runner.py 21.0 KB
Newer Older
C
Chengmo 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
# Copyright (c) 2020 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 os
import time
C
Chengmo 已提交
19
import numpy as np
C
Chengmo 已提交
20
import paddle.fluid as fluid
C
Chengmo 已提交
21

C
Chengmo 已提交
22 23 24 25 26 27 28
from paddlerec.core.utils import envs

__all__ = [
    "RunnerBase", "SingleRunner", "PSRunner", "CollectiveRunner", "PslibRunner"
]


C
Chengmo 已提交
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
def as_numpy(tensor):
    """
    Convert a Tensor to a numpy.ndarray, its only support Tensor without LoD information.
    For higher dimensional sequence data, please use LoDTensor directly.

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          import numpy

          new_scope = fluid.Scope()
          with fluid.scope_guard(new_scope):
              fluid.global_scope().var("data").get_tensor().set(numpy.ones((2, 2)), fluid.CPUPlace())
          tensor = new_scope.find_var("data").get_tensor()
          fluid.executor.as_numpy(tensor) # or numpy.array(new_scope.find_var("data").get_tensor())

    Args:
       tensor(Variable): a instance of Tensor

    Returns:
        numpy.ndarray
    """
    if isinstance(tensor, fluid.core.LoDTensorArray):
        return [as_numpy(t) for t in tensor]
    if isinstance(tensor, list):
        return [as_numpy(t) for t in tensor]
    assert isinstance(tensor, fluid.core.LoDTensor)
    lod = tensor.lod()
    # (todo) need print lod or return it for user
    if tensor._is_initialized():
        return np.array(tensor)
    else:
        return None


C
Chengmo 已提交
65 66 67 68 69 70 71 72 73 74 75 76 77
class RunnerBase(object):
    """R
    """

    def __init__(self, context):
        pass

    def exuctor(self, context):
        pass

    def _run(self, context, model_dict):
        reader_name = model_dict["dataset_name"]
        name = "dataset." + reader_name + "."
T
tangwei 已提交
78

C
Chengmo 已提交
79 80 81 82 83 84 85 86 87
        if envs.get_global_env(name + "type") == "DataLoader":
            self._executor_dataloader_train(model_dict, context)
        else:
            self._executor_dataset_train(model_dict, context)

    def _executor_dataset_train(self, model_dict, context):
        reader_name = model_dict["dataset_name"]
        model_name = model_dict["name"]
        model_class = context["model"][model_dict["name"]]["model"]
88

C
Chengmo 已提交
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108
        fetch_vars = []
        fetch_alias = []
        fetch_period = int(
            envs.get_global_env("runner." + context["runner_name"] +
                                ".print_interval", 20))
        scope = context["model"][model_name]["scope"]
        program = context["model"][model_name]["main_program"]
        reader = context["dataset"][reader_name]

        with fluid.scope_guard(scope):
            if context["is_infer"]:
                metrics = model_class.get_infer_results()
                if metrics:
                    fetch_vars = metrics.values()
                    fetch_alias = metrics.keys()
                context["exe"].infer_from_dataset(
                    program=program,
                    dataset=reader,
                    fetch_list=fetch_vars,
                    fetch_info=fetch_alias,
X
xjqbest 已提交
109 110
                    print_period=fetch_period,
                    debug=envs.get_global_env("debug", False))
C
Chengmo 已提交
111 112 113 114 115 116 117 118 119 120 121
            else:
                metrics = model_class.get_metrics()
                if metrics:
                    fetch_vars = metrics.values()
                    fetch_alias = metrics.keys()
                with fluid.scope_guard(scope):
                    context["exe"].train_from_dataset(
                        program=program,
                        dataset=reader,
                        fetch_list=fetch_vars,
                        fetch_info=fetch_alias,
X
xjqbest 已提交
122 123
                        print_period=fetch_period,
                        debug=envs.get_global_env("debug", False))
C
Chengmo 已提交
124 125 126 127

