model.py 73.4 KB
Newer Older
1
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
#
# 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 absolute_import
from __future__ import division
from __future__ import print_function

import inspect
import os
import pickle
import numpy as np
import six
import warnings
25 26 27
import time
import socket
import contextlib
28 29
from collections import Iterable

30
import paddle
31
from paddle import fluid
32 33
from paddle.fluid import core
from paddle.fluid.framework import in_dygraph_mode, Variable, ParamBase, _current_expected_place
34 35
from paddle.fluid.framework import in_dygraph_mode, Variable
from paddle.fluid.framework import _current_expected_place as _get_device
36 37 38 39
from paddle.fluid.executor import global_scope
from paddle.fluid.io import is_belong_to_optimizer
from paddle.fluid.dygraph.base import to_variable
from paddle.fluid.dygraph.parallel import ParallelEnv
40
from paddle.fluid.dygraph.dygraph_to_static.program_translator import ProgramTranslator, FunctionSpec
41
from paddle.fluid.layers.utils import flatten
42
from paddle.fluid.layers import collective
43 44
from paddle.fluid.incubate.fleet.collective import fleet, DistributedStrategy
from paddle.fluid.incubate.fleet.base import role_maker
45

46 47
from paddle.io import DataLoader, Dataset, DistributedBatchSampler
from paddle.fluid.executor import scope_guard, Executor
48
from paddle.fluid.dygraph.layers import Layer
49
from paddle.metric import Metric
50

51 52
from paddle.static import InputSpec as Input

53 54
from .callbacks import config_callbacks

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200
__all__ = ['Model', ]

_parallel_context_initialized = False


def to_list(value):
    if value is None:
        return value
    if isinstance(value, (list, tuple)):
        return list(value)
    return [value]


def to_numpy(var):
    assert isinstance(var, (Variable, fluid.core.VarBase)), "not a variable"
    if isinstance(var, fluid.core.VarBase):
        return var.numpy()
    t = global_scope().find_var(var.name).get_tensor()
    return np.array(t)


def flatten_list(l):
    assert isinstance(l, list), "not a list"
    outl = []
    splits = []
    for sl in l:
        assert isinstance(sl, list), "sub content not a list"
        splits.append(len(sl))
        outl += sl
    return outl, splits


def restore_flatten_list(l, splits):
    outl = []
    for split in splits:
        assert len(l) >= split, "list length invalid"
        sl, l = l[:split], l[split:]
        outl.append(sl)
    return outl


def extract_args(func):
    if hasattr(inspect, 'getfullargspec'):
        return inspect.getfullargspec(func)[0]
    else:
        return inspect.getargspec(func)[0]


def _all_gather(x, nranks, ring_id=0, use_calc_stream=True):
    return collective._c_allgather(
        x, nranks, ring_id=ring_id, use_calc_stream=use_calc_stream)


def wait_server_ready(endpoints):
    assert not isinstance(endpoints, six.string_types)
    while True:
        all_ok = True
        not_ready_endpoints = []
        for ep in endpoints:
            ip_port = ep.split(":")
            with contextlib.closing(
                    socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
                sock.settimeout(2)
                result = sock.connect_ex((ip_port[0], int(ip_port[1])))
                if result != 0:
                    all_ok = False
                    not_ready_endpoints.append(ep)
        if not all_ok:
            time.sleep(3)
        else:
            break


def init_communicator(program, rank, nranks, wait_port, current_endpoint,
                      endpoints):
    if nranks < 2:
        return
    other_endpoints = endpoints[:]
    other_endpoints.remove(current_endpoint)
    if rank == 0 and wait_port:
        wait_server_ready(other_endpoints)
    block = program.global_block()
    nccl_id_var = block.create_var(
        name=fluid.unique_name.generate('nccl_id'),
        persistable=True,
        type=fluid.core.VarDesc.VarType.RAW)

    block.append_op(
        type='c_gen_nccl_id',
        inputs={},
        outputs={'Out': nccl_id_var},
        attrs={
            'rank': rank,
            'endpoint': current_endpoint,
            'other_endpoints': other_endpoints
        })

    block.append_op(
        type='c_comm_init',
        inputs={'X': nccl_id_var},
        outputs={},
        attrs={
            'nranks': nranks,
            'rank': rank,
            'ring_id': 0,
        })


def prepare_distributed_context(place=None):
    if place is None:
        place = fluid.CUDAPlace(ParallelEnv().dev_id) if ParallelEnv().nranks > 1 \
            else fluid.CUDAPlace(0)

    strategy = fluid.dygraph.parallel.ParallelStrategy()
    strategy.nranks = ParallelEnv().nranks
    strategy.local_rank = ParallelEnv().local_rank
    strategy.trainer_endpoints = ParallelEnv().trainer_endpoints
    strategy.current_endpoint = ParallelEnv().current_endpoint

    if strategy.nranks < 2:
        return

    global _parallel_context_initialized

    if not _parallel_context_initialized and isinstance(place, fluid.CUDAPlace):

        def _init_context():
            communicator_prog = fluid.Program()
            init_communicator(communicator_prog, strategy.local_rank,
                              strategy.nranks, True, strategy.current_endpoint,
                              strategy.trainer_endpoints)
            exe = fluid.Executor(place)
            exe.run(communicator_prog)

        if fluid.in_dygraph_mode():
            fluid.disable_dygraph()
            _init_context()
            fluid.enable_dygraph(place)
        else:
            _init_context()

    else:
        assert ("Only support CUDAPlace for now.")

    _parallel_context_initialized = True
    return strategy
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256


class StaticGraphAdapter(object):
    """
    Model traning/inference with a static graph.
    """

    def __init__(self, model):
        super(StaticGraphAdapter, self).__init__()
        self.model = model
        # with `_build_once` gone, parameters are now created in `__init__`
        # so we need to keep track of the parameters already created
        self._startup_prog = fluid.default_startup_program()
        self._orig_prog = fluid.default_main_program()

        self._label_vars = {}  # label variables
        self._input_vars = {}  # label variables
        self._endpoints = {}
        self._loss_endpoint = None
        self._executor = None
        self._progs = {}
        self._compiled_progs = {}

        self._merge_count = {
            'eval_total': 0,
            'test_total': 0,
            'eval_batch': 0,
            'test_batch': 0
        }

        self._nranks = ParallelEnv().nranks
        self._local_rank = ParallelEnv().local_rank

    @property
    def mode(self):
        return self.model.mode

    @mode.setter
    def mode(self, value):
        self.model.mode = value

    def train_batch(self, inputs, labels=None):
        assert self.model._optimizer, \
            "model not ready, please call `model.prepare()` first"
        self.mode = 'train'
        return self._run(inputs, labels)

    def eval_batch(self, inputs, labels=None):
        self.mode = 'eval'
        return self._run(inputs, labels)

    def test_batch(self, inputs):
        self.mode = 'test'
        return self._run(inputs, None)

    def parameters(self, *args, **kwargs):
257
        return self.model.network.parameters(*args, **kwargs)
258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275

    def save(self, path):
        def _save(state, path):
            if not state:
                return
            state = {
                k: to_numpy(v) if isinstance(v, Variable) else v
                for k, v in state.items()
            }
            with open(path, 'wb') as f:
                pickle.dump(state, f)

        base = os.path.basename(path)
        assert base != "", "path should be of 'dirname/filename' format"
        dir_name = os.path.dirname(path)
        if dir_name and not os.path.exists(dir_name):
            os.makedirs(dir_name)
        param_path = path + ".pdparams"
276
        _save(self.model.network.state_dict(), param_path)
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447
        prog = self._progs.get('train', None)
        if prog is None or self.model._optimizer is None:
            return
        # XXX `optimizer.state_dict()` only work in dygraph mode
        optim_path = path + ".pdopt"
        optim = {
            p.name: p
            for p in filter(is_belong_to_optimizer, prog.list_vars())
        }
        if not optim:
            return

        _save(optim, optim_path)

    def load(self, param_state_pairs, optim_state):
        if self._executor is None:
            executor = fluid.Executor(fluid.CPUPlace())._default_executor
        else:
            executor = self._executor._default_executor

