model.py 86.6 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
import paddle
30
from paddle import fluid
31
from paddle.fluid import core
32 33 34
from paddle.fluid.framework import in_dygraph_mode
from paddle.fluid.framework import Variable
from paddle.fluid.framework import _get_paddle_place
35
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 41
from paddle.fluid.dygraph.io import INFER_MODEL_SUFFIX
from paddle.fluid.dygraph.io import INFER_PARAMS_SUFFIX
42
from paddle.fluid.layers.utils import flatten
43
from paddle.fluid.layers import collective
44

45 46 47
from paddle.io import DataLoader
from paddle.io import Dataset
from paddle.io import DistributedBatchSampler
48
from paddle.metric import Metric
49
from paddle.static import InputSpec as Input
50
import paddle.distributed as dist
J
Jiaqi Liu 已提交
51 52
import paddle.distributed.fleet as fleet
from paddle.distributed.fleet.base import role_maker
53

L
LiuChiachi 已提交
54
from .callbacks import config_callbacks, EarlyStopping
L
LielinJiang 已提交
55
from .model_summary import summary
56

57
__all__ = []
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

_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)
136
    block = program.global_block()
137 138
    if rank == 0 and wait_port:
        wait_server_ready(other_endpoints)
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
    if core.is_compiled_with_cuda():
        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,
            })
    elif core.is_compiled_with_npu():
        hccl_id_var = block.create_var(
Z
zhangchunle 已提交
166
            name=fluid.unique_name.generate('hccl_id'),
167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
            persistable=True,
            type=core.VarDesc.VarType.RAW)
        block.append_op(
            type='c_gen_hccl_id',
            inputs={},
            outputs={'Out': hccl_id_var},
            attrs={
                'rank': rank,
                'endpoint': current_endpoint,
                'other_endpoints': other_endpoints
            })
        block.append_op(
            type='c_comm_init_hccl',
            inputs={'X': hccl_id_var},
            outputs={},
            attrs={
                'rank': rank,
                'ring_id': 0,
                'device_id': int(os.getenv("FLAGS_selected_npus")),
                'rank_ids': nranks
            })
188 189 190 191 192 193 194


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

195
    place = _get_paddle_place(place)
196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228
    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
229 230


L
LiuChiachi 已提交
231
def _update_input_info(inputs):
L
LiuChiachi 已提交
232
    "Get input shape list by given inputs in Model initialization."
233
    shapes = None
L
LiuChiachi 已提交
234
    dtypes = None
L
LiuChiachi 已提交
235 236
    if isinstance(inputs, Input):
        shapes = [list(inputs.shape)]
L
LiuChiachi 已提交
237
        dtypes = [inputs.dtype]
238
    elif isinstance(inputs, (list, tuple)):
239
        shapes = [list(input.shape) for input in inputs]
L
LiuChiachi 已提交
240
        dtypes = [input.dtype for input in inputs]
241 242
    elif isinstance(inputs, dict):
        shapes = [list(inputs[name].shape) for name in inputs]
L
LiuChiachi 已提交
243 244 245 246
        dtypes = [inputs[name].dtype for name in inputs]
    else:
        return None
    return shapes, dtypes
247 248


249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279
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

J
Jiaqi Liu 已提交
280 281 282 283 284
        self._amp_level = "O0"
        self._amp_configs = {}
        self._amp_custom_lists = {}
        self._use_fp16_guard = True

285 286 287 288 289 290 291 292
    @property
    def mode(self):
        return self.model.mode

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

L
lyuwenyu 已提交
293
    def train_batch(self, inputs, labels=None, update=True):
294 295 296
        assert self.model._optimizer, \
            "model not ready, please call `model.prepare()` first"
        self.mode = 'train'
L
update  
lyuwenyu 已提交
297
        assert update is True, "Does not support `update == False` in static mode by now."
298 299 300 301 302 303
        return self._run(inputs, labels)

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

304
    def predict_batch(self, inputs):
305 306 307 308
        self.mode = 'test'
        return self._run(inputs, None)

    def parameters(self, *args, **kwargs):
309
        return self.model.network.parameters(*args, **kwargs)
310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327

    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"
328
        _save(self.model.network.state_dict(), param_path)
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499
        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[:]
500

501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
        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))
524 525 526 527 528

        if num_loss and len(metrics):
            return rets[:num_loss], metrics
        else:
            return rets[:num_loss] if num_loss else metrics
529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559

    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):
560 561
            inputs = self.model._inputs
            labels = self.model._labels if self.model._labels else []
562 563
            inputs = [k._create_feed_layer() for k in to_list(inputs)]
            labels = [k._create_feed_layer() for k in to_list(labels)]
564
            self._label_vars[mode] = labels
565
            outputs = to_list(self.model.network.forward(*inputs))
566

567 568
            if mode != 'test' and self.model._loss:
                losses = self.model._loss(*(outputs + labels))
569 570 571 572 573 574 575 576

            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:
577
                    metrics.append(to_list(metric.compute(*(outputs + labels))))
578 579 580 581 582 583

            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)
J
Jiaqi Liu 已提交
584 585 586 587 588 589 590
                    dist_strategy = fleet.DistributedStrategy()
                    if self._amp_level != 'O0':
                        dist_strategy.amp = True
                        dist_strategy.amp_configs = self._amp_configs.copy()
                        dist_strategy.amp_configs.update(self._amp_custom_lists)
                        dist_strategy.amp_configs[
                            'use_pure_fp16'] = self._amp_level == 'O2'
591 592
                    self.model._optimizer = fleet.distributed_optimizer(
                        self.model._optimizer, strategy=dist_strategy)
J
Jiaqi Liu 已提交
593 594 595 596 597 598 599 600 601 602 603
                elif self._amp_level != "O0" and core.is_compiled_with_cuda:
                    amp_lists = paddle.static.amp.AutoMixedPrecisionLists(
                        **self.
                        _amp_custom_lists) if self._amp_custom_lists else None

                    self.model._optimizer = paddle.static.amp.decorate(
                        self.model._optimizer,
                        amp_lists=amp_lists,
                        use_pure_fp16=self._amp_level == "O2",
                        use_fp16_guard=self._use_fp16_guard,
                        **self._amp_configs)
604 605 606 607 608 609 610 611 612 613 614