    def _executor_dataloader_train(self, model_dict, context):
        model_name = model_dict["name"]
        model_class = context["model"][model_dict["name"]]["model"]
128
        program = self._get_dataloader_program(model_dict, context)
C
Chengmo 已提交
129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152

        fetch_period = int(
            envs.get_global_env("runner." + context["runner_name"] +
                                ".print_interval", 20))
        if context["is_infer"]:
            metrics = model_class.get_infer_results()
        else:
            metrics = model_class.get_metrics()

        metrics_varnames = []
        metrics_format = []
        metrics_format.append("{}: {{}}".format("batch"))
        for name, var in metrics.items():
            metrics_varnames.append(var.name)
            metrics_format.append("{}: {{}}".format(name))
        metrics_format = ", ".join(metrics_format)

        reader = context["model"][model_dict["name"]]["model"]._data_loader
        reader.start()
        batch_id = 0
        scope = context["model"][model_name]["scope"]
        with fluid.scope_guard(scope):
            try:
                while True:
C
Chengmo 已提交
153 154 155 156
                    metrics_tensors = context["exe"].run(
                        program=program,
                        fetch_list=metrics_varnames,
                        return_numpy=False)
C
Chengmo 已提交
157
                    metrics = [batch_id]
C
Chengmo 已提交
158 159 160 161 162

                    metrics_rets = [
                        as_numpy(metrics_tensor)
                        for metrics_tensor in metrics_tensors
                    ]
C
Chengmo 已提交
163 164 165 166 167 168 169 170
                    metrics.extend(metrics_rets)

                    if batch_id % fetch_period == 0 and batch_id != 0:
                        print(metrics_format.format(*metrics))
                    batch_id += 1
            except fluid.core.EOFException:
                reader.reset()

M
update  
malin10 已提交
171 172 173
        if batch_id > 0:
            print(metrics_format.format(*metrics))

174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191
    def _get_dataloader_program(self, model_dict, context):
        model_name = model_dict["name"]
        if context["model"][model_name]["compiled_program"] == None:
            if context["is_infer"]:
                program = context["model"][model_name]["main_program"]
            elif context["is_fleet"]:
                if context["fleet_mode"].upper() == "PS":
                    program = self._get_ps_program(model_dict, context)
                elif context["fleet_mode"].upper() == "COLLECTIVE":
                    program = context["model"][model_name]["main_program"]
            elif not context["is_fleet"]:
                if context["device"].upper() == "CPU":
                    program = self._get_single_cpu_program(model_dict, context)
                elif context["device"].upper() == "GPU":
                    program = self._get_single_gpu_program(model_dict, context)
            context["model"][model_name]["compiled_program"] = program
        return context["model"][model_name]["compiled_program"]

C
Chengmo 已提交
192 193 194 195 196 197 198 199 200 201 202 203 204 205
    def _get_strategy(self, model_dict, context):
        _build_strategy = fluid.BuildStrategy()
        _exe_strategy = fluid.ExecutionStrategy()

        # 0: kCoeffNumDevice; 1: One; 2: Customized
        _gradient_scale_strategy = model_dict.get("gradient_scale_strategy", 0)
        if _gradient_scale_strategy == 0:
            gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.CoeffNumDevice
        elif _gradient_scale_strategy == 1:
            gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.One
        elif _gradient_scale_strategy == 2:
            gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.Customized
        else:
            raise ValueError(
T
tangwei 已提交
206
                "Unsupported config. gradient_scale_strategy must be one of [0, 1, 2]."
C
Chengmo 已提交
207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226
            )
        _build_strategy.gradient_scale_strategy = gradient_scale_strategy

        if "thread_num" in model_dict and model_dict["thread_num"] > 1:
            _build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
            _exe_strategy.num_threads = model_dict["thread_num"]
            os.environ['CPU_NUM'] = str(_exe_strategy.num_threads)