        # restore parameter states
        fluid.core._create_loaded_parameter(
            [param for param, state in param_state_pairs],
            global_scope(), executor)
        for param, state in param_state_pairs:
            self._set_var(param, state)

        # restore optimizer states
        # FIXME what if a different optimizer is used?
        if not self.model._optimizer or not optim_state:
            return
        self._load_optimizer(optim_state, executor)

    def _load_optimizer(self, state, executor):
        prog = self._progs.get('train', None)
        optim = list(filter(is_belong_to_optimizer, prog.list_vars()))
        if not optim:
            return

        fluid.core._create_loaded_parameter(optim, global_scope(), executor)

        converted_state = dict(state)
        for var in optim:
            if var.name in ["@LR_DECAY_COUNTER@", "global_step"]:
                # When using learning rate scheduler, dygraph would name the
                # global step var as "global_step" to save, while static-graph
                # would has a state var named as "@LR_DECAY_COUNTER@".
                # NOTE: dygraph saved global_step is 1 larger than that in
                # static-graph, since the time of global_step to increase is
                # different.
                state_val = (
                    np.array(converted_state.pop("global_step")) - 1
                ) if "global_step" in converted_state else converted_state.pop(
                    "@LR_DECAY_COUNTER@", None)
                if state_val is not None:
                    converted_state[var.name] = state_val
            elif var.name.startswith("learning_rate_"):
                # When using static learning rate, static-graph would make it
                # a persistable var named 'unique_name.generate("learning_rate")',
                # However, dygraph wouldn't save it.
                if var.name not in state:
                    continue
            else:
                # moment and other accumulators
                if var.name not in converted_state:
                    # try to convert from dygraph name
                    opt_name = self.model._optimizer._name
                    opt_cls_name = self.model._optimizer.__class__.__name__
                    opt_unq_name = None
                    for name in self.model._optimizer._accumulators.keys():
                        accum_name = name if opt_name is None else name[len(
                            opt_name) + 1:]
                        for param_name, state_var in self.model._optimizer._accumulators[
                                name].items():
                            if opt_unq_name is None:
                                # can not infer out the exact unique(opt_name),
                                # thus try to extract rather than generate
                                for state_key in sorted(
                                        state.keys(),
                                        key=lambda x: len(x),
                                        reverse=True):
                                    prefix = param_name + "_" + (
                                        opt_cls_name
                                        if opt_name is None else opt_name) + "_"
                                    if state_key.startswith(prefix):
                                        prefix_offset = state_key[len(
                                            prefix):].find("_") + len(prefix)
                                        opt_unq_name = state_key[len(
                                            param_name + "_"):prefix_offset]
                                        # TODO: assert
                                        # assert opt_unq_name is None
                                    # gen(param.name + "_" + gen(opt_name) + "_" + accum_name)
                                    # always end with "_0" since the unique optimizer._name
                            dy_state_name = (param_name + "_" + opt_unq_name +
                                             "_" + accum_name + "_0")
                            converted_state[
                                state_var.name] = converted_state.pop(
                                    dy_state_name)

            assert var.name in converted_state, \
                "variable [{}] is not in optimizer state file".format(var.name)
            self._set_var(var, converted_state[var.name])

    def _set_var(self, var, ndarray):
        t = global_scope().find_var(var.name).get_tensor()
        p = t._place()
        if p.is_cpu_place():
            place = fluid.CPUPlace()
        elif p.is_cuda_pinned_place():
            place = fluid.CUDAPinnedPlace()
        else:
            p = fluid.core.Place()
            p.set_place(t._place())
            place = fluid.CUDAPlace(p.gpu_device_id())

        t.set(ndarray, place)

    def _run(self, inputs, labels=None):
        compiled_prog = self._compiled_progs.get(self.mode, None)
        assert compiled_prog, \
            "Model is not ready, please call `model.prepare()` first"

        inputs = to_list(inputs)
        if labels is not None:
            labels = to_list(labels)
        assert len(inputs) == len(self._input_vars[self.mode]), \
            "number of inputs" \
            + " does not match number of arguments of `forward` method"

        feed = {}
        input_names = [v.name for v in self._input_vars[self.mode]]
        for idx, n in enumerate(input_names):
            # train and test may take different arguments
            if inputs[idx] is not None:
                feed[n] = inputs[idx]
        if labels is not None:
            for idx, v in enumerate(self._label_vars[self.mode]):
                feed[v.name] = labels[idx]

        endpoints = self._endpoints[self.mode]
        if self.mode == 'test':
            fetch_list = endpoints['output']
        else:
            metric_list, metric_splits = flatten_list(endpoints['metric'])
            fetch_list = endpoints['loss'] + metric_list
            num_loss = len(endpoints['loss'])

        # if fetch Variable is same as input Variable, do not fetch
        # from program, get it from input directly
        pruned_fetch_list = []
        pruned_fetch_idx_name_map = [""] * len(fetch_list)
        for i, fetch_var in enumerate(fetch_list):
            if fetch_var.name in feed.keys():
                pruned_fetch_idx_name_map[i] = fetch_var.name
            else:
                pruned_fetch_list.append(fetch_var)

        rets = self._executor.run(compiled_prog,
                                  feed=feed,
                                  fetch_list=pruned_fetch_list,
                                  return_numpy=False)

        # restore pruned fetch_list Variable from feeds
        for i, name in enumerate(pruned_fetch_idx_name_map):
            if len(name) > 0:
                rets.insert(i, feed[name])

        # LoDTensor cannot be fetch as numpy directly
        rets = [np.array(v) for v in rets]
        if self.mode == 'test':
            return rets[:]
448

449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471
        metric_states = restore_flatten_list(rets[num_loss:], metric_splits)
        metrics = []
        for metric, state in zip(self.model._metrics, metric_states):
            # cut off padding size
            if self.mode != 'train' and self.model._test_dataloader is not None \
                    and isinstance(self.model._test_dataloader, DataLoader) \
                    and self._nranks > 1:
                total_size = len(self.model._test_dataloader.dataset)
                # TODO: fixme if have better way to get batch size
                samples = state[0].shape[0]
                current_count = self._merge_count.get(self.mode + '_total', 0)
                if current_count + samples >= total_size:
                    state = [
                        s[:int(total_size - current_count), ...] for s in state
                    ]
                    self._merge_count[self.mode + '_total'] = 0
                    self._merge_count[self.mode + '_batch'] = int(total_size -
                                                                  current_count)
                else:
                    self._merge_count[self.mode + '_total'] += samples
                    self._merge_count[self.mode + '_batch'] = samples

            metrics.append(metric.update(*state))
472 473 474 475 476

        if num_loss and len(metrics):
            return rets[:num_loss], metrics
        else:
            return rets[:num_loss] if num_loss else metrics
477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507

    def prepare(self):
        modes = ['train', 'eval', 'test']
        for mode in modes:
            self._make_program(mode)
            self._compile_and_initialize(self._progs[mode], mode)

    def _make_program(self, mode):
        prog = self._progs.get(mode, None)
        if prog is not None:
            return

        prog = self._orig_prog.clone()
        # NOTE: When defining learning rate scheduling in static-graph, ops to
        # increase the global step var and calculate learning rate would be
        # prepended into _orig_prog. test program maked by `_orig_prog.clone`
        # also would include these ops. Thus must prune these ops in test
        # program, otherwise the global step would be changed in test.
        if mode != 'train':
            for op in list(prog.global_block().ops):
                prog.global_block()._remove_op(0)
        if mode == 'train' and self.model._optimizer \
                and self.model._optimizer._learning_rate_map:
            # HACK workaround learning rate map issue
            lr_var = self.model._optimizer._learning_rate_map[self._orig_prog]
            new_lr_var = prog.global_block().vars[lr_var.name]
            self.model._optimizer._learning_rate_map[prog] = new_lr_var

        losses = []
        metrics = []
        with fluid.program_guard(prog, self._startup_prog):
508 509
            inputs = self.model._inputs
            labels = self.model._labels if self.model._labels else []
510 511
            inputs = [k._create_feed_layer() for k in to_list(inputs)]
            labels = [k._create_feed_layer() for k in to_list(labels)]
512
            self._label_vars[mode] = labels
513
            outputs = to_list(self.model.network.forward(*inputs))
514

