_op.attrs)

911 912 913
        # add table parameter gradient and it's block id to grad_to_block_id
        grad_to_block_id.append(grad_var.name + ":" + str(table_opt_block.idx))

914 915
        return table_opt_block

T
tangwei12 已提交
916 917 918 919 920 921
    def _create_checkpoint_save_block(self, pserver_program, pre_block_idx):
        """
        create a new block to handle save checkpoint.
        """
        import os

T
tangwei12 已提交
922
        pserver_program.global_block().create_var(
T
tangwei12 已提交
923
            name="lookup_table_path",
T
tangwei12 已提交
924 925
            persistable=True,
            type=core.VarDesc.VarType.RAW)
T
tangwei12 已提交
926

T
tangwei12 已提交
927 928 929 930 931
        checkpoint_save_block = pserver_program.create_block(pre_block_idx)
        checkpoint_save_block.append_op(
            type='save',
            inputs={'X': [self.table_name]},
            outputs={},
T
tangwei12 已提交
932 933 934 935
            attrs={
                'file_path':
                "this 'file_path' do not be used in save lookup table variable"
            })
T
tangwei12 已提交
936 937 938

        return checkpoint_save_block.idx

T
typhoonzero 已提交
939 940 941 942 943
    def _create_vars_from_blocklist(self,
                                    program,
                                    block_list,
                                    add_trainer_suffix=False):
        """
944
        Create vars for each split.
T
typhoonzero 已提交
945 946
        NOTE: only grads need to be named for different trainers, use
              add_trainer_suffix to rename the grad vars.
947 948 949 950
        Args:
            program (ProgramDesc): ProgramDesc which gradients blong.
            block_list (list[(varname, block_id, block_size)]): List of gradient blocks.
            add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True.
951 952
        Returns:
            var_mapping (dict(varname->[new_varname_variable])):A dict mapping
953
                from original var name to each var split.
T
typhoonzero 已提交
954
        """
955 956

        # varname->[(block_id, current_block_size)]
T
typhoonzero 已提交
957
        block_map = dict()
958

T
typhoonzero 已提交
959
        var_mapping = dict()
T
typhoonzero 已提交
960 961 962 963 964
        for block_str in block_list:
            varname, offset, size = block_str.split(":")
            if not block_map.has_key(varname):
                block_map[varname] = []
            block_map[varname].append((long(offset), long(size)))
Y
yi.wu 已提交
965 966 967
        # Do not remove this important debug message:
        print("block map: %s" % block_map)

T
typhoonzero 已提交
968
        for varname, splited in block_map.iteritems():
T
typhoonzero 已提交
969
            orig_var = program.global_block().var(varname)
T
typhoonzero 已提交
970
            if len(splited) == 1:
971
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
972 973 974 975 976 977 978 979
                    new_var_name = "%s.trainer_%d" % \
                        (orig_var.name, self.trainer_id)
                    program.global_block().rename_var(varname, new_var_name)
                    var_mapping[varname] = \
                        [program.global_block().var(new_var_name)]
                else:
                    var_mapping[varname] = \
                        [program.global_block().var(orig_var.name)]
T
typhoonzero 已提交
980
                continue
T
typhoonzero 已提交
981 982

            var_mapping[varname] = []
T
typhoonzero 已提交
983 984 985 986
            orig_shape = orig_var.shape
            orig_dim1_flatten = 1
            if len(orig_shape) >= 2:
                orig_dim1_flatten = reduce(lambda x, y: x * y, orig_shape[1:])
T
typhoonzero 已提交
987

T
typhoonzero 已提交
988
            for i, block in enumerate(splited):
T
typhoonzero 已提交
989
                size = block[1]
T
typhoonzero 已提交
990 991 992 993
                rows = size / orig_dim1_flatten
                splited_shape = [rows]
                if len(orig_shape) >= 2:
                    splited_shape.extend(orig_shape[1:])
T
typhoonzero 已提交
994
                new_var_name = ""
995
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
996 997 998 999 1000
                    new_var_name = "%s.block%d.trainer_%d" % \
                        (varname, i, self.trainer_id)
                else:
                    new_var_name = "%s.block%d" % \
                        (varname, i)
T
typhoonzero 已提交
1001
                var = program.global_block().create_var(
T
typhoonzero 已提交
1002 1003
                    name=new_var_name,
                    persistable=False,
T
typhoonzero 已提交
1004
                    dtype=orig_var.dtype,
1005
                    type=orig_var.type,
T
typhoonzero 已提交
1006
                    shape=splited_shape)  # flattend splited var
T
typhoonzero 已提交
1007
                var_mapping[varname].append(var)
T
typhoonzero 已提交
1008
            program.global_block().sync_with_cpp()
T
typhoonzero 已提交
1009
        return var_mapping
T
done  
typhoonzero 已提交
1010

