# 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. from __future__ import print_function import math from ps_dispatcher import RoundRobin, HashName, PSDispatcher from .. import core, framework from ..framework import Program, default_main_program, \ default_startup_program, \ Variable, Parameter, grad_var_name LOOKUP_TABLE_TYPE = "lookup_table" LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad" RPC_CLIENT_VAR_NAME = "RPC_CLIENT_VAR" class VarBlock: def __init__(self, varname, offset, size): self.varname = varname # NOTE: real offset is offset * size self.offset = offset self.size = size def __str__(self): return "%s:%d:%d" % (self.varname, self.offset, self.size) class UnionFind(object): """ Union-find data structure. Union-find is a data structure that keeps track of a set of elements partitioned into a number of disjoint (non-overlapping) subsets. Reference: https://en.wikipedia.org/wiki/Disjoint-set_data_structure Args: elements(list): The initialize element list. """ def __init__(self, elementes=None): self._parents = [] # index -> parent index self._index = {} # element -> index self._curr_idx = 0 if not elementes: elementes = [] for ele in elementes: self._parents.append(self._curr_idx) self._index.update({ele: self._curr_idx}) self._curr_idx += 1 def find(self, x): # Find the root index of given element x, # execute the path compress while findind the root index if not x in self._index: return -1 idx = self._index[x] while idx != self._parents[idx]: t = self._parents[idx] self._parents[idx] = self._parents[t] idx = t return idx def union(self, x, y): # Union two given element x_root = self.find(x) y_root = self.find(y) if x_root == y_root: return self._parents[x_root] = y_root def is_connected(self, x, y): # If two given elements have the same root index, # then they are connected. return self.find(x) == self.find(y) def same_or_split_var(p_name, var_name): return p_name == var_name or p_name.startswith(var_name + ".block") def split_dense_variable(var_list, service_count, min_block_size=8192): """ 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 minimum block size 8K elements (maybe 16bit or 32bit or 64bit). Args: var_list (list): List of variables. service_count (int): Numel of pserver services. A pserver may have two or more listening ports. min_block_size (int): Minimum splitted block size. Returns: blocks (list[(varname, block_id, current_block_size)]): A list of VarBlocks. Each VarBlock specifies a shard of the var. """ blocks = [] for var in var_list: split_count = service_count 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 if max_pserver_count < service_count: 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 # update split_count after aligning 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 def delete_ops(block, ops): try: start = list(block.ops).index(ops[0]) end = list(block.ops).index(ops[-1]) [block.remove_op(start) for _ in xrange(end - start + 1)] except Exception, e: raise e block.program.sync_with_cpp() def find_op_by_input_arg(block, arg_name): for index, op in enumerate(block.ops): if arg_name in op.input_arg_names: return index return -1 def find_op_by_output_arg(block, arg_name): for index, op in enumerate(block.ops): if arg_name in op.output_arg_names: return index return -1 class DistributeTranspiler: def transpile(self, trainer_id, program=None, pservers="127.0.0.1:6174", trainers=1, split_method=RoundRobin, sync_mode=True): """ Transpile the 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. Use different methods to split trainable variables to different parameter servers. 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. 4. append send_op to send splited variables to server and fetch params(splited blocks or origin param) from server. 5. append concat_op to merge splited blocks to update local weights. 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 :param trainer_id: one unique id for each trainer in a job. :type trainer_id: int :param program: program to transpile, default is default_main_program :type program: Program :param pservers: parameter server endpoints like "m1:6174,m2:6174" :type pservers: string :param trainers: total number of workers/trainers in the job :type trainers: int :param split_method: A function to determin how to split variables to different servers equally. :type split_method: function :type sync_mode: boolean default True :param sync_mode: if sync_mode is set True, it means that dist transpiler will transpile the program into sync_mode pserver and trainer program. """ 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 # TODO(typhoonzero): currently trainer_id is fetched from cluster system # like Kubernetes, we should port this to use etcd later when developing # fluid distributed training with fault-tolerance. self.trainer_id = trainer_id pserver_endpoints = pservers.split(",") self.pserver_endpoints = pserver_endpoints self.optimize_ops, params_grads = self._get_optimize_pass() ps_dispatcher = split_method(pserver_endpoints) # 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 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 self.has_distributed_lookup_table = len( distributed_lookup_table_ops) > 0 # step1: For large parameters and gradients, split them into smaller # blocks. param_list = [] grad_list = [] for p, g in params_grads: # skip parameter marked not trainable if type(p) == Parameter and p.trainable == False: continue param_list.append(p) grad_list.append(g) 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] self.table_grad_list = [ program.global_block().create_var( name="%s.trainer_%d.pserver_%d" % (table_grad_var.name, 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)) ] grad_blocks = split_dense_variable(grad_list, len(pserver_endpoints)) param_blocks = split_dense_variable(param_list, len(pserver_endpoints)) assert (len(grad_blocks) == len(param_blocks)) # step2: Create new vars for the parameters and gradients blocks and # add ops to do the split. param_var_mapping = self._create_vars_from_blocklist(program, param_blocks) grad_var_mapping = self._create_vars_from_blocklist( program, grad_blocks, add_trainer_suffix=self.trainer_num > 1) 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(":") grad_param_mapping[grad_var_mapping[g_name][int(g_bid)]] = \ param_var_mapping[p_name][int(p_bid)] rpc_client_var = program.global_block().create_var( name=RPC_CLIENT_VAR_NAME, persistable=True, type=core.VarDesc.VarType.RAW) # step 3: transpile trainer side program, insert recv op and send op. # 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 ] # step 3.1: insert send op to send gradient vars to parameter servers ps_dispatcher.reset() send_vars = [] for orig_varname, splited_vars in grad_var_mapping.items(): eplist = ps_dispatcher.dispatch(splited_vars) 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) index += 1 else: AssertionError("Can not insert the send op by original " "variable name :", orig_varname) program.global_block().insert_op( index=index + 1, type="send_vars", inputs={"X": splited_vars}, outputs={"RPCClient": rpc_client_var}, attrs={"epmap": eplist}) for _, var in enumerate(splited_vars): send_vars.append(var) if self.sync_mode: program.global_block().append_op( type="send_barrier", inputs={}, outputs={"RPCClient": rpc_client_var}, attrs={"endpoints": pserver_endpoints}) # step 3.2: insert recv op to receive parameters from parameter server recv_vars = [] for _, var in enumerate(send_vars): recv_vars.append(grad_param_mapping[var]) ps_dispatcher.reset() eplist = ps_dispatcher.dispatch(recv_vars) program.global_block().append_op( type="recv", inputs={}, outputs={"Out": recv_vars, "RPCClient": rpc_client_var}, attrs={"epmap": eplist}) program.global_block().append_op( type="fetch_barrier", inputs={}, outputs={"RPCClient": rpc_client_var}, attrs={"endpoints": pserver_endpoints}) for i, ep in enumerate(eplist): self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i]) self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i]) # TODO(Yancey1989): check dist lookup table if self.has_distributed_lookup_table: self._replace_lookup_table_op_with_prefetch(program, rpc_client_var, eplist) self._split_table_grad_and_add_send_vars(program, rpc_client_var, pserver_endpoints) def get_trainer_program(self): # remove optimize ops and add a send op to main_program delete_ops(self.origin_program.global_block(), self.optimize_ops) # FIXME(typhoonzero): serialize once will fix error occurs when clone. self.origin_program.__str__() return self.origin_program def get_pserver_program(self, endpoint): """ Get pserver side program using the endpoint. 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. """ # step1 pserver_program = Program() # step2: Create vars to receive vars at parameter servers. 