# Copyright (c) 2019 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. """Fleet Utils.""" import collections import json import logging import math import numpy as np import os import sys import time import paddle.fluid as fluid from paddle.fluid.log_helper import get_logger from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler import fleet as fleet_pslib from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler import fleet as fleet_transpiler from . import hdfs from .hdfs import * from . import utils __all__ = ["FleetUtil"] _logger = get_logger( __name__, logging.INFO, fmt='%(asctime)s-%(levelname)s: %(message)s') fleet = fleet_pslib class FleetUtil(object): """ FleetUtil provides some common functions for users' convenience. Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.rank0_print("my log") """ def __init__(self, mode="pslib"): global fleet if mode == "pslib": fleet = fleet_pslib elif mode == "transpiler": fleet = fleet_transpiler else: raise ValueError( "Please choose one mode from [\"pslib\", \"transpiler\"]") def rank0_print(self, s): """ Worker of rank 0 print some log. Args: s(str): string to print Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.rank0_print("my log") """ if fleet.worker_index() != 0: return print(s) sys.stdout.flush() def rank0_info(self, s): """ Worker of rank 0 print some log info. Args: s(str): string to log Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.rank0_info("my log info") """ if fleet.worker_index() != 0: return _logger.info(s) def rank0_error(self, s): """ Worker of rank 0 print some log error. Args: s(str): string to log Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.rank0_error("my log error") """ if fleet.worker_index() != 0: return _logger.error(s) def set_zero(self, var_name, scope=fluid.global_scope(), place=fluid.CPUPlace(), param_type="int64"): """ Set tensor of a Variable to zero. Args: var_name(str): name of Variable scope(Scope): Scope object, default is fluid.global_scope() place(Place): Place object, default is fluid.CPUPlace() param_type(str): param data type, default is int64 Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.set_zero(myvar.name, myscope) """ param = scope.var(var_name).get_tensor() param_array = np.zeros(param._get_dims()).astype(param_type) param.set(param_array, place) def print_global_auc(self, scope=fluid.global_scope(), stat_pos="_generated_var_2", stat_neg="_generated_var_3", print_prefix=""): """ Print global auc of all distributed workers. Args: scope(Scope): Scope object, default is fluid.global_scope() stat_pos(str): name of auc pos bucket Variable stat_neg(str): name of auc neg bucket Variable print_prefix(str): prefix of print auc Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.print_global_auc(myscope, stat_pos=stat_pos.name, stat_neg=stat_neg.name) # below is part of model emb = my_slot_net(slots, label) # emb can be fc layer of size 1 similarity_norm = fluid.layers.sigmoid(fluid.layers.clip(\ emb, min=-15.0, max=15.0), name="similarity_norm")\ binary_predict = fluid.layers.concat(input=[\ fluid.layers.elementwise_sub(\ fluid.layers.ceil(similarity_norm), similarity_norm),\ similarity_norm], axis=1) auc, batch_auc, [batch_stat_pos, batch_stat_neg, stat_pos, \ stat_neg] = fluid.layers.auc(input=binary_predict,\ label=label, curve='ROC',\ num_thresholds=4096) """ auc_value = self.get_global_auc(scope, stat_pos, stat_neg) self.rank0_print(print_prefix + " global auc = %s" % auc_value) def get_global_auc(self, scope=fluid.global_scope(), stat_pos="_generated_var_2", stat_neg="_generated_var_3"): """ Get global auc of all distributed workers. Args: scope(Scope): Scope object, default is fluid.global_scope() stat_pos(str): name of auc pos bucket Variable stat_neg(str): name of auc neg bucket Variable Returns: auc_value(float), total_ins_num(int) Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() auc_value, _ = fleet_util.get_global_auc(myscope, stat_pos=stat_pos, stat_neg=stat_neg) """ if scope.find_var(stat_pos) is None or scope.find_var(stat_neg) is None: self.rank0_print("not found auc bucket") return None fleet._role_maker._barrier_worker() # auc pos bucket pos = np.array(scope.find_var(stat_pos).get_tensor()) # auc pos bucket shape old_pos_shape = np.array(pos.shape) # reshape to one dim pos = pos.reshape(-1) global_pos = np.copy(pos) * 0 # mpi allreduce fleet._role_maker._all_reduce(pos, global_pos) # reshape to its original shape global_pos = global_pos.reshape(old_pos_shape) # auc neg bucket neg = np.array(scope.find_var(stat_neg).get_tensor()) old_neg_shape = np.array(neg.shape) neg = neg.reshape(-1) global_neg = np.copy(neg) * 0 fleet._role_maker._all_reduce(neg, global_neg) global_neg = global_neg.reshape(old_neg_shape) # calculate auc num_bucket = len(global_pos[0]) area = 0.0 pos = 0.0 neg = 0.0 new_pos = 0.0 new_neg = 0.0 total_ins_num = 0 for i in xrange(num_bucket): index = num_bucket - 1 - i new_pos = pos + global_pos[0][index] total_ins_num += global_pos[0][index] new_neg = neg + global_neg[0][index] total_ins_num += global_neg[0][index] area += (new_neg - neg) * (pos + new_pos) / 2 pos = new_pos neg = new_neg auc_value = None if pos * neg == 0 or total_ins_num == 0: auc_value = 0.5 else: auc_value = area / (pos * neg) fleet._role_maker._barrier_worker() return auc_value def load_fleet_model_one_table(self, table_id, path): """ load pslib model to one table Args: table_id(int): load model to one table, default is None, which mean load all table. path(str): model path Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.load_fleet_model("hdfs:/my/model/path", table_id=1) """ fleet.load_one_table(table_id, path) def load_fleet_model(self, path, mode=0): """ load pslib model Args: path(str): model path mode(str): 0 or 1, which means load checkpoint or delta model, default is 0 Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.load_fleet_model("hdfs:/my/model/path") fleet_util.load_fleet_model("hdfs:/my/model/path", mode=0) """ fleet.init_server(path, mode=mode) def save_fleet_model(self, path, mode=0): """ save pslib model Args: path(str): model path mode(str): 0 or 1, which means save checkpoint or delta model, default is 0 Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_fleet_model("hdfs:/my/model/path") """ fleet.save_persistables(None, path, mode=mode) def _get_xbox_str(self, output_path, day, model_path, xbox_base_key, data_path, hadoop_fs_name, monitor_data={}, mode="patch"): xbox_dict = collections.OrderedDict() if mode == "base": xbox_dict["id"] = str(xbox_base_key) elif mode == "patch": xbox_dict["id"] = str(int(time.time())) else: print("warning: unknown mode %s, set it to patch" % mode) mode = "patch" xbox_dict["id"] = str(int(time.time())) xbox_dict["key"] = str(xbox_base_key) if model_path.startswith("hdfs:") or model_path.startswith("afs:"): model_path = model_path[model_path.find(":") + 1:] xbox_dict["input"] = hadoop_fs_name + model_path.rstrip("/") + "/000" xbox_dict["record_count"] = "111111" xbox_dict["partition_type"] = "2" xbox_dict["job_name"] = "default_job_name" xbox_dict["ins_tag"] = "feasign" xbox_dict["ins_path"] = data_path job_id_with_host = os.popen("echo -n ${JOB_ID}").read().strip() instance_id = os.popen("echo -n ${INSTANCE_ID}").read().strip() start_pos = instance_id.find(job_id_with_host) end_pos = instance_id.find("--") if start_pos != -1 and end_pos != -1: job_id_with_host = instance_id[start_pos:end_pos] xbox_dict["job_id"] = job_id_with_host # currently hard code here, set monitor_data empty string xbox_dict["monitor_data"] = "" xbox_dict["monitor_path"] = output_path.rstrip("/") + "/monitor/" \ + day + ".txt" xbox_dict["mpi_size"] = str(fleet.worker_num()) return json.dumps(xbox_dict) def write_model_donefile(self, output_path, day, pass_id, xbox_base_key, hadoop_fs_name, hadoop_fs_ugi, hadoop_home="$HADOOP_HOME", donefile_name="donefile.txt"): """ write donefile when save model Args: output_path(str): output path day(str|int): training day pass_id(str|int): training pass id xbox_base_key(str|int): xbox base key hadoop_fs_name(str): hdfs/afs fs name hadoop_fs_ugi(str): hdfs/afs fs ugi hadoop_home(str): hadoop home, default is "$HADOOP_HOME" donefile_name(str): donefile name, default is "donefile.txt" Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.write_model_donefile(output_path="hdfs:/my/output", model_path="hdfs:/my/model", day=20190723, pass_id=66, xbox_base_key=int(time.time()), hadoop_fs_name="hdfs://xxx", hadoop_fs_ugi="user,passwd") """ day = str(day) pass_id = str(pass_id) xbox_base_key = int(xbox_base_key) if pass_id != "-1": suffix_name = "/%s/%s/" % (day, pass_id) model_path = output_path.rstrip("/") + suffix_name else: suffix_name = "/%s/0/" % day model_path = output_path.rstrip("/") + suffix_name if fleet.worker_index() == 0: donefile_path = output_path + "/" + donefile_name content = "%s\t%lu\t%s\t%s\t%d" % (day, xbox_base_key,\ model_path, pass_id, 0) configs = { "fs.default.name": hadoop_fs_name, "hadoop.job.ugi": hadoop_fs_ugi } client = HDFSClient(hadoop_home, configs) if client.is_file(donefile_path): pre_content = client.cat(donefile_path) pre_content_list = pre_content.split("\n") day_list = [i.split("\t")[0] for i in pre_content_list] pass_list = [i.split("\t")[3] for i in pre_content_list] exist = False for i in range(len(day_list)): if int(day) == int(day_list[i]) and \ int(pass_id) == int(pass_list[i]): exist = True break if not exist: with open(donefile_name, "w") as f: f.write(pre_content + "\n") f.write(content + "\n") client.delete(donefile_path) client.upload( output_path, donefile_name, multi_processes=1, overwrite=False) self.rank0_error("write %s/%s %s succeed" % \ (day, pass_id, donefile_name)) else: self.rank0_error("not write %s because %s/%s already " "exists" % (donefile_name, day, pass_id)) else: with open(donefile_name, "w") as f: f.write(content + "\n") client.upload( output_path, donefile_name, multi_processes=1, overwrite=False) self.rank0_error("write %s/%s %s succeed" % \ (day, pass_id, donefile_name)) fleet._role_maker._barrier_worker() def