515 516
            if mode != 'test' and self.model._loss:
                losses = self.model._loss(*(outputs + labels))
517 518 519 520 521 522 523 524

            if self._nranks > 1 and mode != 'train':
                outputs = [_all_gather(o, self._nranks) for o in outputs]
                if mode != 'test':
                    labels = [_all_gather(l, self._nranks) for l in labels]

            if mode != 'test':
                for metric in self.model._metrics:
525
                    metrics.append(to_list(metric.compute(*(outputs + labels))))
526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547

            if mode == 'train' and self.model._optimizer:
                self._loss_endpoint = fluid.layers.sum(losses)
                if self._nranks > 1:
                    role = role_maker.PaddleCloudRoleMaker(is_collective=True)
                    fleet.init(role)
                    dist_strategy = DistributedStrategy()
                    dist_strategy.mode = "collective"
                    dist_strategy.collective_mode = "grad_allreduce"
                    self.model._optimizer = fleet.distributed_optimizer(
                        self.model._optimizer, strategy=dist_strategy)

                self.model._optimizer.minimize(self._loss_endpoint)

        if mode != 'train':  # clone again to put it in test mode
            prog = prog.clone(for_test=True)

        self._input_vars[mode] = inputs

        self._progs[mode] = prog
        self._endpoints[mode] = {
            "output": outputs,
548
            "loss": to_list(losses),
549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606
            "metric": metrics
        }

    def _compile_and_initialize(self, prog, mode):
        compiled_prog = self._compiled_progs.get(mode, None)
        if compiled_prog is not None:
            return compiled_prog

        assert self.model._place is not None, \
            "device is not set, please call `model.prepare()` first"

        place = self.model._place

        # XXX *ALL WEIGHTS* should be initialized upon model construction
        # even if `forward()` may run different code path for different mode
        # therefore startup program only needs to run once
        if self._executor is None:
            self._executor = fluid.Executor(place)
            # XXX incremental initialization
            uninitialized = []
            for var_py in self._startup_prog.list_vars():
                var = fluid.global_scope().find_var(var_py.name)
                if not var_py.name.startswith('nccl_id') and var and \
                        var.get_tensor()._is_initialized():
                    continue

                uninitialized.append(var_py)
            if uninitialized:
                startup_prog = self._startup_prog._prune(uninitialized)
                self._executor.run(startup_prog)

        if self._nranks < 2:
            compiled_prog = fluid.CompiledProgram(prog)
        else:
            compiled_prog = prog

        self._compiled_progs[mode] = compiled_prog


class DynamicGraphAdapter(object):
    def __init__(self, model):
        super(DynamicGraphAdapter, self).__init__()
        self.model = model
        self._nranks = ParallelEnv().nranks
        self._local_rank = ParallelEnv().local_rank
        self._merge_count = {
            'eval_total': 0,
            'test_total': 0,
            'eval_batch': 0,
            'test_batch': 0
        }

        if self._nranks > 1:
            stradegy = fluid.dygraph.parallel.ParallelStrategy()
            stradegy.nranks = ParallelEnv().nranks
            stradegy.local_rank = ParallelEnv().local_rank
            stradegy.trainer_endpoints = ParallelEnv().trainer_endpoints
            stradegy.current_endpoint = ParallelEnv().current_endpoint
607 608
            self.ddp_model = fluid.dygraph.parallel.DataParallel(
                self.model.network, stradegy)
609 610 611 612 613 614 615 616 617 618 619 620 621

    @property
    def mode(self):
        return self.model.mode

    @mode.setter
    def mode(self, value):
        self.model.mode = value

    # TODO multi device in dygraph mode not implemented at present time
    def train_batch(self, inputs, labels=None):
        assert self.model._optimizer, \
            "model not ready, please call `model.prepare()` first"
622
        self.model.network.train()
623 624
        self.mode = 'train'
        inputs = to_list(inputs)
625 626 627
        labels = labels or []
        labels = [to_variable(l) for l in to_list(labels)]

628 629
        if self._nranks > 1:
            outputs = self.ddp_model.forward(* [to_variable(x) for x in inputs])
630
            losses = self.model._loss(*(to_list(outputs) + labels))
631
            losses = to_list(losses)
632 633 634 635 636
            final_loss = fluid.layers.sum(losses)
            final_loss = self.ddp_model.scale_loss(final_loss)
            final_loss.backward()
            self.ddp_model.apply_collective_grads()
        else:
637 638
            outputs = self.model.network.forward(
                * [to_variable(x) for x in inputs])
639
            losses = self.model._loss(*(to_list(outputs) + labels))
640
            losses = to_list(losses)
641 642 643 644
            final_loss = fluid.layers.sum(losses)
            final_loss.backward()

        self.model._optimizer.minimize(final_loss)
645
        self.model.network.clear_gradients()
646 647
        metrics = []
        for metric in self.model._metrics:
648
            metric_outs = metric.compute(*(to_list(outputs) + labels))
649 650 651 652 653 654 655
            m = metric.update(* [to_numpy(m) for m in to_list(metric_outs)])
            metrics.append(m)

        return ([to_numpy(l) for l in losses], metrics) \
            if len(metrics) > 0 else [to_numpy(l) for l in losses]

    def eval_batch(self, inputs, labels=None):
656
        self.model.network.eval()
657 658
        self.mode = 'eval'
        inputs = to_list(inputs)
659 660 661
        labels = labels or []
        labels = [to_variable(l) for l in to_list(labels)]

662
        outputs = self.model.network.forward(* [to_variable(x) for x in inputs])
663 664
        if self.model._loss:
            losses = self.model._loss(*(to_list(outputs) + labels))
665 666
            losses = to_list(losses)

667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691
        if self._nranks > 1:
            outputs = [_all_gather(o, self._nranks) for o in to_list(outputs)]
            labels = [_all_gather(l, self._nranks) for l in labels]
        metrics = []
        for metric in self.model._metrics:
            # cut off padding value.
            if self.model._test_dataloader is not None and self._nranks > 1 \
                    and isinstance(self.model._test_dataloader, DataLoader):
                total_size = len(self.model._test_dataloader.dataset)
                samples = outputs[0].shape[0]
                current_count = self._merge_count.get(self.mode + '_total', 0)
                if current_count + samples >= total_size:
                    outputs = [
                        o[:int(total_size - current_count)] for o in outputs
                    ]
                    labels = [
                        l[:int(total_size - current_count)] for l in labels
                    ]
                    self._merge_count[self.mode + '_total'] = 0
                    self._merge_count[self.mode + '_batch'] = int(total_size -
                                                                  current_count)
                else:
                    self._merge_count[self.mode + '_total'] += samples
                    self._merge_count[self.mode + '_batch'] = samples

692
            metric_outs = metric.compute(*(to_list(outputs) + labels))
693 694 695
            m = metric.update(* [to_numpy(m) for m in to_list(metric_outs)])
            metrics.append(m)

696
        if self.model._loss and len(metrics):
697
            return [to_numpy(l) for l in losses], metrics
698
        elif self.model._loss:
699 700 701
            return [to_numpy(l) for l in losses]
        else:
            return metrics
702 703

    def test_batch(self, inputs):
704
        self.model.network.eval()
705 706
        self.mode = 'test'
        inputs = [to_variable(x) for x in to_list(inputs)]
707
        outputs = self.model.network.forward(*inputs)
708 709 710 711 712 713
        if self._nranks > 1 and isinstance(self.model._place, fluid.CUDAPlace):
            outputs = [_all_gather(o, self._nranks) for o in to_list(outputs)]

        return [to_numpy(o) for o in to_list(outputs)]

    def parameters(self, *args, **kwargs):
714
        return self.model.network.parameters(*args, **kwargs)
715 716

    def save(self, path):
717
        params = self.model.network.state_dict()
718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733
        fluid.save_dygraph(params, path)
        if self.model._optimizer is None:
            return
        if self.model._optimizer.state_dict():
            optim = self.model._optimizer.state_dict()
            fluid.save_dygraph(optim, path)

    def load(self, param_state_pairs, optim_state):
        # restore parameter states
        for param, state in param_state_pairs:
            param.set_value(state)