                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,
615
            "loss": to_list(losses),
616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646
            "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)

J
Jiaqi Liu 已提交
647 648 649 650
        if self._amp_level == "O2" and mode == 'train' and core.is_compiled_with_cuda(
        ):
            self.model._optimizer.amp_init(place)

651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671
        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
        }

L
LiuChiachi 已提交
672
        self._input_info = None
J
Jiaqi Liu 已提交
673 674 675 676 677
        self._amp_level = "O0"
        self._amp_configs = {}
        self._amp_custom_lists = {}
        self._use_fp16_guard = True

678
        if self._nranks > 1:
679
            dist.init_parallel_env()
680 681 682 683 684
            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
685 686
            self.ddp_model = fluid.dygraph.parallel.DataParallel(
                self.model.network, stradegy)
687 688 689 690 691 692 693 694 695 696

    @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
L
lyuwenyu 已提交
697
    def train_batch(self, inputs, labels=None, update=True):
698 699
        assert self.model._optimizer, \
            "model not ready, please call `model.prepare()` first"
700
        self.model.network.train()
701 702
        self.mode = 'train'
        inputs = to_list(inputs)
L
LiuChiachi 已提交
703
        self._input_info = _update_input_info(inputs)
704 705 706
        labels = labels or []
        labels = [to_variable(l) for l in to_list(labels)]

J
Jiaqi Liu 已提交
707 708 709 710 711 712
        if self._amp_level != "O0":
            scaler = paddle.amp.GradScaler(**self._amp_configs)
        with paddle.amp.auto_cast(
                enable=self._amp_level != 'O0', **self._amp_custom_lists):
            if self._nranks > 1:
                outputs = self.ddp_model.forward(
Z
zhangchunle 已提交
713
                    *[to_variable(x) for x in inputs])
J
Jiaqi Liu 已提交
714 715
            else:
                outputs = self.model.network.forward(
Z
zhangchunle 已提交
716
                    *[to_variable(x) for x in inputs])
717

J
Jiaqi Liu 已提交
718 719 720
            losses = self.model._loss(*(to_list(outputs) + labels))
            losses = to_list(losses)
            final_loss = fluid.layers.sum(losses)
721

J
Jiaqi Liu 已提交
722 723 724
        if self._amp_level != "O0":
            scaled = scaler.scale(final_loss)
            scaled.backward()
L
lyuwenyu 已提交
725 726 727
            if update:
                scaler.minimize(self.model._optimizer, scaled)
                self.model.network.clear_gradients()
J
Jiaqi Liu 已提交
728 729
        else:
            final_loss.backward()
L
lyuwenyu 已提交
730 731 732
            if update:
                self.model._optimizer.minimize(final_loss)
                self.model.network.clear_gradients()
L
update  
lyuwenyu 已提交
733

734 735
        metrics = []
        for metric in self.model._metrics:
736
            metric_outs = metric.compute(*(to_list(outputs) + labels))
Z
zhangchunle 已提交
737
            m = metric.update(*[to_numpy(m) for m in to_list(metric_outs)])
738 739 740 741 742 743
            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):
744
        self.model.network.eval()
745 746
        self.mode = 'eval'
        inputs = to_list(inputs)
L
LiuChiachi 已提交
747
        self._input_info = _update_input_info(inputs)
748 749 750
        labels = labels or []
        labels = [to_variable(l) for l in to_list(labels)]

Z
zhangchunle 已提交
751
        outputs = self.model.network.forward(*[to_variable(x) for x in inputs])
752 753
        if self.model._loss:
            losses = self.model._loss(*(to_list(outputs) + labels))
754 755
            losses = to_list(losses)

756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780
        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

781
            metric_outs = metric.compute(*(to_list(outputs) + labels))
Z
zhangchunle 已提交
782
            m = metric.update(*[to_numpy(m) for m in to_list(metric_outs)])
783 784
            metrics.append(m)

785
        if self.model._loss and len(metrics):
786
            return [to_numpy(l) for l in losses], metrics
787
        elif self.model._loss:
788 789 790
            return [to_numpy(l) for l in losses]
        else:
            return metrics
791

792
    def predict_batch(self, inputs):
793
        self.model.network.eval()
794 795
        self.mode = 'test'
        inputs = [to_variable(x) for x in to_list(inputs)]
L
LiuChiachi 已提交
796
        self._input_info = _update_input_info(inputs)
797
        outputs = self.model.network.forward(*inputs)
798 799 800 801 802 803
        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):
804
        return self.model.network.parameters(*args, **kwargs)
805 806

    def save(self, path):
807
        params = self.model.network.state_dict()
808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823
        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

824 825
        # 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
826 827 828 829 830 831 832 833 834 835 836
        # 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
837
        param_names = [param.name for param in self.model.network.parameters()]
838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868
        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

869 870
        if not hasattr(self.model._optimizer, 'set_state_dict'):
            warnings.warn(
871
                "paddle.fluid.optimizer is deprecated in API 2.0, please use paddle.optimizer instead."
872 873 874 875
            )
            self.model._optimizer.set_dict(converted_state)
        else:
            self.model._optimizer.set_state_dict(converted_state)
876 877


878
class Model(object):
879 880 881
    """
    An Model object is network with training and inference features.
    Dynamic graph and static graph are supported at the same time,
882
    switched by `paddle.enable_static()`. The usage is as follows.
883
    But note, the switching between dynamic and static should be before
884
    instantiating a Model. The input description, i.e, paddle.static.InputSpec,
885
    must be required for static graph.
886