write_xbox_donefile(self, output_path, day, pass_id, xbox_base_key, data_path, hadoop_fs_name, hadoop_fs_ugi, monitor_data={}, hadoop_home="$HADOOP_HOME", donefile_name=None): """ write delta donefile or xbox base donefile Args: output_path(str): output path day(str|int): training day of model pass_id(str|int): training pass id of model xbox_base_key(str|int): xbox base key data_path(str|list): training data path hadoop_fs_name(str): hdfs/afs fs name hadoop_fs_ugi(str): hdfs/afs fs ugi monitor_data(dict): metrics hadoop_home(str): hadoop home, default is "$HADOOP_HOME" donefile_name(str): donefile name, default is None" Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.write_xbox_donefile( output_path="hdfs:/my/output/", model_path="hdfs:/my/output/20190722/01", day=20190722, pass_id=1, xbox_base_key=int(time.time()), data_path="hdfs:/my/data/", hadoop_fs_name="hdfs://xxx", hadoop_fs_ugi="user,passwd", monitor_data={} ) """ day = str(day) pass_id = str(pass_id) xbox_base_key = int(xbox_base_key) mode = None if pass_id != "-1": mode = "patch" suffix_name = "/%s/delta-%s/" % (day, pass_id) model_path = output_path.rstrip("/") + suffix_name if donefile_name is None: donefile_name = "xbox_patch_done.txt" else: mode = "base" suffix_name = "/%s/base/" % day model_path = output_path.rstrip("/") + suffix_name if donefile_name is None: donefile_name = "xbox_base_done.txt" if isinstance(data_path, list): data_path = ",".join(data_path) if fleet.worker_index() == 0: donefile_path = output_path + "/" + donefile_name xbox_str = self._get_xbox_str(output_path, day, model_path, \ xbox_base_key, data_path, hadoop_fs_name, monitor_data={}, mode=mode) configs = { "fs.default.name": hadoop_fs_name, "hadoop.job.ugi": hadoop_fs_ugi } client = HDFSClient(hadoop_home, configs) if client.is_file(donefile_path): pre_content = client.cat(donefile_path) last_dict = json.loads(pre_content.split("\n")[-1]) last_day = last_dict["input"].split("/")[-3] last_pass = last_dict["input"].split("/")[-2].split("-")[-1] exist = False if int(day) < int(last_day) or \ int(day) == int(last_day) and \ int(pass_id) <= int(last_pass): exist = True if not exist: with open(donefile_name, "w") as f: f.write(pre_content + "\n") f.write(xbox_str + "\n") client.delete(donefile_path) client.upload( output_path, donefile_name, multi_processes=1, overwrite=False) self.rank0_error("write %s/%s %s succeed" % \ (day, pass_id, donefile_name)) else: self.rank0_error("not write %s because %s/%s already " "exists" % (donefile_name, day, pass_id)) else: with open(donefile_name, "w") as f: f.write(xbox_str + "\n") client.upload( output_path, donefile_name, multi_processes=1, overwrite=False) self.rank0_error("write %s/%s %s succeed" % \ (day, pass_id, donefile_name)) fleet._role_maker._barrier_worker() def write_cache_donefile(self, output_path, day, pass_id, key_num, hadoop_fs_name, hadoop_fs_ugi, hadoop_home="$HADOOP_HOME", donefile_name="sparse_cache.meta", **kwargs): """ write cache donefile Args: output_path(str): output path day(str|int): training day of model pass_id(str|int): training pass id of model key_num(str|int): save cache return value hadoop_fs_name(str): hdfs/afs fs name hadoop_fs_ugi(str): hdfs/afs fs ugi hadoop_home(str): hadoop home, default is "$HADOOP_HOME" donefile_name(str): donefile name, default is "sparse_cache.meta" kwargs(dict): user defined properties file_num(int): cache file num table_id(int): cache table id Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.write_cache_donefile( output_path="hdfs:/my/output/", day=20190722, pass_id=1, key_num=123456, hadoop_fs_name="hdfs://xxx", hadoop_fs_ugi="user,passwd", ) """ day = str(day) pass_id = str(pass_id) key_num = int(key_num) file_num = kwargs.get("file_num", 16) table_id = kwargs.get("table_id", 0) if pass_id != "-1": suffix_name = "/%s/delta-%s/%03d_cache" % (day, pass_id, table_id) model_path = output_path.rstrip("/") + suffix_name else: suffix_name = "/%s/base/%03d_cache" % (day, table_id) model_path = output_path.rstrip("/") + suffix_name if fleet.worker_index() == 0: donefile_path = model_path + "/" + donefile_name configs = { "fs.default.name": hadoop_fs_name, "hadoop.job.ugi": hadoop_fs_ugi } client = HDFSClient(hadoop_home, configs) if client.is_file(donefile_path): self.rank0_error( \ "not write because %s already exists" % donefile_path) else: meta_str = "file_prefix:part\npart_num:%s\nkey_num:%d\n" \ % (file_num, key_num) with open(donefile_name, "w") as f: f.write(meta_str) client.upload( model_path, donefile_name, multi_processes=1, overwrite=False) self.rank0_error("write %s succeed" % donefile_path) fleet._role_maker._barrier_worker() def load_model(self, output_path, day, pass_id): """ load pslib model Args: output_path(str): output path day(str|int): training day pass_id(str|int): training pass id Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.load_model("hdfs:/my/path", 20190722, 88) """ day = str(day) pass_id = str(pass_id) suffix_name = "/%s/%s/" % (day, pass_id) load_path = output_path + suffix_name self.rank0_error("going to load_model %s" % load_path) self.load_fleet_model(load_path) self.rank0_error("load_model