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 suff_idx = v.name.find(".trainer_") if suff_idx >= 0: orig_var_name = v.name[:suff_idx] else: orig_var_name = v.name # 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) if self.sync_mode and self.trainer_num > 1: for trainer_id in xrange(self.trainer_num): 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) # step 3 # Create a union-find data structure from optimize ops, # If two ops are connected, we could add these two ops # into one set. ufind = self._create_ufind(self.optimize_ops) # step 3.2 # Iterate through the ops and append optimize op which # located on current pserver opt_op_on_pserver = [] for _, op in enumerate(self.optimize_ops): if self._is_opt_op(op) and self._is_opt_op_on_pserver(endpoint, op): opt_op_on_pserver.append(op) # step 3.3 # Iterate through the ops, and if an op and the optimize ops # which located on current pserver are in one set, then # append it into the sub program. # We try to put optimization program run parallelly, assume # optimization program always looks like: # # prevop -> prevop -> opt op -> following op -> following op; -> # prevop -> prevop -> opt op -> following op -> following op; -> # global op -> global op # # we put operators that can run parallelly to many program blocks. # in above example, we seperate ops by the ";". Global ops must run # after all the optimize ops finished. global_ops = [] # HACK: optimization global ops only used to scale beta1 and beta2 # replace it with dependency engine. for op in self.optimize_ops: if self._is_adam_connected_op(op): global_ops.append(op) def __append_optimize_op__(op, block, grad_to_block_id): if self._is_opt_op(op): self._append_pserver_ops(block, op, endpoint, grad_to_block_id, self.origin_program) else: self._append_pserver_non_opt_ops(block, op) # append lr decay ops to the child block if exists lr_ops = self._get_lr_ops() if len(lr_ops) > 0: lr_decay_block = pserver_program.create_block( pserver_program.num_blocks - 1) for _, op in enumerate(lr_ops): self._append_pserver_non_opt_ops(lr_decay_block, op) # append op to the current block grad_to_block_id = [] pre_block_idx = pserver_program.num_blocks - 1 for idx, opt_op in enumerate(opt_op_on_pserver): per_opt_block = pserver_program.create_block(pre_block_idx) for _, op in enumerate(self.optimize_ops): # optimizer is connected to itself if ufind.is_connected(op, opt_op) and op not in global_ops: __append_optimize_op__(op, per_opt_block, grad_to_block_id) # append global ops if global_ops: opt_state_block = pserver_program.create_block( pserver_program.num_blocks - 1) for glb_op in global_ops: __append_optimize_op__(glb_op, opt_state_block, grad_to_block_id) # NOT USED: single block version: # # for _, op in enumerate(self.optimize_ops): # for _, opt_op in enumerate(opt_op_on_pserver): # if ufind.is_connected(op, opt_op): # __append_optimize_op__(glb_op, optimize_block) # break # process distributed lookup_table prefetch_block = None if self.has_distributed_lookup_table: pserver_index = self.pserver_endpoints.index(endpoint) table_opt_block = self._create_table_optimize_block( pserver_index, pserver_program, pre_block_idx) prefetch_block = self._create_prefetch_block( pserver_index, pserver_program, table_opt_block) # 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: assert prefetch_block is not None else: assert prefetch_block is None prefetch_block = pserver_program.global_block() # step5 append the listen_and_serv op pserver_program.global_block().append_op( type="listen_and_serv", inputs={'X': recv_inputs}, outputs={}, attrs={ "OptimizeBlock": pserver_program.block(1), "endpoint": endpoint, "Fanin": self.trainer_num, "PrefetchBlock": prefetch_block, "sync_mode": self.sync_mode, "grad_to_block_id": grad_to_block_id }) 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. """ s_prog = Program() orig_s_prog = default_startup_program() 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(): tmpvar = s_prog.global_block().clone_variable(var) 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 # transpiler function for dis lookup_table def _replace_lookup_table_op_with_prefetch(self, program, rpc_client_var, eplist): # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op self.prefetch_input_vars = None self.prefetch_output_vars = None 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 op_index = list(all_ops).index(op) ids_name = op.input("Ids") out_name = op.output("Out") if self.prefetch_input_vars is None: ids_var = program.global_block().vars[ids_name[0]] self.prefetch_input_vars = self.create_splited_vars( source_var=ids_var, block=program.global_block(), tag="_prefetch_in_") if self.prefetch_output_vars is None: out_var = program.global_block().vars[out_name[0]] self.prefetch_output_vars = self.create_splited_vars( source_var=out_var, block=program.global_block(), tag="_prefetch_out_") # insert split_ids_op program.global_block().insert_op( index=op_index, type="split_ids", inputs={ 'Ids': [ program.global_block().vars[varname] for varname in ids_name ] }, outputs={"Out": self.prefetch_input_vars}) # insert prefetch_op program.global_block().insert_op( index=op_index + 1, type="prefetch", inputs={'X': self.prefetch_input_vars}, outputs={ "Out": self.prefetch_output_vars, "RPCClient": rpc_client_var }, attrs={"epmap": eplist}) # insert concat_op program.global_block().insert_op( index=op_index + 2, type="concat", inputs={'X': self.prefetch_output_vars}, outputs={ "Out": [ program.global_block().vars[varname] for varname in out_name ] }, attrs={"axis": 0}) # delete lookup_table_op delete_ops(program.global_block(), [op]) # break for loop break def _split_table_grad_and_add_send_vars(self, program, rpc_client_var, pserver_endpoints): # 2. add split_ids_op and send_vars_op to send gradient to pservers # there should only be one table_name all_ops = program.global_block().ops table_grad_name = grad_var_name(self.table_name) 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]] }, outputs={"Out": self.table_grad_list}) program.global_block().insert_op( index=op_index + 2, type="send_vars", inputs={'X': self.table_grad_list}, outputs={"RPCClient": rpc_client_var}, attrs={"sync_send": True, "epmap": pserver_endpoints}) 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] prefetch_block = pserver_program.create_block(optimize_block.idx) trainer_ids = self.prefetch_input_vars[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.prefetch_output_vars[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 }) return prefetch_block def _create_table_optimize_block(self, pserver_index, pserver_program, pre_block_idx): def _clone_var(block, var, persistable=True): assert isinstance(var, Variable) return block.create_var( name=var.name, shape=var.shape, dtype=var.dtype, type=var.type, persistable=persistable) # STEP: create table optimize block # create table param and grad var in pserver program 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) grad_var = _clone_var( pserver_program.global_block(), self.origin_program.global_block().vars[grad_var_name( self.table_name)], persistable=False) # 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] table_opt_block = pserver_program.create_block(pre_block_idx) # only support sgd now assert table_opt_op.type == "sgd" if self.sync_mode: # create grad vars in pserver program table_grad_var = self.table_param_grad[1] table_grad_list = [ 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) ] # append sum op for table_grad_list table_opt_block.append_op( type="sum", inputs={"X": table_grad_list}, outputs={"Out": [grad_var]}) 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) return table_opt_block # ====================== private transpiler functions ===================== def _create_vars_from_blocklist(self, program, block_list, add_trainer_suffix=False): """ Create vars for each split. NOTE: only grads need to be named for different trainers, use add_trainer_suffix to rename the grad vars. 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. Returns: var_mapping (dict(varname->[new_varname_variable])):A dict mapping from original var name to each var split. """ # varname->[(block_id, current_block_size)] block_map = dict() var_mapping = dict() 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))) for varname, splited in block_map.iteritems(): orig_var = program.global_block().var(varname) if len(splited) == 1: if self.sync_mode and add_trainer_suffix: 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)] continue var_mapping[varname] = [] 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:]) for i, block in enumerate(splited): size = block[1] rows = size / orig_dim1_flatten splited_shape = [rows] if len(orig_shape) >= 2: splited_shape.extend(orig_shape[1:]) new_var_name = "" if self.sync_mode and add_trainer_suffix: new_var_name = "%s.block%d.trainer_%d" % \ (varname, i, self.trainer_id) else: new_var_name = "%s.block%d" % \ (varname, i) var = program.global_block().create_var( name=new_var_name, persistable=False, dtype=orig_var.dtype, type=orig_var.type, shape=splited_shape) # flattend splited var var_mapping[varname].append(var) program.global_block().sync_with_cpp() return var_mapping 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): 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, persistable=persistable) def _insert_split_op(self, program, orig_var, index, splited_vars): 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]") 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 Param and Grad is split to multiple servers. """ # 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 def _orig_varname(self, varname): suff_idx = varname.find(".trainer_") orig_var_name = "" if suff_idx >= 0: orig_var_name = varname[:suff_idx] else: orig_var_name = varname return orig_var_name def _append_pserver_ops(self, optimize_block, opt_op, endpoint, grad_to_block_id, origin_program): program = optimize_block.program pserver_block = program.global_block() new_inputs = dict() # update param/grad shape first, then other inputs like # moment can use the updated shape for key in opt_op.input_names: if key == "Grad": grad_block = None for g in self.param_grad_ep_mapping[endpoint]["grads"]: if same_or_split_var( self._orig_varname(g.name), self._orig_varname(opt_op.input(key)[0])): grad_block = g break if not grad_block: # do not append this op if current endpoint # is not dealing with this grad block return merged_var = \ pserver_block.vars[self._orig_varname(grad_block.