J
Jiaqi Liu 已提交
887 888 889 890 891
    When training on GPU, auto mixed precision (AMP) training is supported, and
    pure float16 training is also supported in static mode while using Adam,
    AdamW and Momentum optimizer. Before using pure float16 training,
    `multi_precision` could be set to True when creating optimizer, which can
    avoid poor accuracy or slow convergence in a way, and inputs of dtype float
892 893 894 895
    should be cast to float16 by users. `paddle.static.amp.fp16_guard` API
    should be also used to limit the range of pure float16 training, otherwise,
    'use_fp16_guard' should be set to False by users. However, limiting the
    range of is not supported during training using AMP.
J
Jiaqi Liu 已提交
896

897
    Args:
898 899
        network (paddle.nn.Layer): The network is an instance of
            paddle.nn.Layer.
900 901
        inputs (InputSpec|list|tuple|dict|None): `inputs`, entry points of network,
            could be a InputSpec instance, or list/tuple of InputSpec instances,
902 903
            or dict ({name: InputSpec}), and it couldn't be None in static
            graph.
904 905
        labels (InputSpec|list|tuple|None): `labels`, entry points of network,
            could be a InputSpec instnace or list/tuple of InputSpec instances,
906
            or None. For static graph, if labels is required in loss,
907 908 909
            labels must be set. Otherwise, it could be None.


910
    Examples:
J
Jiaqi Liu 已提交
911 912
        1. A common example

913 914
        .. code-block:: python

915 916 917 918 919 920
          import paddle
          import paddle.nn as nn
          import paddle.vision.transforms as T
          from paddle.static import InputSpec
  
          device = paddle.set_device('cpu') # or 'gpu'
J
Jiaqi Liu 已提交
921

922 923 924 925 926 927 928 929 930 931 932 933 934
          net = nn.Sequential(
              nn.Flatten(1),
              nn.Linear(784, 200),
              nn.Tanh(),
              nn.Linear(200, 10))
  
          # inputs and labels are not required for dynamic graph.
          input = InputSpec([None, 784], 'float32', 'x')
          label = InputSpec([None, 1], 'int64', 'label')
          
          model = paddle.Model(net, input, label)
          optim = paddle.optimizer.SGD(learning_rate=1e-3,
              parameters=model.parameters())
J
Jiaqi Liu 已提交
935

936 937 938 939 940 941 942 943 944 945
          model.prepare(optim,
                        paddle.nn.CrossEntropyLoss(),
                        paddle.metric.Accuracy())
          
          transform = T.Compose([
              T.Transpose(),
              T.Normalize([127.5], [127.5])
          ])
          data = paddle.vision.datasets.MNIST(mode='train', transform=transform)
          model.fit(data, epochs=2, batch_size=32, verbose=1)
J
Jiaqi Liu 已提交
946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978


        2. An example using mixed precision training.

        .. code-block:: python

          import paddle
          import paddle.nn as nn
          import paddle.vision.transforms as T

          def run_example_code():
            device = paddle.set_device('gpu')

            net = nn.Sequential(nn.Flatten(1), nn.Linear(784, 200), nn.Tanh(),
                                nn.Linear(200, 10))

            model = paddle.Model(net)
            optim = paddle.optimizer.SGD(learning_rate=1e-3, parameters=model.parameters())

            amp_configs = {
                "level": "O1",
                "custom_white_list": {'conv2d'},
                "use_dynamic_loss_scaling": True
            }
            model.prepare(optim,
                paddle.nn.CrossEntropyLoss(),
                paddle.metric.Accuracy(),
                amp_configs=amp_configs)

            transform = T.Compose([T.Transpose(), T.Normalize([127.5], [127.5])])
            data = paddle.vision.datasets.MNIST(mode='train', transform=transform)
            model.fit(data, epochs=2, batch_size=32, verbose=1)

979
          # mixed precision training is only supported on GPU now.
J
Jiaqi Liu 已提交
980 981 982
          if paddle.is_compiled_with_cuda():
            run_example_code()

983 984
    """

985
    def __init__(self, network, inputs=None, labels=None):
986
        self.mode = 'train'
987
        self.network = network
988 989
        self._inputs = None
        self._labels = None
990
        self._loss = None
991 992
        self._loss_weights = None
        self._optimizer = None
L
LiuChiachi 已提交
993
        self._input_info = None
994
        self._is_shape_inferred = False
995
        self._test_dataloader = None
L
LiuChiachi 已提交
996
        self.stop_training = False
997

998
        if not in_dygraph_mode():
999
            if not isinstance(inputs, (list, tuple, dict, Input)):
1000
                raise TypeError(
1001 1002
                    "'inputs' must be list or tuple or dict, and couldn't be None."
                )
1003
        elif inputs:
L
LiuChiachi 已提交
1004
            self._input_info = _update_input_info(inputs)
L
LielinJiang 已提交
1005

1006
        self._inputs = self._verify_spec(inputs, is_input=True)
1007
        self._labels = self._verify_spec(labels)
1008

1009 1010 1011 1012 1013 1014
        # init backend
        if fluid.in_dygraph_mode():
            self._adapter = DynamicGraphAdapter(self)
        else:
            self._adapter = StaticGraphAdapter(self)

L
lyuwenyu 已提交
1015
    def train_batch(self, inputs, labels=None, update=True):
1016
        """
L
lyuwenyu 已提交
1017 1018
        Run one training step on one batch of data. And using `update` indicates
        whether optimizer update gradients computing by this batch.
1019 1020

        Args:
1021 1022 1023 1024 1025 1026 1027
            inputs (numpy.ndarray|Tensor|list): Batch of input data. It could 
                be a numpy array or paddle.Tensor, or a list of arrays or 
                tensors (in case the model has multiple inputs).
            labels (numpy.ndarray|Tensor|list): Batch of labels. It could be 
                a numpy array or paddle.Tensor, or a list of arrays or tensors 
                (in case the model has multiple labels). If has no labels, 
                set None. Default is None.
L
lyuwenyu 已提交
1028 1029
            update (bool): Whether update parameters after loss.backward() computing.
                Using it to accumulate gradients. Default is True.
1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040

        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
1041
              import paddle
1042 1043
              import paddle.nn as nn
              from paddle.static import InputSpec
1044

1045
              device = paddle.set_device('cpu') # or 'gpu'
1046

1047 1048 1049 1050 1051 1052 1053 1054
              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)
1055
              optim = paddle.optimizer.SGD(learning_rate=1e-3,
1056
                  parameters=model.parameters())
1057
              model.prepare(optim, paddle.nn.CrossEntropyLoss())
1058 1059 1060 1061 1062
              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)
        """
L
lyuwenyu 已提交
1063
        loss = self._adapter.train_batch(inputs, labels, update)
L
LiuChiachi 已提交
1064
        if fluid.in_dygraph_mode() and self._input_info is None:
L
LiuChiachi 已提交
1065
            self._update_inputs()
1066
        return loss
1067

1068
    @paddle.no_grad()
1069 1070 1071 1072 1073
    def eval_batch(self, inputs, labels=None):
        """
        Run one evaluating step on a batch of data.