done") def save_model(self, output_path, day, pass_id): """ save pslib model Args: output_path(str): output path day(str|int): training day pass_id(str|int): training pass id Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_model("hdfs:/my/path", 20190722, 88) """ day = str(day) pass_id = str(pass_id) suffix_name = "/%s/%s/" % (day, pass_id) model_path = output_path + suffix_name self.rank0_print("going to save_model %s" % model_path) self.save_fleet_model(model_path) self.rank0_print("save_model done") def save_batch_model(self, output_path, day): """ save batch model Args: output_path(str): output path day(str|int): training day Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_batch_model("hdfs:/my/path", 20190722) """ day = str(day) suffix_name = "/%s/0/" % day model_path = output_path + suffix_name self.rank0_print("going to save_model %s" % model_path) fleet.save_persistables(None, model_path, mode=3) self.rank0_print("save_batch_model done") def save_delta_model(self, output_path, day, pass_id): """ save delta model Args: output_path(str): output path day(str|int): training day pass_id(str|int): training pass id Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_batch_model("hdfs:/my/path", 20190722, 88) """ day = str(day) pass_id = str(pass_id) suffix_name = "/%s/delta-%s/" % (day, pass_id) model_path = output_path + suffix_name self.rank0_print("going to save_delta_model %s" % model_path) fleet.save_persistables(None, model_path, mode=1) self.rank0_print("save_delta_model done") def save_xbox_base_model(self, output_path, day): """ save xbox base model Args: output_path(str): output path day(str|int): training day Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_xbox_base_model("hdfs:/my/path", 20190722, 88) """ day = str(day) suffix_name = "/%s/base/" % day model_path = output_path + suffix_name self.rank0_print("going to save_xbox_base_model " + model_path) fleet.save_persistables(None, model_path, mode=2) self.rank0_print("save_xbox_base_model done") def save_cache_model(self, output_path, day, pass_id, mode=1, **kwargs): """ save cache model Args: output_path(str): output path day(str|int): training day pass_id(str|int): training pass id mode(str|int): save mode kwargs(dict): user defined properties table_id(int): table id to save cache Returns: key_num(int): cache key num Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_cache_model("hdfs:/my/path", 20190722, 88) """ day = str(day) pass_id = str(pass_id) mode = int(mode) table_id = kwargs.get("table_id", 0) suffix_name = "/%s/delta-%s" % (day, pass_id) model_path = output_path.rstrip("/") + suffix_name self.rank0_print("going to save_cache_model %s" % model_path) key_num = fleet.save_cache_model( None, model_path, mode=mode, table_id=table_id) self.rank0_print("save_cache_model done") return key_num def save_cache_base_model(self, output_path, day, **kwargs): """ save cache model Args: output_path(str): output path day(str|int): training day pass_id(str|int): training pass id kwargs(dict): user defined properties table_id(int): table id to save cache Returns: key_num(int): cache key num Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_cache_base_model("hdfs:/my/path", 20190722) """ day = str(day) table_id = kwargs.get("table_id", 0) suffix_name = "/%s/base" % day model_path = output_path.rstrip("/") + suffix_name self.rank0_print("going to save_cache_base_model %s" % model_path) key_num = fleet.save_cache_model( None, model_path, mode=2, table_id=table_id) self.rank0_print("save_cache_base_model done") return key_num def pull_all_dense_params(self, scope, program): """ pull all dense params in trainer of rank 0 Args: scope(Scope): fluid Scope program(Program): fluid Program Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.pull_all_dense_params(my_scope, my_program) """ fleet._role_maker._barrier_worker() if fleet._role_maker.is_first_worker(): prog_id = str(id(program)) tables = fleet._opt_info["program_id_to_worker"][prog_id].\ get_desc().dense_table prog_conf = fleet._opt_info['program_configs'][prog_id] prog_tables = {} for key in prog_conf: if "dense" not in key: continue for table_id in prog_conf[key]: prog_tables[int(table_id)] = 0 for table in tables: if int(table.table_id) not in prog_tables: continue var_name_list = [] for i in range(0, len(table.dense_variable_name)): var_name = table.dense_variable_name[i] if scope.find_var(var_name) is None: raise ValueError("var " + var_name + " not found in scope " + "when pull dense") var_name_list.append(var_name) fleet._fleet_ptr.pull_dense(scope, int(table.table_id), var_name_list) fleet._role_maker._barrier_worker() def save_paddle_inference_model(self, executor, scope, program, feeded_vars, target_vars, output_path, day, pass_id, hadoop_fs_name, hadoop_fs_ugi, hadoop_home="$HADOOP_HOME", save_combine=True): """ save paddle inference model, and upload to hdfs dnn_plugin path Args: executor(Executor): fluid Executor scope(Scope): fluid Scope program(Program): fluid Program feeded_vars(list[Variable]): feed