        # resotre optimizer states
        if not self.model._optimizer or not optim_state:
            return

734 735
        # If optimizer performs set_state_dict when state vars haven't been created,
        # which would happen when set_state_dict before minimize, the state would be
736 737 738 739 740 741 742 743 744 745 746
        # stored in optimizer._accumulators_holder and loaded lazily.
        # To contrive this when loading from static-graph saved states, extend
        # state dict to include keys named accoring to dygraph naming rules.
        # TODO: if len(self.model._optimizer._accumulators) > 0
        converted_state = dict(optim_state)
        opt_unq_name = self.model._optimizer._name
        if opt_unq_name is None:
            opt_unq_name = ''

        opt_cls_name = self.model._optimizer.__class__.__name__
        opt_name = opt_unq_name[:opt_unq_name.rfind("_")]  # remove suffix idx
747
        param_names = [param.name for param in self.model.network.parameters()]
748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778
        for var_name, state_var in sorted(
                optim_state.items(), key=lambda x: len(x[0]), reverse=True):
            if var_name in ["@LR_DECAY_COUNTER@", "global_step"]:
                # NOTE: dygraph saved global_step is 1 larger than that in
                # static-graph, since the time of global_step to increase is
                # different.
                if var_name == "@LR_DECAY_COUNTER@":
                    converted_state["global_step"] = np.array(
                        converted_state.pop("@LR_DECAY_COUNTER@")) + 1
            else:
                # moment and other accumulators
                # extend state dict to include promising dygraph names
                for param_name in param_names:
                    if var_name.startswith(param_name + "_" + opt_name):
                        # when init optimizer with name
                        accum_name = var_name[len(param_name + "_" + opt_name +
                                                  "_"):]
                    elif var_name.startswith(param_name +
                                             "_") and opt_name == opt_cls_name:
                        # when init optimizer without name
                        accum_name = var_name[len(param_name + "_"):]
                    else:
                        continue
                    # remove suffix idx
                    accum_name = accum_name[:accum_name.rfind("_")]
                    # state names always end with "_0" in dygraph because of the
                    # unique optimizer._name
                    dy_state_name = (param_name + "_" + opt_unq_name + "_" +
                                     accum_name + "_0")
                    converted_state[dy_state_name] = state_var

779 780 781 782 783 784 785
        if not hasattr(self.model._optimizer, 'set_state_dict'):
            warnings.warn(
                "paddle.fluid.optimizer is deprecated in API 2.0, please use paddle.optimizer instead"
            )
            self.model._optimizer.set_dict(converted_state)
        else:
            self.model._optimizer.set_state_dict(converted_state)
786 787


788
class Model(object):
789 790 791
    """
    An Model object is network with training and inference features.
    Dynamic graph and static graph are supported at the same time,
792
    switched by `paddle.disable_static()`. The usage is as follows.
793
    But note, the switching between dynamic and static should be before
794
    instantiating a Model. The input description, i.e, paddle.static.InputSpec,
795 796
    must be required for static graph.

797
    Args:
798 799
        network (paddle.nn.Layer): The network is an instance of
            paddle.nn.Layer.
800 801 802
        inputs (InputSpec|list|dict|None): `inputs`, entry points of network,
            could be a InputSpec instance, or lits of InputSpec instances,
            or dict ({name: InputSpec}), or None. For static graph,
803
            inputs must be set. For dynamic graph, it could be None.
804 805 806
        labels (InputSpec|list|None): `labels`, entry points of network,
            could be a InputSpec instnace or lits of InputSpec instances,
            or None. For static graph, if labels is required in loss,
807 808 809
            labels must be set. Otherwise, it could be None.


810
    Examples:
811 812
        .. code-block:: python

813
        import paddle
814 815 816 817
        import paddle.nn as nn
        from paddle.static import InputSpec

        device = paddle.set_device('cpu') # or 'gpu'
818
        # if use static graph, do not set
819
        paddle.disable_static(device)
820 821 822 823 824 825

        net = nn.Sequential(
            nn.Linear(784, 200),
            nn.Tanh(),
            nn.Linear(200, 10))

826
        # inputs and labels are not required for dynamic graph.
827 828
        input = InputSpec([None, 784], 'float32', 'x')
        label = InputSpec([None, 1], 'int64', 'label')
829
        
830
        model = paddle.Model(net, input, label)
831
        optim = paddle.optimizer.SGD(learning_rate=1e-3,
832
            parameters=model.parameters())
833
        model.prepare(optim,
834
                      paddle.nn.CrossEntropyLoss(),
835
                      paddle.metric.Accuracy())
836
        
837 838
        data = paddle.vision.datasets.MNIST(mode='train', chw_format=False)
        model.fit(data, epochs=2, batch_size=32, verbose=1)
839 840
    """

841
    def __init__(self, network, inputs=None, labels=None):
842
        self.mode = 'train'
843
        self.network = network
844 845
        self._inputs = None
        self._labels = None
846
        self._loss = None
847 848 849 850 851
        self._loss_weights = None
        self._optimizer = None
        self._optimizer = None
        self._test_dataloader = None

852 853 854 855
        if not in_dygraph_mode():
            if not isinstance(inputs, (list, dict, Input)):
                raise TypeError(
                    "'inputs' must be list or dict in static graph mode")
856 857
        self._inputs = self._verify_spec(inputs, True)
        self._labels = self._verify_spec(labels)
858

859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884
        # init backend
        if fluid.in_dygraph_mode():
            self._adapter = DynamicGraphAdapter(self)
        else:
            self._adapter = StaticGraphAdapter(self)

    def train_batch(self, inputs, labels=None):
        """
        Run one training step on a batch of data.

        Args:
            inputs (list): A list of numpy.ndarray, each is a batch of
                input data.
            labels (list): A list of numpy.ndarray, each is a batch of
                input label. If has no labels, set None. Default is None.

        Returns:
            A list of scalar training loss if the model has no metrics,
            or a tuple (list of scalar loss, list of metrics) if the model
            set metrics.

        Examples:

            .. code-block:: python
            
              import numpy as np
885
              import paddle
886 887
              import paddle.nn as nn
              from paddle.static import InputSpec
888

889
              device = paddle.set_device('cpu') # or 'gpu'
890
              paddle.disable_static(device)
891

892 893 894 895 896 897 898 899
              net = nn.Sequential(
                  nn.Linear(784, 200),
                  nn.Tanh(),
                  nn.Linear(200, 10))

              input = InputSpec([None, 784], 'float32', 'x')
              label = InputSpec([None, 1], 'int64', 'label')
              model = paddle.Model(net, input, label)
900
              optim = paddle.optimizer.SGD(learning_rate=1e-3,
901
                  parameters=model.parameters())
902
              model.prepare(optim, paddle.nn.CrossEntropyLoss())
903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929
              data = np.random.random(size=(4,784)).astype(np.float32)
              label = np.random.randint(0, 10, size=(4, 1)).astype(np.int64)
              loss = model.train_batch([data], [label])
              print(loss)
        """
        return self._adapter.train_batch(inputs, labels)

    def eval_batch(self, inputs, labels=None):
        """
        Run one evaluating step on a batch of data.

        Args:
            inputs (list): A list of numpy.ndarray, each is a batch of
                input data.
            labels (list): A list of numpy.ndarray, each is a batch of
                input label. If has no labels, set None. Default is None.

        Returns:
            A list of scalar testing loss if the model has no metrics,
            or a tuple (list of scalar loss, list of metrics) if the model
            set metrics.

        Examples:

            .. code-block:: python
            
              import numpy as np
930
              import paddle
931 932
              import paddle.nn as nn
              from paddle.static import InputSpec
933

934
              device = paddle.set_device('cpu') # or 'gpu'
935
              paddle.disable_static(device)
936

937 938 939 940 941 942 943 944
              net = nn.Sequential(
                  nn.Linear(784, 200),
                  nn.Tanh(),
                  nn.Linear(200, 10))

              input = InputSpec([None, 784], 'float32', 'x')
              label = InputSpec([None, 1], 'int64', 'label')
              model = paddle.Model(net, input, label)
945
              optim = paddle.optimizer.SGD(learning_rate=1e-3,
946
                  parameters=model.parameters())
947
              model.prepare(optim,
948
                            paddle.nn.CrossEntropyLoss())
949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972
              data = np.random.random(size=(4,784)).astype(np.float32)
              label = np.random.randint(0, 10, size=(4, 1)).astype(np.int64)
              loss = model.eval_batch([data], [label])
              print(loss)
        """
        return self._adapter.eval_batch(inputs, labels)

    def test_batch(self, inputs):
        """
        Run one testing step on a batch of data.