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" % \ (self._orig_varname(grad_block.name), i) vars2merge.append(pserver_block.vars[per_trainer_name]) optimize_block.append_op( type="sum", inputs={"X": vars2merge}, outputs={"Out": merged_var}) # 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)}) 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"]: if same_or_split_var(p.name, opt_op.input(key)[0]): param_block = p break if not param_block: return tmpvar = pserver_block.create_var( name=param_block.name, persistable=True, dtype=param_block.dtype, shape=param_block.shape) new_inputs[key] = tmpvar elif key == "LearningRate": # learning rate variable has already be created by non-optimize op, # don't create it once again. 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 for key in opt_op.input_names: new_shape = None if key in ["Param", "Grad", "LearningRate"]: continue var = self.origin_program.global_block().vars[opt_op.input(key)[0]] # 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) tmpvar = pserver_block.create_var( name=var.name, persistable=var.persistable, dtype=var.dtype, shape=new_shape) new_inputs[key] = tmpvar # change output's ParamOut variable outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, opt_op) outputs["ParamOut"] = new_inputs["Param"] optimize_block.append_op( type=opt_op.type, inputs=new_inputs, outputs=outputs, attrs=opt_op.attrs) def _append_pserver_non_opt_ops(self, optimize_block, opt_op): program = optimize_block.program # Append the ops for parameters that do not need to be optimized/updated inputs = self._get_input_map_from_op( self.origin_program.global_block().vars, opt_op) for varlist in inputs.itervalues(): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: if not program.global_block().vars.has_key(var.name): program.global_block().create_var( name=var.name, persistable=var.persistable, dtype=var.dtype, shape=var.shape) outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, opt_op) for varlist in outputs.itervalues(): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: program.global_block().clone_variable(var) optimize_block.append_op( type=opt_op.type, inputs=inputs, outputs=outputs, attrs=opt_op.attrs) 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. 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()) op1_output_names = op1.desc.output_arg_names() op2_input_names = _append_inname_remove_beta(op2.desc.input_arg_names()) op2_output_names = op2.desc.output_arg_names() 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 def _is_opt_op(self, op): # NOTE: It's a HACK implement. # optimize op: SGDOptimize, MomentumOptimizer, AdamOptimizer and etc... if "Param" in op.input_names and \ "LearningRate" in op.input_names: 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"] ] if op.input("Param")[0] in param_names: return True else: for n in param_names: param = op.input("Param")[0] if same_or_split_var(n, param) and n != param: return True return False def _get_input_map_from_op(self, varmap, op): """Returns a dict from op input name to the vars in varmap.""" 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): """Returns a dict from op output name to the vars in varmap.""" 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 def _get_lr_ops(self): lr_ops = [] # find learning rate variables by optimize op lr_vars = set() for op in self.optimize_ops: if self._is_opt_op(op): lr_vars.add(op.input("LearningRate")[0]) find_ops = [] # find ops which output is lr var block = self.origin_program.global_block() 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) 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 \ not self._is_opt_op(op1) and not self._is_opt_op(op2): 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) # we only need to append op for once break return lr_ops def _get_optimize_pass(self): """ Get optimizer operators, paramters and gradients from origin_program Returns: opt_ops (list): optimize operators. params_grads (dict): paramter->gradient. """ block = self.origin_program.global_block() opt_ops = [] params_grads = [] for op in block.ops: if self._is_opt_op(op): opt_ops.append(op) params_grads.append((self.origin_program.global_block().var( op.input("Param")[0]), self.origin_program.global_block().var( op.input("Grad")[0]))) elif self._is_adam_connected_op(op): opt_ops.append(op) else: pass return opt_ops, params_grads 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