        Args:
1074 1075 1076 1077 1078 1079 1080
            inputs (numpy.ndarray|Tensor|list): Batch of input data. It could 
                be a numpy array or paddle.Tensor, or a list of arrays or 
                tensors (in case the model has multiple inputs).
            labels (numpy.ndarray|Tensor|list): Batch of labels. It could be 
                a numpy array or paddle.Tensor, or a list of arrays or tensors 
                (in case the model has multiple labels). If has no labels, 
                set None. Default is None.
1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091

        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
1092
              import paddle
1093 1094
              import paddle.nn as nn
              from paddle.static import InputSpec
1095

1096
              device = paddle.set_device('cpu') # or 'gpu'
1097

1098 1099 1100 1101 1102 1103 1104 1105
              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)
1106
              optim = paddle.optimizer.SGD(learning_rate=1e-3,
1107
                  parameters=model.parameters())
1108
              model.prepare(optim,
1109
                            paddle.nn.CrossEntropyLoss())
1110 1111 1112 1113 1114
              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)
        """
1115
        loss = self._adapter.eval_batch(inputs, labels)
L
LiuChiachi 已提交
1116
        if fluid.in_dygraph_mode() and self._input_info is None:
L
LiuChiachi 已提交
1117
            self._update_inputs()
1118
        return loss
1119

1120
    @paddle.no_grad()
1121
    def predict_batch(self, inputs):
1122
        """
1123
        Run one predicting step on a batch of data.
1124 1125

        Args:
1126 1127 1128
            inputs (numpy.ndarray|Tensor|list): Batch of input data. It could 
                be a numpy array or paddle.Tensor, or a list of arrays or 
                tensors (in case the model has multiple inputs).
1129 1130 1131 1132 1133 1134 1135 1136 1137 1138

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

        Examples:

            .. code-block:: python
            
              import numpy as np
1139
              import paddle
1140
              import paddle.nn as nn
L
LielinJiang 已提交
1141
              from paddle.static import InputSpec
1142

1143
              device = paddle.set_device('cpu') # or 'gpu'
L
LielinJiang 已提交
1144 1145 1146
              
              input = InputSpec([None, 784], 'float32', 'x')
              label = InputSpec([None, 1], 'int64', 'label')
1147

1148 1149 1150 1151 1152 1153
              net = nn.Sequential(
                  nn.Linear(784, 200),
                  nn.Tanh(),
                  nn.Linear(200, 10),
                  nn.Softmax())

L
LielinJiang 已提交
1154
              model = paddle.Model(net, input, label)
1155
              model.prepare()
1156
              data = np.random.random(size=(4,784)).astype(np.float32)
1157
              out = model.predict_batch([data])
1158 1159
              print(out)
        """
1160
        loss = self._adapter.predict_batch(inputs)
L
LiuChiachi 已提交
1161
        if fluid.in_dygraph_mode() and self._input_info is None:
L
LiuChiachi 已提交
1162
            self._update_inputs()
1163
        return loss
1164

1165 1166 1167 1168 1169
    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`.
1170

1171 1172
        If `training` is set to True, the parameters saved contain all 
        the trainable Variable, will save to a file with suffix ".pdparams".
1173 1174 1175 1176
        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).
1177
        This function will silently overwrite existing file at the target location.
1178

1179
        If `training` is set to False, only inference model will be saved.
1180 1181

        Args:
1182 1183 1184
            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.
1185 1186
            training (bool, optional): Whether to save for training. If not, save
                for inference only. Default: True.
1187 1188 1189 1190 1191 1192 1193

        Returns:
            None

        Examples:

            .. code-block:: python
1194

1195
                import paddle
1196
                import paddle.nn as nn
1197
                import paddle.vision.transforms as T
1198
                from paddle.static import InputSpec
1199

1200
                class Mnist(nn.Layer):
1201
                    def __init__(self):
1202
                        super(Mnist, self).__init__()
1203
                        self.net = nn.Sequential(
L
LielinJiang 已提交
1204
                            nn.Flatten(1),
1205 1206 1207 1208
                            nn.Linear(784, 200),
                            nn.Tanh(),
                            nn.Linear(200, 10),
                            nn.Softmax())
1209

1210
                    def forward(self, x):
1211
                        return self.net(x)
1212

1213
                dynamic = True  # False
1214
                # if use static graph, do not set
1215 1216
                if not dynamic:
                    paddle.enable_static()
1217

1218 1219 1220
                input = InputSpec([None, 784], 'float32', 'x')
                label = InputSpec([None, 1], 'int64', 'label')
                model = paddle.Model(Mnist(), input, label)
1221
                optim = paddle.optimizer.SGD(learning_rate=1e-3,
1222
                    parameters=model.parameters())
1223
                model.prepare(optim, paddle.nn.CrossEntropyLoss())
1224 1225 1226 1227 1228 1229 1230
                
                transform = T.Compose([
                    T.Transpose(),
                    T.Normalize([127.5], [127.5])
                ])
                data = paddle.vision.datasets.MNIST(mode='train', transform=transform)
                
1231
                model.fit(data, epochs=1, batch_size=32, verbose=0)
1232 1233
                model.save('checkpoint/test')  # save for training
                model.save('inference_model', False)  # save for inference
1234
        """
1235