        Args:
            inputs (list): A list of numpy.ndarray, each is a batch of
                input data.

        Returns:
            A list of numpy.ndarray of predictions, that is the outputs
            of Model forward.

        Examples:

            .. code-block:: python
            
              import numpy as np
973
              import paddle
974
              import paddle.nn as nn
975

976
              device = paddle.set_device('cpu') # or 'gpu'
977
              paddle.disable_static(device)
978

979 980 981 982 983 984 985
              net = nn.Sequential(
                  nn.Linear(784, 200),
                  nn.Tanh(),
                  nn.Linear(200, 10),
                  nn.Softmax())

              model = paddle.Model(net)
986
              model.prepare()
987
              data = np.random.random(size=(4,784)).astype(np.float32)
988
              out = model.test_batch([data])
989 990 991 992
              print(out)
        """
        return self._adapter.test_batch(inputs)

993 994 995 996 997
    def save(self, path, training=True):
        """  
        This function saves parameters, optimizer information or model and 
        paramters only for inference to path. It depends on the parameter
        `training`.
998

999 1000
        If `training` is set to True, the parameters saved contain all 
        the trainable Variable, will save to a file with suffix ".pdparams".
1001 1002 1003 1004
        The optimizer information contains all the variable used by optimizer.
        For Adam optimizer, contains beta1, beta2, momentum etc. All the
        information will save to a file with suffix ".pdopt". (If the optimizer
        have no variable need to save (like SGD), the fill will not generated).
1005
        This function will silently overwrite existing file at the target location.
1006

1007 1008 1009 1010 1011
        If `training` is set to False, only inference model will be saved. It 
        should be noted that before using `save`, you should run the model, and 
        the shape of input you saved is as same as the input of its running.
        `@paddle.jit.to_static` must be added on `forward` function of your layer 
        in dynamic mode now and these will be optimized later.
1012 1013 1014 1015 1016

        Args:
            path (str): The file prefix to save model. The format is
                'dirname/file_prefix' or 'file_prefix'. if empty str. A exception
                 will be raised.
1017 1018
            training (bool, optional): Whether to save for training. If not, save
                for inference only. Default: True.
1019 1020 1021 1022 1023 1024 1025

        Returns:
            None

        Examples:

            .. code-block:: python
1026

1027
                import paddle
1028 1029
                import paddle.nn as nn
                from paddle.static import InputSpec
1030

1031
                class Mnist(nn.Layer):
1032
                    def __init__(self):
1033
                        super(Mnist, self).__init__()
1034 1035 1036 1037 1038
                        self.net = nn.Sequential(
                            nn.Linear(784, 200),
                            nn.Tanh(),
                            nn.Linear(200, 10),
                            nn.Softmax())
1039

1040 1041 1042
                    # If save for inference in dygraph, need this
                    @paddle.jit.to_static
                    def forward(self, x):
1043
                        return self.net(x)
1044

1045
                dynamic = True  # False
1046
                device = paddle.set_device('cpu')
1047 1048 1049
                # if use static graph, do not set
                paddle.disable_static(device) if dynamic else None
                # inputs and labels are not required for dynamic graph.
1050 1051 1052
                input = InputSpec([None, 784], 'float32', 'x')
                label = InputSpec([None, 1], 'int64', 'label')
                model = paddle.Model(Mnist(), input, label)
1053
                optim = paddle.optimizer.SGD(learning_rate=1e-3,
1054
                    parameters=model.parameters())
1055
                model.prepare(optim, paddle.nn.CrossEntropyLoss())
1056 1057
                data = paddle.vision.datasets.MNIST(mode='train', chw_format=False)
                model.fit(data, epochs=1, batch_size=32, verbose=0)
1058 1059
                model.save('checkpoint/test')  # save for training
                model.save('inference_model', False)  # save for inference
1060
        """
1061

1062
        if ParallelEnv().local_rank == 0:
1063 1064 1065 1066
            if not training:
                self._save_inference_model(path)
            else:
                self._adapter.save(path)
1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100

    def load(self, path, skip_mismatch=False, reset_optimizer=False):
        """
        Load from files storing the model states and optimizer states. The file
        for optimizer states is not necessary if no need to restore the optimizer.

        NOTE: parameters are retrieved out from the file storing model states
        accoring to their structured names.

        For fine-tuning or transfer-learning models where some of the layers have
        changed, keep parameters needed to restore have same structured names in
        the pre-trained model and fine-tuning model.

        Args:
            path (str): The prefix of files storing the model states and
                optimizer states. The files would be `path.pdparams` and
                `path.pdopt` separately, and the latter is not necessary
                when no need to restore.
            skip_mismatch (bool): Whether to skip the loading of mismatch
                parameter or raise an error when mismatch happens (not found
                the parameter in file storing model states of or receives a
                mismatch shape).
            reset_optimizer (bool): If True, ignore the providing file storing
                optimizer states and initialize optimizer states from scratch.
                Otherwise, restore optimizer states from `path.pdopt` if
                a optimizer has been set to the model. Default False.

        Returns:
            None

        Examples:

            .. code-block:: python
            
1101
              import paddle
1102
              import paddle.nn as nn
1103
              
1104
              device = paddle.set_device('cpu')
1105
              paddle.disable_static(device)
1106 1107 1108 1109 1110 1111 1112

              model = paddle.Model(nn.Sequential(
                  nn.Linear(784, 200),
                  nn.Tanh(),
                  nn.Linear(200, 10),
                  nn.Softmax()))
              model.save('checkpoint/test')
1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144
              model.load('checkpoint/test')
        """

        def _load_state_from_path(path):
            if not os.path.exists(path):
                return
            with open(path, 'rb') as f:
                return pickle.load(f) if six.PY2 else pickle.load(
                    f, encoding='latin1')

        def _check_match(key, param):
            state = param_state.get(key, None)
            if state is None:
                raise ValueError(
                    "{} is not found in the providing file.".format(key))
            if list(state.shape) != list(param.shape):
                raise ValueError(
                    "{} receives a shape {}, but the expected shape is {}.".
                    format(key, list(state.shape), list(param.shape)))
            return param, state

        def _strip_postfix(path):
            path, ext = os.path.splitext(path)
            assert ext in ['', '.pdparams', '.pdopt', '.pdmodel'], \
                    "Unknown postfix {} from weights".format(ext)
            return path

        path = _strip_postfix(path)
        param_state = _load_state_from_path(path + ".pdparams")
        assert param_state, "Failed to load parameters, please check path."

        matched_param_state = []
1145
        for key, param in self.network.state_dict().items():
1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173
            try:
                match_res = _check_match(key, param)
            except ValueError as err:
                if skip_mismatch:
                    warnings.warn(
                        ("Skip loading for {}. ".format(key) + str(err)))
                    # reset optimizer when mismatch happens
                    reset_optimizer = True
                else:
                    raise err
            matched_param_state.append(match_res)

        optim_state = None if reset_optimizer else _load_state_from_path(
            path + ".pdopt")
        return self._adapter.load(matched_param_state, optim_state)

    def parameters(self, *args, **kwargs):
        """
        Returns a list of parameters of the model.

        Returns:
            A list of Parameter in static graph.
            A list of ParamBase in dynamic graph.

        Examples:

            .. code-block:: python

1174
              import paddle
1175
              import paddle.nn as nn
1176

1177
              paddle.disable_static()
1178 1179 1180 1181 1182

              model = paddle.Model(nn.Sequential(
                  nn.Linear(784, 200),
                  nn.Tanh(),
                  nn.Linear(200, 10)))
1183 1184 1185 1186
              params = model.parameters()
        """
        return self._adapter.parameters()

1187
    def prepare(self, optimizer=None, loss=None, metrics=None):
1188 1189 1190 1191 1192 1193 1194
        """
        Configures the model before runing.