        return _exe_strategy, _build_strategy

    def _get_single_gpu_program(self, model_dict, context):
        model_name = model_dict["name"]
        return context["model"][model_name]["main_program"].clone()

    def _get_single_cpu_program(self, model_dict, context):
        model_name = model_dict["name"]
        model_class = context["model"][model_dict["name"]]["model"]
        program = context["model"][model_name]["main_program"].clone()
        _exe_strategy, _build_strategy = self._get_strategy(model_dict,
                                                            context)
M
update  
malin10 已提交
227

C
Chengmo 已提交
228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
        program = fluid.compiler.CompiledProgram(program).with_data_parallel(
            loss_name=model_class.get_avg_cost().name,
            build_strategy=_build_strategy,
            exec_strategy=_exe_strategy)
        return program

    def _get_ps_program(self, model_dict, context):
        model_name = model_dict["name"]
        model_class = context["model"][model_dict["name"]]["model"]
        program = context["model"][model_name]["main_program"].clone()

        _build_strategy = context["strategy"].get_build_strategy()
        _exe_strategy = context["strategy"].get_execute_strategy()

        if "thread_num" in model_dict and model_dict["thread_num"] > 1:
            _build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
            _exe_strategy.num_threads = model_dict["thread_num"]
            os.environ['CPU_NUM'] = str(_exe_strategy.num_threads)

        _gradient_scale_strategy = model_dict.get("gradient_scale_strategy", 0)
        if _gradient_scale_strategy == 0:
            gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.CoeffNumDevice
        elif _gradient_scale_strategy == 1:
            gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.One
        elif _gradient_scale_strategy == 2:
            gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.Customized
        else:
            raise ValueError(
                "Unsurpported config. gradient_scale_strategy must be one of [0, 1, 2]."
            )
        _build_strategy.gradient_scale_strategy = gradient_scale_strategy

        program = fluid.compiler.CompiledProgram(program).with_data_parallel(
            loss_name=model_class.get_avg_cost().name,
            build_strategy=_build_strategy,
            exec_strategy=_exe_strategy)
        return program

    def save(self, epoch_id, context, is_fleet=False):
        def need_save(epoch_id, epoch_interval, is_last=False):
268 269 270 271 272
            name = "runner." + context["runner_name"] + "."
            total_epoch = int(envs.get_global_env(name + "epochs", 1))
            if epoch_id + 1 == total_epoch:
                is_last = True

C
Chengmo 已提交
273 274 275 276 277
            if is_last:
                return True
            if epoch_id == -1:
                return False

278
            return (epoch_id + 1) % epoch_interval == 0
C
Chengmo 已提交
279 280 281 282 283 284 285 286 287 288 289 290

        def save_inference_model():
            name = "runner." + context["runner_name"] + "."
            save_interval = int(
                envs.get_global_env(name + "save_inference_interval", -1))
            if not need_save(epoch_id, save_interval, False):
                return
            feed_varnames = envs.get_global_env(
                name + "save_inference_feed_varnames", [])
            fetch_varnames = envs.get_global_env(
                name + "save_inference_fetch_varnames", [])
            if feed_varnames is None or fetch_varnames is None or feed_varnames == "" or fetch_varnames == "" or \
C
Chengmo 已提交
291
                    len(feed_varnames) == 0 or len(fetch_varnames) == 0:
C
Chengmo 已提交
292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340
                return
            fetch_vars = [
                fluid.default_main_program().global_block().vars[varname]
                for varname in fetch_varnames
            ]
            dirname = envs.get_global_env(name + "save_inference_path", None)

            assert dirname is not None
            dirname = os.path.join(dirname, str(epoch_id))

            if is_fleet:
                context["fleet"].save_inference_model(
                    context["exe"], dirname, feed_varnames, fetch_vars)
            else:
                fluid.io.save_inference_model(dirname, feed_varnames,
                                              fetch_vars, context["exe"])