1236
        if ParallelEnv().local_rank == 0:
1237 1238 1239 1240
            if not training:
                self._save_inference_model(path)
            else:
                self._adapter.save(path)
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

    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
            
1275
              import paddle
1276
              import paddle.nn as nn
L
LielinJiang 已提交
1277 1278
              from paddle.static import InputSpec

1279
              device = paddle.set_device('cpu')
L
LielinJiang 已提交
1280 1281

              input = InputSpec([None, 784], 'float32', 'x')
1282 1283 1284 1285 1286

              model = paddle.Model(nn.Sequential(
                  nn.Linear(784, 200),
                  nn.Tanh(),
                  nn.Linear(200, 10),
L
LielinJiang 已提交
1287 1288
                  nn.Softmax()), input)

1289
              model.save('checkpoint/test')
1290 1291 1292 1293 1294 1295 1296
              model.load('checkpoint/test')
        """

        def _load_state_from_path(path):
            if not os.path.exists(path):
                return
            with open(path, 'rb') as f:
T
tianshuo78520a 已提交
1297
                return pickle.load(f, encoding='latin1')
1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320

        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 = []
1321
        for key, param in self.network.state_dict().items():
1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349
            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

1350
              import paddle
1351
              import paddle.nn as nn
L
LielinJiang 已提交
1352
              from paddle.static import InputSpec
1353

L
LielinJiang 已提交
1354 1355
              input = InputSpec([None, 784], 'float32', 'x')
              
1356 1357 1358
              model = paddle.Model(nn.Sequential(
                  nn.Linear(784, 200),
                  nn.Tanh(),
L
LielinJiang 已提交
1359 1360
                  nn.Linear(200, 10)), input)

1361 1362 1363 1364
              params = model.parameters()
        """
        return self._adapter.parameters()

J
Jiaqi Liu 已提交
1365 1366 1367 1368 1369
    def _prepare_amp(self, amp_configs):
        def _check_pure_fp16_configs():
            # pure float16 training has some restricts now
            if self._adapter._amp_level == "O2":
                if in_dygraph_mode():
1370 1371 1372
                    warnings.warn(
                        "Pure float16 training is not supported in dygraph mode now, and it will be supported in future version."
                    )
J
Jiaqi Liu 已提交
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
                else:
                    # grad clip is not supported in pure fp16 training now
                    assert self._optimizer._grad_clip is None, \
                        "Grad clip is not supported in pure float16 training now, and it will be supported in future version."

        self._adapter._amp_custom_lists = {}
        self._adapter._amp_configs = {}

        # check and get level of mixed precision training
        if not amp_configs:
            self._adapter._amp_level = 'O0'
            return
        elif isinstance(amp_configs, str):
            if amp_configs not in ('O0', 'O1', 'O2'):
                raise ValueError(
                    "The level of amp_configs should be 'O0', 'O1' or 'O2'.")
            self._adapter._amp_level = amp_configs
            _check_pure_fp16_configs()
            return
        else:
            if 'level' not in amp_configs:
                self._adapter._amp_level = 'O1'
            elif amp_configs['level'] not in ('O0', 'O1', 'O2'):
                raise ValueError(
                    "amp_configs['level'] should be 'O0', 'O1' or 'O2'.")
            else:
                self._adapter._amp_level = amp_configs['level']
        amp_config_key_set = set(amp_configs.keys()) - {'level'}
        if not amp_config_key_set or self._adapter._amp_level == 'O0':
            return

        if 'use_pure_fp16' in amp_configs:
            raise ValueError(
1406
                "'use_pure_fp16' is an invalid parameter, the level of mixed precision training only depends on 'O1' or 'O2'."
J
Jiaqi Liu 已提交
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
            )

        _check_pure_fp16_configs()

        # construct amp_custom_lists
        if self._adapter._amp_level != 'O0' and amp_config_key_set:
            for param_name in [
                    'custom_white_list', 'custom_black_list',
                    'custom_black_varnames'
            ]:
                if param_name in amp_config_key_set:
                    self._adapter._amp_custom_lists[param_name] = amp_configs[
                        param_name]
                    amp_config_key_set -= {param_name}

        def _check_amp_configs(amp_config_key_set):
            accepted_param_set = {
                'init_loss_scaling',
                'incr_ratio',
                'decr_ratio',
                'incr_every_n_steps',
                'decr_every_n_nan_or_inf',
                'use_dynamic_loss_scaling',
                'use_fp16_guard',
            }
            if amp_config_key_set - accepted_param_set:
                raise ValueError(
1434 1435
                    "Except for 'level', the keys of 'amp_configs' must be accepted by mixed precision APIs, but {} could not be recognized.".
                    format(tuple(amp_config_key_set - accepted_param_set)))
J
Jiaqi Liu 已提交
1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451

            if 'use_fp16_guard' in amp_config_key_set:
                if in_dygraph_mode():
                    raise ValueError(
                        "'use_fp16_guard' is supported in static mode only.")
                self._adapter._use_fp16_guard = amp_configs['use_fp16_guard']
                amp_config_key_set.remove('use_fp16_guard')

            return amp_config_key_set

        amp_configs_set = _check_amp_configs(amp_config_key_set)
        for key in amp_configs_set:
            self._adapter._amp_configs[key] = amp_configs[key]