        Args:
            optimizer (Optimizer|None): Optimizer must be set in training
                and should be a Optimizer instance. It can be None in eval
                and test mode.
1195 1196
            loss (Loss|callable function|None): Loss function can
                be a `paddle.nn.Layer` instance or any callable function
1197 1198
                taken the predicted values and ground truth values as input.
                It can be None when there is no loss.
1199 1200 1201 1202 1203 1204 1205
            metrics (Metric|list of Metric|None): If metrics is set, all
                metrics will be calculated and output in train/eval mode.

        Returns:
            None
        """

1206 1207
        self._place = _get_device()
        if isinstance(self._place, fluid.CUDAPlace):
1208 1209 1210 1211 1212 1213 1214
            global _parallel_context_initialized
            if ParallelEnv().nranks > 1 and not _parallel_context_initialized:
                if fluid.in_dygraph_mode():
                    main_prog_seed = fluid.default_main_program().random_seed
                    startup_prog_seed = fluid.default_startup_program(
                    ).random_seed
                    fluid.disable_dygraph()
1215
                    paddle.disable_static(self._place)
1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226
                    # enable_dygraph would create and switch to a new program,
                    # thus also copy seed to the new program
                    fluid.default_main_program().random_seed = main_prog_seed
                    fluid.default_startup_program(
                    ).random_seed = startup_prog_seed
                    fluid.dygraph.parallel.prepare_context()
                else:
                    prepare_distributed_context(self._place)
                _parallel_context_initialized = True

        self._optimizer = optimizer
1227 1228 1229 1230 1231
        if loss is not None:
            if not isinstance(loss, paddle.nn.Layer) and not callable(loss):
                raise TypeError("'loss' must be sub classes of " \
                    "`paddle.nn.Layer` or any callable function.")
        self._loss = loss
1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309

        metrics = metrics or []
        for metric in to_list(metrics):
            assert isinstance(metric, Metric), \
                "{} is not sub class of Metric".format(
                    metric.__class__.__name__)
        self._metrics = to_list(metrics)

        if not in_dygraph_mode():
            self._adapter.prepare()

    def fit(
            self,
            train_data=None,
            eval_data=None,
            batch_size=1,
            epochs=1,
            eval_freq=1,
            log_freq=10,
            save_dir=None,
            save_freq=1,
            verbose=2,
            drop_last=False,
            shuffle=True,
            num_workers=0,
            callbacks=None, ):
        """
        Trains the model for a fixed number of epochs. If `eval_data` is set,
        evaluation will be done at the end of each epoch.

        Args:
            train_data (Dataset|DataLoader): An iterable data loader is used for 
                train. An instance of paddle paddle.io.Dataset or 
                paddle.io.Dataloader is recomended. Default: None.
            eval_data (Dataset|DataLoader): An iterable data loader is used for
                evaluation at the end of epoch. If None, will not do evaluation. 
                An instance of paddle.io.Dataset or paddle.io.Dataloader 
                is recomended. Default: None.
            batch_size (int): Integer number. The batch size of train_data
                and eval_data. When train_data and eval_data are both the
                instance of Dataloader, this parameter will be ignored.
                Default: 1.
            epochs (int): Integer number. The number of epochs to train
                the model. Default: 1.
            eval_freq (int): The frequency, in number of epochs, an evalutation
                is performed. Default: 1.
            log_freq (int): The frequency, in number of steps, the training logs
                are printed. Default: 10.
            save_dir(str|None): The directory to save checkpoint during training.
                If None, will not save checkpoint. Default: None.
            save_freq (int): The frequency, in number of epochs, to save
                checkpoint. Default: 1.
            verbose (int): The verbosity mode, should be 0, 1, or 2. 0 = silent,
                1 = progress bar, 2 = one line per epoch. Default: 2.
            drop_last (bool): Whether drop the last incomplete batch of
                train_data when dataset size is not divisible by the batch size.
                When train_data is an instance of Dataloader, this parameter
                will be ignored. Default: False.
            shuffle (bool): Whther to shuffle train_data. When train_data is
                an instance of Dataloader, this parameter will be ignored.
                Default: True.
            num_workers (int): The number of subprocess to load data, 0 for no
                subprocess used and loading data in main process.
                When train_data and eval_data are both the instance of
                Dataloader, this parameter will be ignored. Default: 0.
            callbacks (Callback|None): A list of `Callback` instances to apply
                during training. If None, `ProgBarLogger` and `ModelCheckpoint`
                are automatically inserted. Default: None.

        Returns:
            None

        Examples:
            1. An example use Dataset and set btch size, shuffle in fit.
               How to make a batch is done internally.

            .. code-block:: python

1310
              import paddle
1311
              from paddle.static import InputSpec
1312 1313

              dynamic = True
1314
              device = paddle.set_device('cpu') # or 'gpu'
1315
              paddle.disable_static(device) if dynamic else None
1316
           
1317 1318
              train_dataset = paddle.vision.datasets.MNIST(mode='train')
              val_dataset = paddle.vision.datasets.MNIST(mode='test')
1319
           
1320 1321
              input = InputSpec([None, 1, 28, 28], 'float32', 'image')
              label = InputSpec([None, 1], 'int64', 'label')
1322
           
1323 1324
              model = paddle.Model(
                  paddle.vision.models.LeNet(classifier_activation=None),
1325
                  input, label)
1326 1327
              optim = paddle.optimizer.Adam(
                  learning_rate=0.001, parameters=model.parameters())
1328 1329
              model.prepare(
                  optim,
1330
                  paddle.nn.CrossEntropyLoss(),
1331
                  paddle.metric.Accuracy(topk=(1, 2)))
1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342
              model.fit(train_dataset,
                        val_dataset,
                        epochs=2,
                        batch_size=64,
                        save_dir='mnist_checkpoint')

            2. An example use DataLoader, batch size and shuffle is set in
               DataLoader.

            .. code-block:: python

1343
              import paddle
1344
              from paddle.static import InputSpec
1345 1346

              dynamic = True
1347
              device = paddle.set_device('cpu') # or 'gpu'
1348
              paddle.disable_static(device) if dynamic else None
1349
           
1350
              train_dataset = paddle.vision.datasets.MNIST(mode='train')
1351
              train_loader = paddle.io.DataLoader(train_dataset,
1352
                  places=device, batch_size=64)
1353
              val_dataset = paddle.vision.datasets.MNIST(mode='test')
1354
              val_loader = paddle.io.DataLoader(val_dataset,
1355 1356
                  places=device, batch_size=64)
           
1357 1358
              input = InputSpec([None, 1, 28, 28], 'float32', 'image')
              label = InputSpec([None, 1], 'int64', 'label')
1359
           
1360 1361
              model = paddle.Model(
                  paddle.vision.models.LeNet(classifier_activation=None), input, label)
1362 1363
              optim = paddle.optimizer.Adam(
                  learning_rate=0.001, parameters=model.parameters())
1364 1365
              model.prepare(
                  optim,
1366
                  paddle.nn.CrossEntropyLoss(),
1367
                  paddle.metric.Accuracy(topk=(1, 2)))
1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478
              model.fit(train_loader,
                        val_loader,
                        epochs=2,
                        save_dir='mnist_checkpoint')
        """

        assert train_data is not None, \
                "train_data must be given!"

        if isinstance(train_data, Dataset):
            train_sampler = DistributedBatchSampler(
                train_data,
                batch_size=batch_size,
                shuffle=shuffle,
                drop_last=drop_last)
            train_loader = DataLoader(
                train_data,
                batch_sampler=train_sampler,
                places=self._place,
                num_workers=num_workers,
                return_list=True)
        else:
            train_loader = train_data

        if eval_data is not None and isinstance(eval_data, Dataset):
            eval_sampler = DistributedBatchSampler(
                eval_data, batch_size=batch_size)
            eval_loader = DataLoader(
                eval_data,
                batch_sampler=eval_sampler,
                places=self._place,
                num_workers=num_workers,
                return_list=True)
        elif eval_data is not None:
            eval_loader = eval_data
        else:
            eval_loader = None

        do_eval = eval_loader is not None
        self._test_dataloader = eval_loader

        steps = self._len_data_loader(train_loader)
        cbks = config_callbacks(
            callbacks,
            model=self,
            epochs=epochs,
            steps=steps,
            log_freq=log_freq,
            save_freq=save_freq,
            save_dir=save_dir,
            verbose=verbose,
            metrics=self._metrics_name(), )

        cbks.on_begin('train')
        for epoch in range(epochs):

            cbks.on_epoch_begin(epoch)
            logs = self._run_one_epoch(train_loader, cbks, 'train')
            cbks.on_epoch_end(epoch, logs)

            if do_eval and epoch % eval_freq == 0:

                eval_steps = self._len_data_loader(eval_loader)
                cbks.on_begin('eval', {
                    'steps': eval_steps,
                    'metrics': self._metrics_name()
                })

                eval_logs = self._run_one_epoch(eval_loader, cbks, 'eval')

                cbks.on_end('eval', eval_logs)

        cbks.on_end('train', logs)
        self._test_dataloader = None

    def evaluate(
            self,
            eval_data,
            batch_size=1,
            log_freq=10,
            verbose=2,
            num_workers=0,
            callbacks=None, ):
        """
        Evaluate the loss and metrics of the model on input dataset.