        def save_persistables():
            name = "runner." + context["runner_name"] + "."
            save_interval = int(
                envs.get_global_env(name + "save_checkpoint_interval", -1))
            if not need_save(epoch_id, save_interval, False):
                return
            dirname = envs.get_global_env(name + "save_checkpoint_path", None)
            if dirname is None or dirname == "":
                return
            dirname = os.path.join(dirname, str(epoch_id))
            if is_fleet:
                context["fleet"].save_persistables(context["exe"], dirname)
            else:
                fluid.io.save_persistables(context["exe"], dirname)

        save_persistables()
        save_inference_model()


class SingleRunner(RunnerBase):
    """R
    """

    def __init__(self, context):
        print("Running SingleRunner.")
        pass

    def run(self, context):
        epochs = int(
            envs.get_global_env("runner." + context["runner_name"] +
                                ".epochs"))
        for epoch in range(epochs):
T
tangwei 已提交
341
            for model_dict in context["phases"]:
C
Chengmo 已提交
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366
                begin_time = time.time()
                self._run(context, model_dict)
                end_time = time.time()
                seconds = end_time - begin_time
                print("epoch {} done, use time: {}".format(epoch, seconds))
                with fluid.scope_guard(context["model"][model_dict["name"]][
                        "scope"]):
                    train_prog = context["model"][model_dict["name"]][
                        "default_main_program"]
                    startup_prog = context["model"][model_dict["name"]][
                        "startup_program"]
                    with fluid.program_guard(train_prog, startup_prog):
                        self.save(epoch, context)
        context["status"] = "terminal_pass"


class PSRunner(RunnerBase):
    def __init__(self, context):
        print("Running PSRunner.")
        pass

    def run(self, context):
        epochs = int(
            envs.get_global_env("runner." + context["runner_name"] +
                                ".epochs"))
T
tangwei 已提交
367
        model_dict = context["env"]["phase"][0]
C
Chengmo 已提交
368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393
        for epoch in range(epochs):
            begin_time = time.time()
            self._run(context, model_dict)
            end_time = time.time()
            seconds = end_time - begin_time
            print("epoch {} done, use time: {}".format(epoch, seconds))
            with fluid.scope_guard(context["model"][model_dict["name"]][
                    "scope"]):
                train_prog = context["model"][model_dict["name"]][
                    "main_program"]
                startup_prog = context["model"][model_dict["name"]][
                    "startup_program"]
                with fluid.program_guard(train_prog, startup_prog):
                    self.save(epoch, context, True)
        context["status"] = "terminal_pass"


class CollectiveRunner(RunnerBase):
    def __init__(self, context):
        print("Running CollectiveRunner.")
        pass

    def run(self, context):
        epochs = int(
            envs.get_global_env("runner." + context["runner_name"] +
                                ".epochs"))
T
tangwei 已提交
394
        model_dict = context["env"]["phase"][0]
C
Chengmo 已提交
395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418
        for epoch in range(epochs):
            begin_time = time.time()
            self._run(context, model_dict)
            end_time = time.time()
            seconds = end_time - begin_time
            print("epoch {} done, use time: {}".format(epoch, seconds))
            with fluid.scope_guard(context["model"][model_dict["name"]][
                    "scope"]):
                train_prog = context["model"][model_dict["name"]][
                    "default_main_program"]
                startup_prog = context["model"][model_dict["name"]][
                    "startup_program"]
                with fluid.program_guard(train_prog, startup_prog):
                    self.save(epoch, context, True)
        context["status"] = "terminal_pass"


class PslibRunner(RunnerBase):
    def __init__(self, context):
        print("Running PSRunner.")
        pass