    def prepare(self, optimizer=None, loss=None, metrics=None,
                amp_configs=None):
1452 1453 1454 1455 1456 1457 1458
        """
        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.
1459 1460
            loss (Loss|callable function|None): Loss function can
                be a `paddle.nn.Layer` instance or any callable function
1461 1462
                taken the predicted values and ground truth values as input.
                It can be None when there is no loss.
1463 1464
            metrics (Metric|list of Metric|None): If metrics is set, all
                metrics will be calculated and output in train/eval mode.
J
Jiaqi Liu 已提交
1465 1466 1467 1468
            amp_configs (str|dict|None): AMP configurations. If AMP or pure
                float16 training is used, the key 'level' of 'amp_configs'
                should be set to 'O1' or 'O2' respectively. Otherwise, the
                value of 'level' defaults to 'O0', which means float32
1469 1470
                training. In addition to 'level', parameters consistent with
                mixed precision API could also be passed in. The supported
J
Jiaqi Liu 已提交
1471 1472 1473 1474
                keys are: 'init_loss_scaling', 'incr_ratio', 'decr_ratio',
                'incr_every_n_steps', 'decr_every_n_nan_or_inf',
                'use_dynamic_loss_scaling', 'custom_white_list',
                'custom_black_list', and 'custom_black_varnames'or
1475 1476 1477 1478 1479 1480
                'use_fp16_guard' is only supported in static mode. Mixed
                precision API documentations  :ref:`api_paddle_amp_auto_cast`
                and  :ref:`api_paddle_amp_GradScaler` could be referenced
                for details. For convenience, 'amp_configs' could be set to
                'O1' or 'O2' if no more parameters are needed. 'amp_configs'
                could be None in float32 training. Default: None.
1481 1482 1483 1484
        Returns:
            None
        """

1485 1486
        self._place = _get_device()
        if isinstance(self._place, fluid.CUDAPlace):
1487 1488 1489 1490 1491 1492 1493
            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()
1494
                    paddle.disable_static(self._place)
1495 1496 1497 1498 1499 1500 1501 1502 1503 1504
                    # 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
                else:
                    prepare_distributed_context(self._place)
                _parallel_context_initialized = True

        self._optimizer = optimizer
1505 1506
        if loss is not None:
            if not isinstance(loss, paddle.nn.Layer) and not callable(loss):
1507 1508 1509
                raise TypeError(
                    "'loss' must be sub classes of `paddle.nn.Layer` or any callable function."
                )
1510
        self._loss = loss
1511 1512 1513 1514 1515 1516 1517

        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)
J
Jiaqi Liu 已提交
1518
        self._prepare_amp(amp_configs)
1519 1520 1521 1522

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

1523
    def fit(self,
1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535
            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,
L
update  
lyuwenyu 已提交
1536
            callbacks=None,
1537 1538
            accumulate_grad_batches=1,
            num_iters=None):
1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580
        """
        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.
L
lyuwenyu 已提交
1581 1582
            accumulate_grad_batches (int): The number of batches to accumulate gradident 
                during training process before optimizer updates. It can mimic large batch
L
lyuwenyu 已提交
1583
                size. Default: 1.
1584 1585 1586
            num_iters (int|None): Integer number. The number of iterations to train
                the model. If None, follow `epochs` to train the model, otherwise, train
                the model `num_iters` times. Default: None.
L
lyuwenyu 已提交
1587
            
1588 1589 1590 1591 1592 1593 1594 1595 1596
        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

1597
              import paddle
1598
              import paddle.vision.transforms as T
1599
              from paddle.vision.datasets import MNIST
1600
              from paddle.static import InputSpec
1601 1602

              dynamic = True
1603 1604 1605
              if not dynamic:
                  paddle.enable_static()

1606 1607 1608 1609
              transform = T.Compose([
                  T.Transpose(),
                  T.Normalize([127.5], [127.5])
              ])
1610 1611
              train_dataset = MNIST(mode='train', transform=transform)
              val_dataset = MNIST(mode='test', transform=transform)
1612
           
1613 1614
              input = InputSpec([None, 1, 28, 28], 'float32', 'image')
              label = InputSpec([None, 1], 'int64', 'label')
1615
           
1616
              model = paddle.Model(
L
LielinJiang 已提交
1617
                  paddle.vision.models.LeNet(),
1618
                  input, label)
1619 1620
              optim = paddle.optimizer.Adam(
                  learning_rate=0.001, parameters=model.parameters())
1621 1622
              model.prepare(
                  optim,
1623
                  paddle.nn.CrossEntropyLoss(),
1624
                  paddle.metric.Accuracy(topk=(1, 2)))
1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635
              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

1636
              import paddle
1637
              import paddle.vision.transforms as T
1638
              from paddle.vision.datasets import MNIST
1639
              from paddle.static import InputSpec
1640 1641

              dynamic = True
1642 1643
              if not dynamic:
                  paddle.enable_static()
1644 1645 1646 1647 1648
              
              transform = T.Compose([
                    T.Transpose(),
                    T.Normalize([127.5], [127.5])
                ])
1649
              train_dataset = MNIST(mode='train', transform=transform)
1650
              train_loader = paddle.io.DataLoader(train_dataset,
1651 1652
                  batch_size=64)
              val_dataset = MNIST(mode='test', transform=transform)
1653
              val_loader = paddle.io.DataLoader(val_dataset,
1654
                  batch_size=64)
1655
           
1656 1657
              input = InputSpec([None, 1, 28, 28], 'float32', 'image')
              label = InputSpec([None, 1], 'int64', 'label')
1658
           
1659
              model = paddle.Model(
L
LielinJiang 已提交
1660
                  paddle.vision.models.LeNet(), input, label)
1661 1662
              optim = paddle.optimizer.Adam(
                  learning_rate=0.001, parameters=model.parameters())
1663 1664
              model.prepare(
                  optim,
1665
                  paddle.nn.CrossEntropyLoss(),
1666
                  paddle.metric.Accuracy(topk=(1, 2)))
1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706
              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
L
update  
lyuwenyu 已提交
1707

L
lyuwenyu 已提交
1708
        self._accumulate = accumulate_grad_batches
L
update  
lyuwenyu 已提交
1709

1710
        steps = self._len_data_loader(train_loader)
1711 1712 1713 1714 1715
        self.num_iters = num_iters
        if num_iters is not None and isinstance(num_iters, int):
            assert num_iters > 0, "num_iters must be greater than 0!"
            epochs = (num_iters // steps) + 1
            steps = min(num_iters, steps)
1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726
        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(), )

L
LiuChiachi 已提交
1727 1728 1729
        if any(isinstance(k, EarlyStopping) for k in cbks) and not do_eval:
            warnings.warn("EarlyStopping needs validation data.")