1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021
    def create_splited_vars(self, source_var, block, tag):
        return [
            block.create_var(
                name=str(source_var.name + tag + str(index)),
                type=source_var.type,
                shape=source_var.shape,
                dtype=source_var.dtype)
            for index in range(len(self.pserver_endpoints))
        ]

    def _clone_var(self, block, var, persistable=True):
T
done  
typhoonzero 已提交
1022 1023 1024 1025 1026 1027 1028
        assert isinstance(var, Variable)
        return block.create_var(
            name=var.name,
            shape=var.shape,
            dtype=var.dtype,
            type=var.type,
            lod_level=var.lod_level,
1029
            persistable=persistable)
T
done  
typhoonzero 已提交
1030

Y
Yancey1989 已提交
1031
    def _insert_split_op(self, program, orig_var, index, splited_vars):
Y
update  
Yancey1989 已提交
1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055
        if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS:
            height_sections = []
            for v in splited_vars:
                height_sections.append(v.shape[0])
            program.global_block().insert_op(
                index=index + 1,
                type="split_selected_rows",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
                attrs={"height_sections": height_sections})
        elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR:
            sections = []
            for v in splited_vars:
                sections.append(v.shape[0])
            program.global_block().insert_op(
                index=index + 1,
                type="split_byref",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
                attrs={"sections": sections}  # assume split evenly
            )
        else:
            AssertionError("Variable type should be in set "
                           "[LOD_TENSOR, SELECTED_ROWS]")
T
done  
typhoonzero 已提交
1056

T
typhoonzero 已提交
1057 1058 1059 1060
    def _get_optimizer_input_shape(self, op_type, varkey, orig_shape,
                                   param_shape):
        """
        Returns the shape for optimizer inputs that need to be reshaped when
1061
        Param and Grad is split to multiple servers.
T
typhoonzero 已提交
1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083
        """
        # HACK(typhoonzero): Should use functions of corresponding optimizer in
        # optimizer.py to get the shape, do not  bind this in the transpiler.
        if op_type == "adam":
            if varkey in ["Moment1", "Moment2"]:
                return param_shape
        elif op_type == "adagrad":
            if varkey == "Moment":
                return param_shape
        elif op_type == "adamax":
            if varkey in ["Moment", "InfNorm"]:
                return param_shape
        elif op_type == "momentum":
            if varkey == "Velocity":
                return param_shape
        elif op_type == "":
            if varkey == "Moment":
                return param_shape
        elif op_type == "sgd":
            pass
        return orig_shape

1084 1085
    def _get_varname_parts(self, varname):
        # returns origin, blockid, trainerid
T
typhoonzero 已提交
1086
        orig_var_name = ""
1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
        trainer_part = ""
        block_part = ""
        trainer_idx = varname.find(".trainer_")
        if trainer_idx >= 0:
            trainer_part = varname[trainer_idx + 1:]
        else:
            trainer_idx = len(varname)
        block_index = varname.find(".block")
        if block_index >= 0:
            block_part = varname[block_index + 1:trainer_idx]
T
typhoonzero 已提交
1097
        else:
1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124
            block_index = len(varname)
        orig_var_name = varname[0:min(block_index, trainer_idx)]
        return orig_var_name, block_part, trainer_part

    def _orig_varname(self, varname):
        orig, _, _ = self._get_varname_parts(varname)
        return orig

    def _append_pserver_grad_merge_ops(self, optimize_block,
                                       grad_varname_for_block, endpoint,
                                       grad_to_block_id, origin_program):
        program = optimize_block.program
        pserver_block = program.global_block()
        grad_block = None
        for g in self.param_grad_ep_mapping[endpoint]["grads"]:
            if self._orig_varname(g.name) == \
                    self._orig_varname(grad_varname_for_block):
                grad_block = g
                break
        if not grad_block:
            # do not append this op if current endpoint
            # is not dealing with this grad block
            return
        orig_varname, block_name, trainer_name = self._get_varname_parts(
            grad_block.name)
        if block_name:
            merged_var_name = '.'.join([orig_varname, block_name])
T
typhoonzero 已提交
1125
        else:
1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139
            merged_var_name = orig_varname
        merged_var = \
            pserver_block.vars[merged_var_name]
        grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx))
        if self.sync_mode and self.trainer_num > 1:
            vars2merge = []
            for i in xrange(self.trainer_num):
                per_trainer_name = "%s.trainer_%d" % \
                (merged_var_name, i)
                vars2merge.append(pserver_block.vars[per_trainer_name])