        Args:
            eval_data (Dataset|DataLoader): An iterable data loader is used for
                evaluation. An instance of paddle.io.Dataset or 
                paddle.io.Dataloader is recomended.
            batch_size (int): Integer number. The batch size of train_data
                and eval_data.  When eval_data is the instance of Dataloader,
                this argument will be ignored. Default: 1.
            log_freq (int): The frequency, in number of steps, the eval logs
                are printed. Default: 10.
            verbose (int): The verbosity mode, should be 0, 1, or 2. 0 = silent,
                1 = progress bar, 2 = one line per epoch. Default: 2.
            num_workers (int): The number of subprocess to load data,
                0 for no subprocess used and loading data in main process. When
                train_data and eval_data are both the instance of Dataloader,
                this parameter will be ignored. Default: 0.
            callbacks (Callback|None): A list of `Callback` instances to apply
                during training. If None, `ProgBarLogger` and `ModelCheckpoint`
                are automatically inserted. Default: None.
        Returns:
            dict: Result of metric. The key is the names of Metric,
                value is a scalar or numpy.array.

        Examples:
        .. code-block:: python

1479
            import paddle
1480
            from paddle.static import InputSpec
1481

1482
            # declarative mode
1483
            val_dataset = paddle.vision.datasets.MNIST(mode='test')
1484

1485 1486 1487
            input = InputSpec([-1, 1, 28, 28], 'float32', 'image')
            label = InputSpec([None, 1], 'int64', 'label')
            model = paddle.Model(paddle.vision.models.LeNet(), input, label)
1488
            model.prepare(metrics=paddle.metric.Accuracy())
1489 1490 1491 1492
            result = model.evaluate(val_dataset, batch_size=64)
            print(result)

            # imperative mode
1493
            paddle.disable_static()
1494
            model = paddle.Model(paddle.vision.models.LeNet())
1495
            model.prepare(metrics=paddle.metric.Accuracy())
1496 1497
            result = model.evaluate(val_dataset, batch_size=64)
            print(result)
1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557
                
        """

        if eval_data is not None and isinstance(eval_data, Dataset):
            eval_sampler = DistributedBatchSampler(
                eval_data, batch_size=batch_size)
            eval_loader = DataLoader(
                eval_data,
                batch_sampler=eval_sampler,
                places=self._place,
                num_workers=num_workers,
                return_list=True)
        else:
            eval_loader = eval_data

        self._test_dataloader = eval_loader

        cbks = config_callbacks(
            callbacks,
            model=self,
            log_freq=log_freq,
            verbose=verbose,
            metrics=self._metrics_name(), )

        eval_steps = self._len_data_loader(eval_loader)
        cbks.on_begin('eval',
                      {'steps': eval_steps,
                       'metrics': self._metrics_name()})

        logs = self._run_one_epoch(eval_loader, cbks, 'eval')

        cbks.on_end('eval', logs)

        self._test_dataloader = None

        eval_result = {}
        for k in self._metrics_name():
            eval_result[k] = logs[k]

        return eval_result

    def predict(self,
                test_data,
                batch_size=1,
                num_workers=0,
                stack_outputs=False,
                callbacks=None):
        """
        Compute the output predictions on testing data.

        Args:
            test_data (Dataset|DataLoader): An iterable data loader is used for
                predict. An instance of paddle.io.Dataset or paddle.io.Dataloader
                is recomended.
            batch_size (int): Integer number. The batch size of train_data and eval_data.
                When train_data and eval_data are both the instance of Dataloader, this
                argument will be ignored. Default: 1.
            num_workers (int): The number of subprocess to load data, 0 for no subprocess 
                used and loading data in main process. When train_data and eval_data are
                both the instance of Dataloader, this argument will be ignored. Default: 0.
1558
            stack_outputs (bool): Whether stack output field like a batch, as for an output
1559 1560 1561 1562 1563
                filed of a sample is in shape [X, Y], test_data contains N samples, predict
                output field will be in shape [N, X, Y] if stack_output is True, and will
                be a length N list in shape [[X, Y], [X, Y], ....[X, Y]] if stack_outputs
                is False. stack_outputs as False is used for LoDTensor output situation,
                it is recommended set as True if outputs contains no LoDTensor. Default: False.
1564
            callbacks(Callback): A Callback instance, default None.
1565 1566 1567 1568 1569 1570 1571
        Returns:
            list: output of models.

        Examples:
        .. code-block:: python

            import numpy as np
1572
            import paddle
1573
            from paddle.static import InputSpec
1574

1575
            class MnistDataset(paddle.vision.datasets.MNIST):
1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590
                def __init__(self, mode, return_label=True):
                    super(MnistDataset, self).__init__(mode=mode)
                    self.return_label = return_label

                def __getitem__(self, idx):
                    img = np.reshape(self.images[idx], [1, 28, 28])
                    if self.return_label:
                        return img, np.array(self.labels[idx]).astype('int64')
                    return img,

                def __len__(self):
                    return len(self.images)

            test_dataset = MnistDataset(mode='test', return_label=False)

1591
            # declarative mode
1592 1593
            input = InputSpec([-1, 1, 28, 28], 'float32', 'image')
            model = paddle.Model(paddle.vision.models.LeNet(), input)
1594
            model.prepare()
1595 1596

            result = model.predict(test_dataset, batch_size=64)
1597
            print(len(result[0]), result[0][0].shape)
1598 1599

            # imperative mode
1600
            device = paddle.set_device('cpu')
1601
            paddle.disable_static(device)
1602
            model = paddle.Model(paddle.vision.models.LeNet())
1603 1604 1605
            model.prepare()
            result = model.predict(test_dataset, batch_size=64)
            print(len(result[0]), result[0][0].shape)
1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644
        """

        if test_data is not None and isinstance(test_data, Dataset):
            test_sampler = DistributedBatchSampler(
                test_data, batch_size=batch_size)
            test_loader = DataLoader(
                test_data,
                batch_sampler=test_sampler,
                places=self._place,
                num_workers=num_workers,
                return_list=True)
        else:
            test_loader = test_data

        self._test_dataloader = test_loader

        cbks = config_callbacks(callbacks, model=self, verbose=1)

        test_steps = self._len_data_loader(test_loader)
        logs = {'steps': test_steps}

        cbks.on_begin('test', logs)

        outputs = []

        logs, outputs = self._run_one_epoch(test_loader, cbks, 'test')

        outputs = list(zip(*outputs))

        # NOTE: for lod tensor output, we should not stack outputs
        # for stacking may lose its detail info
        if stack_outputs:
            outputs = [np.vstack(outs) for outs in outputs]

        self._test_dataloader = None

        cbks.on_end('test', logs)
        return outputs

1645 1646 1647 1648 1649
    def _save_inference_model(self,
                              save_dir,
                              model_filename=None,
                              params_filename=None,
                              model_only=False):
1650
        """
1651 1652 1653 1654 1655
        Save inference model can be in static or dynamic mode.
        It should be noted that before using `save_inference_model`, you should
        run the model, and the shape you saved is as same as the input of its
        running. `@paddle.jit.to_static` must be added on `forward` function of
        your layer in dynamic mode now and these will be optimized later.
1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671

        Args:
            save_dir (str): The directory path to save the inference model.
            model_filename (str|None): The name of file to save the inference
                model itself. If is set None, a default filename
                :code:`__model__` will be used.
            params_filename (str|None): The name of file to save all related
                parameters. If it is set None, parameters will be saved
                in separate files .
            model_only (bool): If True, It will save inference model only,
                and do not save parameters. Default: False.