1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746
        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)
L
LiuChiachi 已提交
1747 1748
                if self.stop_training:
                    break
1749 1750 1751

        cbks.on_end('train', logs)
        self._test_dataloader = None
L
update  
lyuwenyu 已提交
1752

1753 1754 1755 1756 1757 1758 1759 1760
    def evaluate(self,
                 eval_data,
                 batch_size=1,
                 log_freq=10,
                 verbose=2,
                 num_workers=0,
                 callbacks=None,
                 num_iters=None):
1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781
        """
        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.
1782 1783 1784
            num_iters (int|None): Integer number. The number of iterations to
                evaluate the model. If None, evaluate on whole input dataset,
                otherwise, evaluate `num_iters` times. Default: None.
1785 1786 1787 1788 1789
        Returns:
            dict: Result of metric. The key is the names of Metric,
                value is a scalar or numpy.array.

        Examples:
1790 1791

          .. code-block:: python
1792

1793
            import paddle
1794
            import paddle.vision.transforms as T
1795
            from paddle.static import InputSpec
1796

1797
            # declarative mode
1798 1799 1800 1801 1802
            transform = T.Compose([
                    T.Transpose(),
                    T.Normalize([127.5], [127.5])
                ])
            val_dataset = paddle.vision.datasets.MNIST(mode='test', transform=transform)
1803

1804 1805 1806
            input = InputSpec([-1, 1, 28, 28], 'float32', 'image')
            label = InputSpec([None, 1], 'int64', 'label')
            model = paddle.Model(paddle.vision.models.LeNet(), input, label)
1807
            model.prepare(metrics=paddle.metric.Accuracy())
1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833
            result = model.evaluate(val_dataset, batch_size=64)
            print(result)
        """

        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)
1834 1835 1836 1837 1838
        self.num_iters = num_iters
        if num_iters is not None and isinstance(num_iters, int):
            assert num_iters > 0, "num_iters must be greater than 0!"
            eval_steps = min(num_iters, eval_steps)
            self.num_iters = eval_steps
1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859
        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,
1860
                verbose=1,
1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874
                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.
1875
            stack_outputs (bool): Whether stack output field like a batch, as for an output
1876 1877 1878 1879 1880
                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.
1881 1882
            verbose (int): The verbosity mode, should be 0, 1, or 2. 0 = silent,
                1 = progress bar, 2 = one line per batch. Default: 1.
1883
            callbacks(Callback): A Callback instance, default None.
1884

1885 1886 1887 1888
        Returns:
            list: output of models.

        Examples:
1889 1890

          .. code-block:: python
1891 1892

            import numpy as np
1893
            import paddle
1894
            from paddle.static import InputSpec
1895

1896
            class MnistDataset(paddle.vision.datasets.MNIST):
1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911
                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)

L
LielinJiang 已提交
1912
            # imperative mode
1913 1914
            input = InputSpec([-1, 1, 28, 28], 'float32', 'image')
            model = paddle.Model(paddle.vision.models.LeNet(), input)
1915
            model.prepare()
1916
            result = model.predict(test_dataset, batch_size=64)
1917
            print(len(result[0]), result[0][0].shape)
1918

L
LielinJiang 已提交
1919
            # declarative mode
1920
            device = paddle.set_device('cpu')
L
LielinJiang 已提交
1921 1922 1923
            paddle.enable_static()
            input = InputSpec([-1, 1, 28, 28], 'float32', 'image')
            model = paddle.Model(paddle.vision.models.LeNet(), input)
1924
            model.prepare()
L
LielinJiang 已提交
1925

1926 1927
            result = model.predict(test_dataset, batch_size=64)
            print(len(result[0]), result[0][0].shape)
1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943
        """

        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

1944
        cbks = config_callbacks(callbacks, model=self, verbose=verbose)
1945 1946 1947 1948

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

1949
        cbks.on_begin('predict', logs)
1950 1951 1952

        outputs = []

1953
        logs, outputs = self._run_one_epoch(test_loader, cbks, 'predict')
1954 1955 1956 1957 1958 1959 1960 1961 1962 1963

        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

1964
        cbks.on_end('predict', logs)
1965 1966
        return outputs

1967
    def _save_inference_model(self, path):
1968
        """
1969
        Save inference model can be used in static or dynamic mode.
1970 1971

        Args:
1972 1973
            path (str): The path prefix to save model. The format is
                ``dirname/file_prefix`` or ``file_prefix``.
1974
        Returns:
1975
            None
1976 1977
        """

1978
        if fluid.in_dygraph_mode():
1979 1980
            with fluid.framework._dygraph_guard(None):
                layer = self.network
L
LiuChiachi 已提交
1981
                if self._input_info is None:  # No provided or inferred
1982
                    raise RuntimeError(
L
LiuChiachi 已提交
1983
                        "Saving inference model needs 'inputs' or running before saving. Please specify 'inputs' in Model initialization or input training data and perform a training for shape derivation."
1984 1985 1986 1987
                    )
                if self._is_shape_inferred:
                    warnings.warn(
                        "'inputs' was not specified when Model initialization, so the input shape to be saved will be the shape derived from the user's actual inputs. The input shape to be saved is %s. For saving correct input shapes, please provide 'inputs' for Model initialization."
L
LiuChiachi 已提交
1988 1989
                        % self._input_info[0])

1990
                paddle.jit.save(layer, path, input_spec=self._inputs)
1991

1992
        else:
1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
            # path check
            file_prefix = os.path.basename(path)
            if file_prefix == "":
                raise ValueError(
                    "The input path MUST be format of dirname/file_prefix "
                    "[dirname\\file_prefix in Windows system], but received "
                    "file_prefix is empty string.")

            dirname = os.path.dirname(path)
            if dirname and not os.path.exists(dirname):
                os.makedirs(dirname)

            model_path = dirname
            model_filename = file_prefix + INFER_MODEL_SUFFIX
            params_filename = file_prefix + INFER_PARAMS_SUFFIX

2009 2010 2011 2012 2013 2014 2015 2016 2017
            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']