            optimize_block.append_op(
                type="sum",
                inputs={"X": vars2merge},
1140 1141
                outputs={"Out": merged_var},
                attrs={"use_mkldnn": False})
1142 1143 1144 1145 1146 1147 1148 1149
            # TODO(panyx0718): What if it's SELECTED_ROWS.
            if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS:
                optimize_block.append_op(
                    type="scale",
                    inputs={"X": merged_var},
                    outputs={"Out": merged_var},
                    attrs={"scale": 1.0 / float(self.trainer_num)})
        return merged_var
T
typhoonzero 已提交
1150

1151
    def _append_pserver_ops(self, optimize_block, opt_op, endpoint,
1152
                            grad_to_block_id, origin_program, merged_var):
1153
        program = optimize_block.program
T
typhoonzero 已提交
1154
        pserver_block = program.global_block()
T
typhoonzero 已提交
1155
        new_inputs = dict()
T
typhoonzero 已提交
1156 1157
        # update param/grad shape first, then other inputs like
        # moment can use the updated shape
T
typhoonzero 已提交
1158
        for key in opt_op.input_names:
T
typhoonzero 已提交
1159 1160 1161 1162 1163 1164
            if key == "Grad":
                new_inputs[key] = merged_var
            elif key == "Param":
                # param is already created on global program
                param_block = None
                for p in self.param_grad_ep_mapping[endpoint]["params"]:
T
typhoonzero 已提交
1165
                    if same_or_split_var(p.name, opt_op.input(key)[0]):
T
typhoonzero 已提交
1166 1167 1168 1169
                        param_block = p
                        break
                if not param_block:
                    return
T
typhoonzero 已提交
1170
                tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1171
                    name=param_block.name,
T
typhoonzero 已提交
1172
                    persistable=True,
T
typhoonzero 已提交
1173 1174 1175
                    dtype=param_block.dtype,
                    shape=param_block.shape)
                new_inputs[key] = tmpvar
1176
            elif key == "LearningRate":
1177
                # learning rate variable has already be created by non-optimize op,
1178
                # don't create it once again.
1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189
                lr_varname = opt_op.input(key)[0]
                if pserver_block.vars.has_key(lr_varname):
                    new_inputs[key] = pserver_block.vars[opt_op.input(key)[0]]
                else:
                    origin_var = origin_program.global_block().vars[lr_varname]
                    tmpvar = pserver_block.create_var(
                        name=origin_var.name,
                        persistable=origin_var.persistable,
                        dtype=origin_var.dtype,
                        shape=origin_var.shape)
                    new_inputs[key] = tmpvar
T
typhoonzero 已提交
1190

T
typhoonzero 已提交
1191
        for key in opt_op.input_names:
1192 1193
            new_shape = None
            if key in ["Param", "Grad", "LearningRate"]:
T
typhoonzero 已提交
1194
                continue
1195
            var = self.origin_program.global_block().vars[opt_op.input(key)[0]]
T
typhoonzero 已提交
1196 1197 1198 1199
            # update accumulator variable shape
            param_shape = new_inputs["Param"].shape
            new_shape = self._get_optimizer_input_shape(opt_op.type, key,
                                                        var.shape, param_shape)
T
typhoonzero 已提交
1200
            tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1201 1202 1203 1204 1205
                name=var.name,
                persistable=var.persistable,
                dtype=var.dtype,
                shape=new_shape)
            new_inputs[key] = tmpvar
T
typhoonzero 已提交
1206

1207
        # change output's ParamOut variable
1208 1209
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1210
        outputs["ParamOut"] = new_inputs["Param"]
T
typhoonzero 已提交
1211

1212
        optimize_block.append_op(
T
typhoonzero 已提交
1213 1214
            type=opt_op.type,
            inputs=new_inputs,
T
typhoonzero 已提交
1215
            outputs=outputs,
T
typhoonzero 已提交
1216 1217
            attrs=opt_op.attrs)

1218 1219 1220 1221 1222 1223 1224 1225 1226
    def _is_splited_grad_var(self, var, var_dict):
        grad_block = None
        for _, g in var_dict.iteritems():
            if self._orig_varname(g.name) == self._orig_varname(var.name):
                if g.name.find(".trainer_") == -1:
                    grad_block = g
                    break
        return grad_block

Q
Qiyang Min 已提交
1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245
    def _clone_lr_op(self, program, block, op):
        inputs = self._get_input_map_from_op(
            self.origin_program.global_block().vars, op)
        for key, varlist in inputs.iteritems():
            if not isinstance(varlist, list):
                varlist = [varlist]
            for var in varlist:
                if var not in program.global_block().vars:
                    block.clone_variable(var)