vars target_vars(list[variable]): fetch vars output_path(str): hdfs/afs output path day(str|int): training day pass_id(str|int): training pass hadoop_fs_name(str): hadoop fs name hadoop_fs_ugi(str): hadoop fs ugi hadoop_home(str): hadoop home, default is "$HADOOP_HOME" save_combine(bool): whether to save in a file or separate files, default is True Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_paddle_inference_model(exe, join_scope, join_program, feeded_vars, target_vars, "hdfs:/my/output/path/", day=20190727, pass_id=6, hadoop_fs_name="xxx", hadoop_fs_ugi="xxx,xxx") """ day = str(day) pass_id = str(pass_id) feeded_var_names = [i.name for i in feeded_vars] model_name = "inference_model" # pull dense before save self.pull_all_dense_params(scope, program) if fleet.worker_index() == 0: with fluid.scope_guard(scope): if save_combine: fluid.io.save_inference_model( dirname=model_name, feeded_var_names=feeded_var_names, target_vars=target_vars, executor=executor, main_program=program.clone(), params_filename="params") else: fluid.io.save_inference_model( dirname=model_name, feeded_var_names=feeded_var_names, target_vars=target_vars, executor=executor, main_program=program.clone()) configs = { "fs.default.name": hadoop_fs_name, "hadoop.job.ugi": hadoop_fs_ugi } client = HDFSClient(hadoop_home, configs) if pass_id == "-1": dest = "%s/%s/base/dnn_plugin/" % (output_path, day) else: dest = "%s/%s/delta-%s/dnn_plugin/" % (output_path, day, pass_id) if not client.is_exist(dest): client.makedirs(dest) client.upload(dest, model_name) fleet._role_maker._barrier_worker() def save_paddle_params(self, executor, scope, program, model_name, output_path, day, pass_id, hadoop_fs_name, hadoop_fs_ugi, hadoop_home="$HADOOP_HOME", var_names=None, save_combine=True): """ save paddle model, and upload to hdfs dnn_plugin path Args: executor(Executor): fluid Executor scope(Scope): fluid Scope program(Program): fluid Program model_name(str): save model local dir or filename output_path(str): hdfs/afs output path day(str|int): training day pass_id(str|int): training pass hadoop_fs_name(str): hadoop fs name hadoop_fs_ugi(str): hadoop fs ugi hadoop_home(str): hadoop home, default is "$HADOOP_HOME" var_names(list): save persistable var names, default is None save_combine(bool): whether to save in a file or separate files, default is True Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.save_paddle_params(exe, join_scope, join_program, "paddle_dense.model.0", "hdfs:/my/output/path/", day=20190727, pass_id=6, hadoop_fs_name="xxx", hadoop_fs_ugi="xxx,xxx", var_names=join_all_var_names) fleet_util.save_paddle_params(exe, join_scope, join_program, "paddle_dense.model.usr.0", "hdfs:/my/output/path/", day=20190727, pass_id=6, hadoop_fs_name="xxx", hadoop_fs_ugi="xxx,xxx", var_names=join_user_var_names) fleet_util.save_paddle_params(exe, join_scope, join_program, "paddle_dense.model.item.0", "hdfs:/my/output/path/", day=20190727, pass_id=6, hadoop_fs_name="xxx", hadoop_fs_ugi="xxx,xxx", var_names=join_user_item_names) """ day = str(day) pass_id = str(pass_id) # pull dense before save self.pull_all_dense_params(scope, program) if fleet.worker_index() == 0: vars = [program.global_block().var(i) for i in var_names] with fluid.scope_guard(scope): if save_combine: fluid.io.save_vars( executor, "./", program, vars=vars, filename=model_name) else: fluid.io.save_vars(executor, model_name, program, vars=vars) configs = { "fs.default.name": hadoop_fs_name, "hadoop.job.ugi": hadoop_fs_ugi } client = HDFSClient(hadoop_home, configs) if pass_id == "-1": dest = "%s/%s/base/dnn_plugin/" % (output_path, day) else: dest = "%s/%s/delta-%s/dnn_plugin/" % (output_path, day, pass_id) if not client.is_exist(dest): client.makedirs(dest) if os.path.isdir(model_name): client.upload_dir(dest, model_name) else: client.upload(dest, model_name) fleet._role_maker._barrier_worker() def get_last_save_xbox_base(self, output_path, hadoop_fs_name, hadoop_fs_ugi, hadoop_home="$HADOOP_HOME"): """ get last saved base xbox info from xbox_base_done.txt Args: output_path(str): output path hadoop_fs_name(str): hdfs/afs fs_name hadoop_fs_ugi(str): hdfs/afs fs_ugi hadoop_home(str): hadoop home, default is "$HADOOP_HOME" Returns: [last_save_day, last_path, xbox_base_key] last_save_day(int): day of saved model last_path(str): model path xbox_base_key(int): xbox key Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() last_save_day, last_path, xbox_base_key = \ fleet_util.get_last_save_xbox_base("hdfs:/my/path", 20190722, 88) """ donefile_path = output_path + "/xbox_base_done.txt" configs = { "fs.default.name": hadoop_fs_name, "hadoop.job.ugi": hadoop_fs_ugi } client = HDFSClient(hadoop_home, configs) if not client.is_file(donefile_path): return [-1, -1, int(time.time())] pre_content = client.cat(donefile_path) last_dict = json.loads(pre_content.split("\n")[-1]) last_day = int(last_dict["input"].split("/")[-3]) last_path = "/".join(last_dict["input"].split("/")[:-1]) xbox_base_key = int(last_dict["key"]) return [last_day, last_path, xbox_base_key] def