    def run(self, context):
        context["fleet"].init_worker()
T
tangwei 已提交
419
        model_dict = context["env"]["phase"][0]
C
Chengmo 已提交
420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439
        epochs = int(
            envs.get_global_env("runner." + context["runner_name"] +
                                ".epochs"))
        for epoch in range(epochs):
            begin_time = time.time()
            self._run(context, model_dict)
            end_time = time.time()
            seconds = end_time - begin_time
            print("epoch {} done, use time: {}".format(epoch, seconds))
        """
        # online Training Can do more, As shown below:

        begin_day = datetime.datetime.strptime("begin_day_d", '%Y%m%d')
        days = int(
            envs.get_global_env("runner." + context["runner_name"] + ".days"))
        for day in range(days):
            for hour in range(24):
                day = begin_day + datetime.timedelta(days=day, hours=hour)
                day_s = day.strftime('%Y%m%d/%H')

T
tangwei 已提交
440
                for dataset in envs.get_global_env("dataset"):
C
Chengmo 已提交
441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469
                    if dataset["type"] != "DataLoader":
                        name = dataset["name"]
                        train_data_path = envs.get_global_env(name +
                                                              "data_path")
                        train_data_path = os.path.join(train_data_path, day_s)

                        file_list = [
                            os.path.join(train_data_path, x)
                            for x in os.listdir(train_data_path)
                        ]
                        context["dataset"][name].set_filelist(file_list)

                for epoch in range(epochs):
                    begin_time = time.time()
                    self._run(context, model_dict)
                    end_time = time.time()
                    seconds = end_time - begin_time
                    print("epoch {} done, use time: {}".format(epoch, seconds))
                    with fluid.scope_guard(context["model"][model_dict["name"]]
                                           ["scope"]):
                        train_prog = context["model"][model_dict["name"]][
                            "default_main_program"]
                        startup_prog = context["model"][model_dict["name"]][
                            "startup_program"]
                        with fluid.program_guard(train_prog, startup_prog):
                            self.save(epoch, context, True)

        """
        context["status"] = "terminal_pass"
470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503


class SingleInferRunner(RunnerBase):
    def __init__(self, context):
        print("Running SingleInferRunner.")
        pass

    def run(self, context):
        self._dir_check(context)

        for index, epoch_name in enumerate(self.epoch_model_name_list):
            for model_dict in context["phases"]:
                self._load(context, model_dict,
                           self.epoch_model_path_list[index])
                begin_time = time.time()
                self._run(context, model_dict)
                end_time = time.time()
                seconds = end_time - begin_time
                print("Infer {} of {} done, use time: {}".format(model_dict[
                    "name"], epoch_name, seconds))
        context["status"] = "terminal_pass"

    def _load(self, context, model_dict, model_path):
        if model_path is None or model_path == "":
            return
        print("load persistables from", model_path)

        with fluid.scope_guard(context["model"][model_dict["name"]]["scope"]):
            train_prog = context["model"][model_dict["name"]]["main_program"]
            startup_prog = context["model"][model_dict["name"]][
                "startup_program"]
            with fluid.program_guard(train_prog, startup_prog):
                fluid.io.load_persistables(
                    context["exe"], model_path, main_program=train_prog)
M
update  
malin10 已提交
504 505 506 507
            clear_metrics = context["model"][model_dict["name"]][
                "model"].get_clear_metrics()
            for var in clear_metrics:
                var.clear()
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523

    def _dir_check(self, context):
        dirname = envs.get_global_env(
            "runner." + context["runner_name"] + ".init_model_path", None)
        self.epoch_model_path_list = []
        self.epoch_model_name_list = []

        for file in os.listdir(dirname):
            file_path = os.path.join(dirname, file)
            if os.path.isdir(file_path):
                self.epoch_model_path_list.append(file_path)
                self.epoch_model_name_list.append(file)

        if len(self.epoch_model_path_list) == 0:
            self.epoch_model_path_list.append(dirname)
            self.epoch_model_name_list.append(dirname)