        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, op)
        for key, varlist in outputs.iteritems():
            if not isinstance(varlist, list):
                varlist = [varlist]
            for var in varlist:
                if var not in program.global_block().vars:
                    block.clone_variable(var)

Y
Yancey1989 已提交
1246
        return block.append_op(
Q
Qiyang Min 已提交
1247 1248 1249
            type=op.type, inputs=inputs, outputs=outputs, attrs=op.attrs)

    def _append_pserver_non_opt_ops(self, optimize_block, opt_op):
1250
        program = optimize_block.program
1251
        # Append the ops for parameters that do not need to be optimized/updated
1252 1253
        inputs = self._get_input_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1254
        for key, varlist in inputs.iteritems():
1255 1256
            if not isinstance(varlist, list):
                varlist = [varlist]
T
typhoonzero 已提交
1257
            for var in varlist:
1258 1259 1260 1261 1262 1263 1264
                # for ops like clipping and weight decay, get the splited var
                # for inputs/outputs
                grad_block = self._is_splited_grad_var(
                    var, program.global_block().vars)
                if grad_block:
                    inputs[key] = grad_block
                elif not program.global_block().vars.has_key(var.name):
1265
                    program.global_block().create_var(
T
typhoonzero 已提交
1266 1267 1268 1269 1270
                        name=var.name,
                        persistable=var.persistable,
                        dtype=var.dtype,
                        shape=var.shape)

1271 1272
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1273
        for key, varlist in outputs.iteritems():
1274 1275 1276
            if not isinstance(varlist, list):
                varlist = [varlist]
            for var in varlist:
1277 1278 1279 1280 1281 1282
                grad_block = self._is_splited_grad_var(
                    var, program.global_block().vars)
                if grad_block:
                    outputs[key] = grad_block
                elif not program.global_block().vars.has_key(var.name):
                    program.global_block().clone_variable(var)
1283

Y
Yancey1989 已提交
1284
        return optimize_block.append_op(
T
typhoonzero 已提交
1285
            type=opt_op.type,
T
typhoonzero 已提交
1286 1287
            inputs=inputs,
            outputs=outputs,
T
typhoonzero 已提交
1288 1289
            attrs=opt_op.attrs)

1290 1291 1292 1293
    def _is_op_connected(self, op1, op2):
        # If one op's input is another op's output or
        # one op's output is another op's input, we say
        # the two operator is connected.
T
typhoonzero 已提交
1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306
        def _append_inname_remove_beta(varname_list):
            op_input_names = []
            for in_name in varname_list:
                # HACK: remove beta1 and beta2 to avoid let all
                # ops connected.
                if in_name.startswith("beta2_pow_acc") or \
                    in_name.startswith("beta1_pow_acc"):
                    continue
                else:
                    op_input_names.append(in_name)
            return op_input_names

        op1_input_names = _append_inname_remove_beta(op1.desc.input_arg_names())
T
typhoonzero 已提交
1307 1308
        op1_output_names = op1.desc.output_arg_names()

T
typhoonzero 已提交
1309
        op2_input_names = _append_inname_remove_beta(op2.desc.input_arg_names())
T
typhoonzero 已提交
1310
        op2_output_names = op2.desc.output_arg_names()
1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327

        if set(op1_output_names) & set(op2_input_names) or \
           set(op1_input_names) & set(op2_output_names):
            return True
        return False

    def _create_ufind(self, optimize_ops):
        # Create a unit find data struct by optimize ops
        ufind = UnionFind(optimize_ops)
        for i in xrange(len(optimize_ops)):
            for j in xrange(i, len(optimize_ops)):
                op1 = optimize_ops[i]
                op2 = optimize_ops[j]
                if self._is_op_connected(op1, op2):
                    ufind.union(op1, op2)
        return ufind

1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338
    def _is_opt_role_op(self, op):
        # NOTE: depend on oprole to find out whether this op is for
        # optimize
        op_maker = core.op_proto_and_checker_maker
        optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize
        if op_maker.kOpRoleAttrName() in op.attrs and \
            int(op.attrs[op_maker.kOpRoleAttrName()]) == int(optimize_role):
            return True
        return False

    def _is_optimizer_op(self, op):
T
typhoonzero 已提交
1339 1340
        if "Param" in op.input_names and \
            "LearningRate" in op.input_names:
1341 1342 1343 1344 1345 1346 1347
            return True
        return False

    def _is_opt_op_on_pserver(self, endpoint, op):
        param_names = [
            p.name for p in self.param_grad_ep_mapping[endpoint]["params"]
        ]
T
typhoonzero 已提交
1348
        if op.input("Param")[0] in param_names:
1349 1350 1351
            return True
        else:
            for n in param_names:
T
typhoonzero 已提交
1352
                param = op.input("Param")[0]
T
typhoonzero 已提交
1353
                if same_or_split_var(n, param) and n != param:
1354 1355 1356
                    return True
            return False