        Returns:
            list: The fetch variables' name list
        """

1672 1673 1674 1675 1676 1677
        def get_inout_spec(all_vars, return_name=False):
            result_list = []
            valid_vars = [var for var in all_vars if isinstance(var, Variable)]
            result_list = valid_vars
            if return_name:
                result_list = [var.name for var in result_list]
1678

1679
            return result_list
1680

1681 1682 1683 1684 1685 1686
        # TODO:
        # 1. Make it Unnecessary to run model before calling `save_inference_model` for users in dygraph.
        # 2. Save correct shape of input, now the interface stores the shape that the user sent to 
        #    the inputs of the model in running.
        # 3. Make it Unnecessary to add `@paddle.jit.to_static` for users in dynamic mode.
        if fluid.in_dygraph_mode():
1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746
            with fluid.framework._dygraph_guard(None):
                layer = self.network

                # 1. input check
                prog_translator = ProgramTranslator()
                if not prog_translator.enable_declarative:
                    raise RuntimeError(
                        "save_inference_model doesn't work when setting ProgramTranslator.enable=False."
                    )
                if not isinstance(layer, Layer):
                    raise TypeError(
                        "The input layer should be 'Layer', but received layer type is %s."
                        % type(layer))

                # 2. get program of declarative Layer.forward
                concrete_program = layer.forward.concrete_program

                # NOTE: we maintain the mapping of variable name to
                # structured name, the buffer variable (non-persistable)
                # saved to inference program may not need by dygraph Layer,
                # we only record the state_dict variable's structured name
                state_names_dict = dict()
                for structured_name, var in layer.state_dict().items():
                    state_names_dict[var.name] = structured_name

                # 3. share parameters from Layer to scope & record var info
                scope = core.Scope()
                extra_var_info = dict()
                for param_or_buffer in concrete_program.parameters:
                    # share to scope
                    param_or_buffer_tensor = scope.var(
                        param_or_buffer.name).get_tensor()
                    src_tensor = param_or_buffer.value().get_tensor()
                    param_or_buffer_tensor._share_data_with(src_tensor)
                    # record var info
                    extra_info_dict = dict()
                    if param_or_buffer.name in state_names_dict:
                        extra_info_dict['structured_name'] = state_names_dict[
                            param_or_buffer.name]
                    extra_info_dict[
                        'stop_gradient'] = param_or_buffer.stop_gradient
                    if isinstance(param_or_buffer, ParamBase):
                        extra_info_dict['trainable'] = param_or_buffer.trainable
                    extra_var_info[param_or_buffer.name] = extra_info_dict

                # 4. build input & output spec
                input_var_names = get_inout_spec(concrete_program.inputs, True)
                output_vars = get_inout_spec(concrete_program.outputs)

                # 5. save inference model
                with scope_guard(scope):
                    return fluid.io.save_inference_model(
                        dirname=save_dir,
                        feeded_var_names=input_var_names,
                        target_vars=output_vars,
                        executor=Executor(_current_expected_place()),
                        main_program=concrete_program.main_program.clone(),
                        model_filename=model_filename,
                        params_filename=params_filename,
                        program_only=model_only)
1747

1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766
        else:
            prog = self._adapter._progs.get('test', None)
            assert prog, \
                "Model is not ready, please call `model.prepare()` first"

            infer_prog = prog.clone(for_test=True)

            input_names = [v.name for v in self._adapter._input_vars['test']]
            endpoints = self._adapter._endpoints['test']['output']

            return fluid.io.save_inference_model(
                save_dir,
                input_names,
                endpoints,
                self._adapter._executor,
                main_program=infer_prog,
                model_filename=model_filename,
                params_filename=params_filename,
                program_only=model_only)
1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792

    def _run_one_epoch(self, data_loader, callbacks, mode, logs={}):
        outputs = []
        for step, data in enumerate(data_loader):
            # data might come from different types of data_loader and have
            # different format, as following:
            # 1. DataLoader in static graph:
            #    [[input1, input2, ..., label1, lable2, ...]]
            # 2. DataLoader in dygraph
            #    [input1, input2, ..., label1, lable2, ...]
            # 3. custumed iterator yield concated inputs and labels:
            #   [input1, input2, ..., label1, lable2, ...]
            # 4. custumed iterator yield seperated inputs and labels:
            #   ([input1, input2, ...], [label1, lable2, ...])
            # To handle all of these, flatten (nested) list to list.
            data = flatten(data)
            # LoDTensor.shape is callable, where LoDTensor comes from
            # DataLoader in static graph
            batch_size = data[0].shape()[0] if callable(data[
                0].shape) else data[0].shape[0]

            callbacks.on_batch_begin(mode, step, logs)

            if mode != 'test':
                outs = getattr(self, mode + '_batch')(data[:len(self._inputs)],
                                                      data[len(self._inputs):])
1793
                if self._metrics and self._loss:
1794
                    metrics = [[l[0] for l in outs[0]]]
1795
                elif self._loss:
1796 1797 1798
                    metrics = [[l[0] for l in outs]]
                else:
                    metrics = []
1799 1800 1801 1802 1803 1804 1805 1806 1807 1808

                # metrics
                for metric in self._metrics:
                    res = metric.accumulate()
                    metrics.extend(to_list(res))

                assert len(self._metrics_name()) == len(metrics)
                for k, v in zip(self._metrics_name(), metrics):
                    logs[k] = v
            else:
L
LielinJiang 已提交
1809 1810 1811 1812 1813 1814
                if self._inputs is not None:
                    outs = getattr(self,
                                   mode + '_batch')(data[:len(self._inputs)])
                else:
                    outs = getattr(self, mode + '_batch')(data)

1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830
                outputs.append(outs)

            logs['step'] = step
            if mode == 'train' or self._adapter._merge_count.get(
                    mode + '_batch', 0) <= 0:
                logs['batch_size'] = batch_size * ParallelEnv().nranks
            else:
                logs['batch_size'] = self._adapter._merge_count[mode + '_batch']

            callbacks.on_batch_end(mode, step, logs)
        self._reset_metrics()

        if mode == 'test':
            return logs, outputs
        return logs

1831 1832 1833 1834
    def _verify_spec(self, specs, is_input=False):
        out_specs = []

        if specs is None:
1835 1836
            # Note(Aurelius84): If not specific specs of `Input`, using argument names of `forward` function
            # to generate `Input`. But how can we know the actual shape of each input tensor?
1837 1838
            if is_input:
                out_specs = [
1839 1840 1841
                    Input(
                        name=n, shape=[None])
                    for n in extract_args(self.network.forward) if n != 'self'
1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861
                ]
            else:
                out_specs = to_list(specs)
        elif isinstance(specs, dict):
            assert is_input == False
            out_specs = [specs[n] \
                for n in extract_args(self.network.forward) if n != 'self']
        else:
            out_specs = to_list(specs)
        # Note: checks each element has specificed `name`.
        if out_specs is not None:
            for i, spec in enumerate(out_specs):
                assert isinstance(spec, Input)
                if spec.name is None:
                    raise ValueError(
                        "Requires Input[{}].name != None, but receive `None` with {}.".
                        format(i, spec))

        return out_specs

1862 1863 1864 1865 1866
    def _reset_metrics(self):
        for metric in self._metrics:
            metric.reset()

    def _metrics_name(self):
1867
        metrics_name = ['loss'] if self._loss else []
1868 1869 1870 1871 1872 1873 1874 1875 1876 1877
        for m in self._metrics:
            metrics_name.extend(to_list(m.name()))
        return metrics_name

    def _len_data_loader(self, data_loader):
        try:
            steps = len(data_loader)
        except Exception:
            steps = None
        return steps