get_last_save_xbox(self, output_path, hadoop_fs_name, hadoop_fs_ugi, hadoop_home="$HADOOP_HOME"): """ get last saved xbox info from xbox_patch_done.txt Args: output_path(str): output path hadoop_fs_name(str): hdfs/afs fs_name hadoop_fs_ugi(str): hdfs/afs fs_ugi hadoop_home(str): hadoop home, default is "$HADOOP_HOME" Returns: [last_save_day, last_save_pass, last_path, xbox_base_key] last_save_day(int): day of saved model last_save_pass(int): pass id of saved last_path(str): model path xbox_base_key(int): xbox key Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() last_save_day, last_save_pass, last_path, xbox_base_key = \ fleet_util.get_last_save_xbox("hdfs:/my/path", 20190722, 88) """ donefile_path = output_path + "/xbox_patch_done.txt" configs = { "fs.default.name": hadoop_fs_name, "hadoop.job.ugi": hadoop_fs_ugi } client = HDFSClient(hadoop_home, configs) if not client.is_file(donefile_path): return [-1, -1, "", int(time.time())] pre_content = client.cat(donefile_path) last_dict = json.loads(pre_content.split("\n")[-1]) last_day = int(last_dict["input"].split("/")[-3]) last_pass = int(last_dict["input"].split("/")[-2].split("-")[-1]) last_path = "/".join(last_dict["input"].split("/")[:-1]) xbox_base_key = int(last_dict["key"]) return [last_day, last_pass, last_path, xbox_base_key] def get_last_save_model(self, output_path, hadoop_fs_name, hadoop_fs_ugi, hadoop_home="$HADOOP_HOME"): """ get last saved model info from donefile.txt Args: output_path(str): output path hadoop_fs_name(str): hdfs/afs fs_name hadoop_fs_ugi(str): hdfs/afs fs_ugi hadoop_home(str): hadoop home, default is "$HADOOP_HOME" Returns: [last_save_day, last_save_pass, last_path, xbox_base_key] last_save_day(int): day of saved model last_save_pass(int): pass id of saved last_path(str): model path xbox_base_key(int): xbox key Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() last_save_day, last_save_pass, last_path, xbox_base_key = \ fleet_util.get_last_save_model("hdfs:/my/path", 20190722, 88) """ last_save_day = -1 last_save_pass = -1 last_path = "" donefile_path = output_path + "/donefile.txt" configs = { "fs.default.name": hadoop_fs_name, "hadoop.job.ugi": hadoop_fs_ugi } client = HDFSClient(hadoop_home, configs) if not client.is_file(donefile_path): return [-1, -1, "", int(time.time())] content = client.cat(donefile_path) content = content.split("\n")[-1].split("\t") last_save_day = int(content[0]) last_save_pass = int(content[3]) last_path = content[2] xbox_base_key = int(content[1]) return [last_save_day, last_save_pass, last_path, xbox_base_key] def get_online_pass_interval(self, days, hours, split_interval, split_per_pass, is_data_hourly_placed): """ get online pass interval Args: days(str): days to train hours(str): hours to train split_interval(int|str): split interval split_per_pass(int}str): split per pass is_data_hourly_placed(bool): is data hourly placed Returns: online_pass_interval(list) Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() online_pass_interval = fleet_util.get_online_pass_interval( days="{20190720..20190729}", hours="{0..23}", split_interval=5, split_per_pass=2, is_data_hourly_placed=False) """ days = os.popen("echo -n " + days).read().split(" ") hours = os.popen("echo -n " + hours).read().split(" ") split_interval = int(split_interval) split_per_pass = int(split_per_pass) splits_per_day = 24 * 60 / split_interval pass_per_day = splits_per_day / split_per_pass left_train_hour = int(hours[0]) right_train_hour = int(hours[-1]) start = 0 split_path = [] for i in range(splits_per_day): h = start / 60 m = start % 60 if h < left_train_hour or h > right_train_hour: start += split_interval continue if is_data_hourly_placed: split_path.append("%02d" % h) else: split_path.append("%02d%02d" % (h, m)) start += split_interval start = 0 online_pass_interval = [] for i in range(pass_per_day): online_pass_interval.append([]) for j in range(start, start + split_per_pass): online_pass_interval[i].append(split_path[j]) start += split_per_pass return online_pass_interval def get_global_metrics(self, scope=fluid.global_scope(), stat_pos_name="_generated_var_2", stat_neg_name="_generated_var_3", sqrerr_name="sqrerr", abserr_name="abserr", prob_name="prob", q_name="q", pos_ins_num_name="pos", total_ins_num_name="total"): """ get global metrics, including auc, bucket_error, mae, rmse, actual_ctr, predicted_ctr, copc, mean_predict_qvalue, total_ins_num. Args: scope(Scope): Scope object, default is fluid.global_scope() stat_pos_name(str): name of auc pos bucket Variable stat_neg_name(str): name of auc neg bucket Variable sqrerr_name(str): name of sqrerr Variable abserr_name(str): name of abserr Variable prob_name(str): name of prob Variable q_name(str): name of q Variable pos_ins_num_name(str): name of pos ins num Variable total_ins_num_name(str): name of total ins num Variable Returns: [auc, bucket_error, mae, rmse, actual_ctr, predicted_ctr, copc, mean_predict_qvalue, total_ins_num] Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() metric_list = fleet_util.get_global_metrics(myscope, stat_pos.name, stat_neg.name, local_sqrerr.name, local_abserr.name, local_prob.name, local_q.name, local_pos_ins.name, local_total_ins.name) # below is part of example model label = fluid.layers.data(name="click", shape=[-1, 1],\ dtype="int64", lod_level=0, append_batch_size=False) emb = my_slot_net(slots, label) # emb can be fc layer of size 1 similarity_norm = fluid.layers.sigmoid(fluid.layers.clip(\ emb, min=-15.0, max=15.0), name="similarity_norm")\ binary_predict = fluid.layers.concat(input=[\ fluid.layers.elementwise_sub(\ fluid.layers.ceil(similarity_norm), similarity_norm),\ similarity_norm], axis=1) auc, batch_auc, [batch_stat_pos, batch_stat_neg, stat_pos, \ stat_neg] = fluid.layers.auc(input=binary_predict,\ label=label, curve='ROC',\ num_thresholds=4096) local_sqrerr, local_abserr, local_prob, local_q, local_pos_ins,\ local_total_ins = fluid.contrib.layers.ctr_metric_bundle(\ similarity_norm, label) """ if scope.find_var(stat_pos_name) is None or \ scope.find_var(stat_neg_name) is None: self.rank0_print("not found auc bucket") return [None] * 9 elif scope.find_var(sqrerr_name) is None: self.rank0_print("not found sqrerr_name=%s" % sqrerr_name) return [None] * 9 elif scope.find_var(abserr_name) is None: self.rank0_print("not found abserr_name=%s" % abserr_name) return [None] * 9 elif scope.find_var(prob_name) is None: self.rank0_print("not found prob_name=%s" % prob_name) return [None] * 9 elif scope.find_var(q_name) is None: self.rank0_print("not found q_name=%s" % q_name) return [None] * 9 elif scope.find_var(pos_ins_num_name) is None: self.rank0_print("not found pos_ins_num_name=%s" % pos_ins_num_name) return [None] * 9 elif scope.find_var(total_ins_num_name) is None: self.rank0_print("not found total_ins_num_name=%s" % \ total_ins_num_name) return [None] * 9 # barrier worker to ensure all workers finished training fleet._role_maker._barrier_worker() # get auc auc = self.get_global_auc(scope, stat_pos_name, stat_neg_name) pos = np.array(scope.find_var(stat_pos_name).get_tensor()) # auc pos bucket shape old_pos_shape = np.array(pos.shape) # reshape to one dim pos = pos.reshape(-1) global_pos = np.copy(pos) * 0 # mpi allreduce fleet._role_maker._all_reduce(pos, global_pos) # reshape to its original shape global_pos = global_pos.reshape(old_pos_shape) # auc neg bucket neg = np.array(scope.find_var(stat_neg_name).get_tensor()) old_neg_shape = np.array(neg.shape) neg = neg.reshape(-1) global_neg = np.copy(neg) * 0 fleet._role_maker._all_reduce(neg, global_neg) global_neg = global_neg.reshape(old_neg_shape) num_bucket = len(global_pos[0]) def get_metric(name): metric = np.array(scope.find_var(name).get_tensor()) old_metric_shape = np.array(metric.shape) metric = metric.reshape(-1) global_metric = np.copy(metric) * 0 fleet._role_maker._all_reduce(metric, global_metric) global_metric = global_metric.reshape(old_metric_shape) return global_metric[0] global_sqrerr = get_metric(sqrerr_name) global_abserr = get_metric(abserr_name) global_prob = get_metric(prob_name) global_q_value = get_metric(q_name) # note: get ins_num from auc bucket is not actual value, # so get it from metric op pos_ins_num = get_metric(pos_ins_num_name) total_ins_num = get_metric(total_ins_num_name) neg_ins_num = total_ins_num - pos_ins_num mae = global_abserr / total_ins_num rmse = math.sqrt(global_sqrerr / total_ins_num) return_actual_ctr = pos_ins_num / total_ins_num predicted_ctr = global_prob / total_ins_num mean_predict_qvalue = global_q_value / total_ins_num copc = 0.0 if abs(predicted_ctr > 1e-6): copc = return_actual_ctr / predicted_ctr # calculate bucket error last_ctr = -1.0 impression_sum = 0.0 ctr_sum = 0.0 click_sum = 0.0 error_sum = 0.0 error_count = 0.0 click = 0.0 show = 0.0 ctr = 0.0 adjust_ctr = 0.0 relative_error = 0.0 actual_ctr = 0.0 relative_ctr_error = 0.0 k_max_span = 0.01 k_relative_error_bound = 0.05 for i in xrange(num_bucket): click = global_pos[0][i] show = global_pos[0][i] + global_neg[0][i] ctr = float(i) / num_bucket if abs(ctr - last_ctr) > k_max_span: last_ctr = ctr impression_sum = 0.0 ctr_sum = 0.0 click_sum = 0.0 impression_sum += show ctr_sum += ctr * show click_sum += click if impression_sum == 0: continue adjust_ctr = ctr_sum / impression_sum if adjust_ctr == 0: continue relative_error = \ math.sqrt((1 - adjust_ctr) / (adjust_ctr * impression_sum)) if relative_error < k_relative_error_bound: actual_ctr = click_sum / impression_sum relative_ctr_error = abs(actual_ctr / adjust_ctr - 1) error_sum += relative_ctr_error * impression_sum error_count += impression_sum last_ctr = -1 bucket_error = error_sum / error_count if error_count > 0 else 0.0 return [ auc, bucket_error, mae, rmse, return_actual_ctr, predicted_ctr, copc, mean_predict_qvalue, int(total_ins_num) ] def print_global_metrics(self, scope=fluid.global_scope(), stat_pos_name="_generated_var_2", stat_neg_name="_generated_var_3", sqrerr_name="sqrerr", abserr_name="abserr", prob_name="prob", q_name="q", pos_ins_num_name="pos", total_ins_num_name="total", print_prefix=""): """ print global metrics, including