T
typhoonzero 已提交
1357
    def _get_input_map_from_op(self, varmap, op):
1358
        """Returns a dict from op input name to the vars in varmap."""
T
typhoonzero 已提交
1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370
        iomap = dict()
        for key in op.input_names:
            vars = []
            for varname in op.input(key):
                vars.append(varmap[varname])
            if len(vars) == 1:
                iomap[key] = vars[0]
            else:
                iomap[key] = vars
        return iomap

    def _get_output_map_from_op(self, varmap, op):
1371
        """Returns a dict from op output name to the vars in varmap."""
T
typhoonzero 已提交
1372 1373 1374 1375 1376 1377 1378 1379 1380 1381
        iomap = dict()
        for key in op.output_names:
            vars = []
            for varname in op.output(key):
                vars.append(varmap[varname])
            if len(vars) == 1:
                iomap[key] = vars[0]
            else:
                iomap[key] = vars
        return iomap
1382 1383 1384 1385 1386 1387

    def _get_lr_ops(self):
        lr_ops = []
        # find learning rate variables by optimize op
        lr_vars = set()
        for op in self.optimize_ops:
1388
            if self._is_optimizer_op(op):
1389 1390 1391 1392
                lr_vars.add(op.input("LearningRate")[0])

        find_ops = []
        # find ops which output is lr var
1393
        block = self.origin_program.global_block()
1394 1395 1396 1397 1398
        for op in block.ops:
            if set(op.output_arg_names) & lr_vars:
                find_ops.append(op)
        # make a union find struct by the ops in default_main_program
        ufind = UnionFind(block.ops)
1399

1400 1401 1402 1403 1404
        for op1 in block.ops:
            for op2 in block.ops:
                # NOTE: we need to skip all optimize ops, since it is connected
                # with forward/backward ops and lr ops, we only need the lr ops.
                if op1 != op2 and self._is_op_connected(op1, op2) and \
1405
                    not self._is_optimizer_op(op1) and not self._is_optimizer_op(op2):
1406 1407 1408 1409 1410 1411
                    ufind.union(op1, op2)
        # find all ops which is related with lr var
        for op1 in block.ops:
            for op2 in find_ops:
                if ufind.is_connected(op1, op2):
                    lr_ops.append(op1)
1412 1413
                    # we only need to append op for once
                    break
1414
        return lr_ops
Y
Yancey1989 已提交
1415 1416

    def _get_optimize_pass(self):
1417 1418 1419 1420 1421 1422
        """
        Get optimizer operators, paramters and gradients from origin_program
        Returns:
            opt_ops (list): optimize operators.
            params_grads (dict): paramter->gradient.
        """
Y
Yancey1989 已提交
1423 1424 1425
        block = self.origin_program.global_block()
        opt_ops = []
        params_grads = []
1426
        origin_var_dict = self.origin_program.global_block().vars
Y
Yancey1989 已提交
1427
        for op in block.ops:
1428
            if self._is_opt_role_op(op):
Y
Yancey1989 已提交
1429
                opt_ops.append(op)
1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440
                # HACK(wuyi): if we find grad vars from input of optimize
                # ops, we may get the output of clip op. Use syntax "@GRAD"
                # and op_role_var to get the pair.
                for input_name in op.input_arg_names:
                    if input_name.find("@GRAD") != -1 and \
                        op.attrs[RPC_OP_ROLE_ATTR_NAME]:
                        param_name = op.attrs[OP_ROLE_VAR_ATTR_NAME][0]
                        params_grads.append([
                            origin_var_dict[param_name],
                            origin_var_dict[input_name]
                        ])
1441 1442
            elif self._is_adam_connected_op(op):
                opt_ops.append(op)
Y
Yancey1989 已提交
1443 1444 1445
            else:
                pass
        return opt_ops, params_grads
1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457

    def _is_adam_connected_op(self, op):
        """
        A hack function to determinate whether the input operator
        is connected to optimize operator.
        """
        if op.type == "scale":
            for in_name in op.input_arg_names:
                if in_name.startswith("beta1_pow_acc") or \
                        in_name.startswith("beta2_pow_acc"):
                    return True
        return False