2018 2019
            fluid.io.save_inference_model(
                model_path,
2020 2021 2022 2023 2024
                input_names,
                endpoints,
                self._adapter._executor,
                main_program=infer_prog,
                model_filename=model_filename,
2025
                params_filename=params_filename)
2026

L
update  
lyuwenyu 已提交
2027 2028 2029 2030 2031 2032
    def _run_one_epoch(
            self,
            data_loader,
            callbacks,
            mode,
            logs={}, ):
2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048
        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
2049

2050 2051 2052 2053 2054
            batch_size = data[0].shape()[0] if callable(data[
                0].shape) else data[0].shape[0]

            callbacks.on_batch_begin(mode, step, logs)

2055
            if mode != 'predict':
L
update  
lyuwenyu 已提交
2056

L
lyuwenyu 已提交
2057 2058
                _inputs = [data[:len(self._inputs)], data[len(self._inputs):]]
                if mode == 'train':
L
lyuwenyu 已提交
2059 2060
                    _inputs.append((step + 1) % self._accumulate == 0 or
                                   step + 1 == len(data_loader))
L
update  
lyuwenyu 已提交
2061

L
lyuwenyu 已提交
2062
                outs = getattr(self, mode + '_batch')(*_inputs)
L
update  
lyuwenyu 已提交
2063

2064
                if self._metrics and self._loss:
2065
                    metrics = [[l[0] for l in outs[0]]]
2066
                elif self._loss:
2067 2068 2069
                    metrics = [[l[0] for l in outs]]
                else:
                    metrics = []
2070 2071 2072 2073 2074 2075 2076 2077 2078 2079

                # 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
2080
                if self._inputs is not None:
2081
                    outs = self.predict_batch(data[:len(self._inputs)])
L
LielinJiang 已提交
2082
                else:
2083
                    outs = self.predict_batch(data)
L
LielinJiang 已提交
2084

2085 2086 2087 2088 2089 2090 2091 2092 2093 2094
                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)
2095 2096 2097 2098
            if hasattr(self, 'num_iters') and self.num_iters is not None:
                self.num_iters -= 1
                if self.num_iters == 0:
                    break
2099 2100
        self._reset_metrics()

2101
        if mode == 'predict':
2102 2103 2104
            return logs, outputs
        return logs

L
LielinJiang 已提交
2105
    def summary(self, input_size=None, dtype=None):
L
LielinJiang 已提交
2106 2107 2108 2109 2110 2111 2112 2113
        """Prints a string summary of the network.

        Args:
            input_size (tuple|InputSpec|list[tuple|InputSpec], optional): size of input tensor. 
                    if not set, input_size will get from ``self._inputs`` if network only have 
                    one input, input_size can be tuple or InputSpec. if model have multiple 
                    input, input_size must be a list which contain every input's shape. 
                    Default: None.
2114
            dtype (str, optional): if dtype is None, 'float32' will be used, Default: None.
L
LielinJiang 已提交
2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127

        Returns:
            Dict: a summary of the network including total params and total trainable params.

        Examples:
            .. code-block:: python

              import paddle
              from paddle.static import InputSpec
           
              input = InputSpec([None, 1, 28, 28], 'float32', 'image')
              label = InputSpec([None, 1], 'int64', 'label')
           
2128
              model = paddle.Model(paddle.vision.models.LeNet(),
L
LielinJiang 已提交
2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139
                  input, label)
              optim = paddle.optimizer.Adam(
                  learning_rate=0.001, parameters=model.parameters())
              model.prepare(
                  optim,
                  paddle.nn.CrossEntropyLoss())

              params_info = model.summary()
              print(params_info)

        """
2140 2141 2142 2143 2144 2145
        assert (input_size is not None or self._inputs is not None
                ), "'input_size' or 'self._input' must be set"
        if input_size is not None:
            _input_size = input_size
        else:
            _input_size = self._inputs
L
LielinJiang 已提交
2146
        return summary(self.network, _input_size, dtype)
L
LielinJiang 已提交
2147

L
LiuChiachi 已提交
2148
    def _verify_spec(self, specs, shapes=None, dtypes=None, is_input=False):
2149 2150
        out_specs = []

2151 2152 2153 2154 2155 2156
        if specs is None:
            # 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?

            if is_input:
                arg_names = extract_args(self.network.forward)[1:]
L
LiuChiachi 已提交
2157 2158 2159
                # While Saving inference model in dygraph, and providing inputs only in running.
                if shapes is not None and dtypes is not None and fluid.in_dygraph_mode(
                ):
2160 2161
                    out_specs = [
                        Input(
L
LiuChiachi 已提交
2162
                            name=n, dtype=dtypes[i], shape=shapes[i])
2163 2164 2165 2166 2167 2168 2169
                        for i, n in enumerate(arg_names)
                    ]
                else:
                    out_specs = [Input(name=n, shape=[None]) for n in arg_names]
            else:
                out_specs = to_list(specs)
        elif isinstance(specs, dict):
2170 2171 2172 2173 2174
            assert is_input is False
            out_specs = [
                specs[n] for n in extract_args(self.network.forward)
                if n != 'self'
            ]
2175 2176 2177 2178 2179 2180 2181 2182
        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(
2183 2184
                        "Requires Input[{}].name != None, but receive `None` with {}."
                        .format(i, spec))
2185 2186 2187

        return out_specs

2188 2189 2190 2191 2192
    def _reset_metrics(self):
        for metric in self._metrics:
            metric.reset()

    def _metrics_name(self):
2193
        metrics_name = ['loss'] if self._loss else []
2194 2195 2196 2197 2198 2199 2200 2201 2202 2203
        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
L
LiuChiachi 已提交
2204 2205 2206

    def _update_inputs(self):
        "Update self._inputs according to given inputs."
L
LiuChiachi 已提交
2207 2208 2209 2210 2211
        self._input_info = self._adapter._input_info
        if self._input_info is not None and len(self._input_info) == 2:
            self._inputs = self._verify_spec(None, self._input_info[0],
                                             self._input_info[1], True)
            self._is_shape_inferred = True