auc, bucket_error, mae, rmse, actual_ctr, predicted_ctr, copc, mean_predict_qvalue, total_ins_num. Args: scope(Scope): Scope object, default is fluid.global_scope() stat_pos_name(str): name of auc pos bucket Variable stat_neg_name(str): name of auc neg bucket Variable sqrerr_name(str): name of sqrerr Variable abserr_name(str): name of abserr Variable prob_name(str): name of prob Variable q_name(str): name of q Variable pos_ins_num_name(str): name of pos ins num Variable total_ins_num_name(str): name of total ins num Variable print_prefix(str): print prefix Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() fleet_util.print_global_metrics(myscope, stat_pos.name, stat_neg.name, local_sqrerr.name, local_abserr.name, local_prob.name, local_q.name, local_pos_ins.name, local_total_ins.name) # below is part of model label = fluid.layers.data(name="click", shape=[-1, 1],\ dtype="int64", lod_level=0, append_batch_size=False) emb = my_slot_net(slots, label) # emb can be fc layer of size 1 similarity_norm = fluid.layers.sigmoid(fluid.layers.clip(\ emb, min=-15.0, max=15.0), name="similarity_norm")\ binary_predict = fluid.layers.concat(input=[\ fluid.layers.elementwise_sub(\ fluid.layers.ceil(similarity_norm), similarity_norm),\ similarity_norm], axis=1) auc, batch_auc, [batch_stat_pos, batch_stat_neg, stat_pos, \ stat_neg] = fluid.layers.auc(input=binary_predict,\ label=label, curve='ROC',\ num_thresholds=4096) local_sqrerr, local_abserr, local_prob, local_q, local_pos_ins, \ local_total_ins = fluid.contrib.layers.ctr_metric_bundle(\ similarity_norm, label) """ if scope.find_var(stat_pos_name) is None or \ scope.find_var(stat_neg_name) is None: self.rank0_print("not found auc bucket") return elif scope.find_var(sqrerr_name) is None: self.rank0_print("not found sqrerr_name=%s" % sqrerr_name) return elif scope.find_var(abserr_name) is None: self.rank0_print("not found abserr_name=%s" % abserr_name) return elif scope.find_var(prob_name) is None: self.rank0_print("not found prob_name=%s" % prob_name) return elif scope.find_var(q_name) is None: self.rank0_print("not found q_name=%s" % q_name) return elif scope.find_var(pos_ins_num_name) is None: self.rank0_print("not found pos_ins_num_name=%s" % pos_ins_num_name) return elif scope.find_var(total_ins_num_name) is None: self.rank0_print("not found total_ins_num_name=%s" % \ total_ins_num_name) return auc, bucket_error, mae, rmse, actual_ctr, predicted_ctr, copc,\ mean_predict_qvalue, total_ins_num = self.get_global_metrics(\ scope, stat_pos_name, stat_neg_name, sqrerr_name, abserr_name,\ prob_name, q_name, pos_ins_num_name, total_ins_num_name) self.rank0_print("%s global AUC=%.6f BUCKET_ERROR=%.6f MAE=%.6f " "RMSE=%.6f Actural_CTR=%.6f Predicted_CTR=%.6f " "COPC=%.6f MEAN Q_VALUE=%.6f Ins number=%s" % (print_prefix, auc, bucket_error, mae, rmse, actual_ctr, predicted_ctr, copc, mean_predict_qvalue, total_ins_num)) def program_type_trans(self, prog_dir, prog_fn, is_text): return utils.program_type_trans(prog_dir, prog_fn, is_text) def draw_from_program_file(self, model_filename, is_text, output_dir, output_filename): """draw program from file""" program = utils.load_program(model_filename, is_text) utils.graphviz(program.global_block(), output_dir, output_filename) def draw_from_program(self, program, output_dir, output_name): """draw Program""" utils.graphviz(program.global_block(), output_dir, output_name) def check_two_programs(self, config): train_prog = utils.load_program(config.train_prog_path, config.is_text_train_program) pruned_prog = utils.load_program(config.pruned_prog_path, config.is_text_pruned_program) if config.draw: pruned_dir = os.path.dirname(config.pruned_prog_path) self.draw_from_program(pruned_prog, pruned_dir, config.draw_out_name) res = utils.check_pruned_program_vars(train_prog, pruned_prog) if res: _logger.info("check_programs succeed.") else: _logger.info( "check_programs failed. pruned program and train program not match!" ) return res def check_vars_and_dump(self, config): _logger.info("start check_vars_and_dump.") results = utils.check_saved_vars_try_dump( config.dump_model_dir, config.dump_program_filename, config.is_text_dump_program, config.feed_config, config.fetch_config, config.batch_size, config.save_params_filename) _logger.info("check_vars_and_dump succeed.") return results def parse_program_proto(self, prog_path, is_text, output_dir): """ Parse program.proto into a more readable format. This function will generate three files: output_dir/vars_all.log, output_dir/vars_persistable.log, output_dir/ops.log. Args: prog_path(str): proto file path to be parsed. is_text(bool): proto file is human-readale format or not(binary). output_dir(str): output dir. Examples: .. code-block:: python from paddle.fluid.incubate.fleet.utils.fleet_util import FleetUtil fleet_util = FleetUtil() program_path = "./program.pbtxt" is_text = True output_dir = "/tmp/" fleet_util.parse_program_proto(program_path, is_text, output_dir) """ program = utils.load_program(prog_path, is_text) utils.parse_program(program, output_dir)