program.py

# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
#
# 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 os
import time
import datetime
from collections import OrderedDict

import paddle
from paddle import to_tensor
import paddle.nn as nn
import paddle.nn.functional as F

from ppcls.optimizer import LearningRateBuilder
from ppcls.optimizer import OptimizerBuilder
from ppcls.modeling import architectures
from ppcls.modeling.loss import MultiLabelLoss
from ppcls.modeling.loss import CELoss
from ppcls.modeling.loss import MixCELoss
from ppcls.modeling.loss import JSDivLoss
from ppcls.modeling.loss import GoogLeNetLoss
from ppcls.utils.misc import AverageMeter
from ppcls.utils import logger
from ppcls.utils import multi_hot_encode
from ppcls.utils import hamming_distance
from ppcls.utils import accuracy_score


def create_model(architecture, classes_num):
    """
    Create a model

    Args:
        architecture(dict): architecture information,
            name(such as ResNet50) is needed
        image(variable): model input variable
        classes_num(int): num of classes

    Returns:
        out(variable): model output variable
    """
    name = architecture["name"]
    params = architecture.get("params", {})
    return architectures.__dict__[name](class_dim=classes_num, **params)


def create_loss(feeds,
                out,
                architecture,
                classes_num=1000,
                epsilon=None,
                use_mix=False,
                use_distillation=False,
                multilabel=False):
    """
    Create a loss for optimization, such as:
        1. CrossEnotry loss
        2. CrossEnotry loss with label smoothing
        3. CrossEnotry loss with mix(mixup, cutmix, fmix)
        4. CrossEnotry loss with label smoothing and (mixup, cutmix, fmix)
        5. GoogLeNet loss

    Args:
        out(variable): model output variable
        feeds(dict): dict of model input variables
        architecture(dict): architecture information,
            name(such as ResNet50) is needed
        classes_num(int): num of classes
        epsilon(float): parameter for label smoothing, 0.0 <= epsilon <= 1.0
        use_mix(bool): whether to use mix(include mixup, cutmix, fmix)

    Returns:
        loss(variable): loss variable
    """
    if architecture["name"] == "GoogLeNet":
        assert len(out) == 3, "GoogLeNet should have 3 outputs"
        loss = GoogLeNetLoss(class_dim=classes_num, epsilon=epsilon)
        return loss(out[0], out[1], out[2], feeds["label"])

    if use_distillation:
        assert len(out) == 2, ("distillation output length must be 2, "
                               "but got {}".format(len(out)))
        loss = JSDivLoss(class_dim=classes_num, epsilon=epsilon)
        return loss(out[1], out[0])

    if use_mix:
        loss = MixCELoss(class_dim=classes_num, epsilon=epsilon)
        feed_y_a = feeds['y_a']
        feed_y_b = feeds['y_b']
        feed_lam = feeds['lam']
        return loss(out, feed_y_a, feed_y_b, feed_lam)
    else:
        if not multilabel:
            loss = CELoss(class_dim=classes_num, epsilon=epsilon)
        else:
            loss = MultiLabelLoss(class_dim=classes_num, epsilon=epsilon)
        return loss(out, feeds["label"])


def create_metric(out,
                  label,
                  architecture,
                  topk=5,
                  classes_num=1000,
                  use_distillation=False,
                  multilabel=False,
                  mode="train"):
    """
    Create measures of model accuracy, such as top1 and top5

    Args:
        out(variable): model output variable
        feeds(dict): dict of model input variables(included label)
        topk(int): usually top5
        classes_num(int): num of classes
        use_distillation(bool): whether to use distillation training
        mode(str): mode, train/valid

    Returns:
        fetchs(dict): dict of measures
    """
    if architecture["name"] == "GoogLeNet":
        assert len(out) == 3, "GoogLeNet should have 3 outputs"
        out = out[0]
    else:
        # just need student label to get metrics
        if use_distillation:
            out = out[1]
    softmax_out = F.softmax(out)

    fetchs = OrderedDict()
    metric_names = set()
    if not multilabel:
        softmax_out = F.softmax(out)

        # set top1 to fetchs
        top1 = paddle.metric.accuracy(softmax_out, label=label, k=1)
        # set topk to fetchs
        k = min(topk, classes_num)
        topk = paddle.metric.accuracy(softmax_out, label=label, k=k)

        metric_names.add("top1")
        metric_names.add("top{}".format(k))

        fetchs['top1'] = top1
        topk_name = "top{}".format(k)
        fetchs[topk_name] = topk
    else:
        out = F.sigmoid(out)
        preds = multi_hot_encode(out.numpy())
        targets = label.numpy()
        ham_dist = to_tensor(hamming_distance(preds, targets))
        accuracy = to_tensor(accuracy_score(preds, targets, base="label"))

        ham_dist_name = "hamming_distance"
        accuracy_name = "multilabel_accuracy"
        metric_names.add(ham_dist_name)
        metric_names.add(accuracy_name)

        fetchs[accuracy_name] = accuracy
        fetchs[ham_dist_name] = ham_dist

    # multi cards' eval
    if mode != "train" and paddle.distributed.get_world_size() > 1:
        for metric_name in metric_names:
            fetchs[metric_name] = paddle.distributed.all_reduce(
                fetchs[metric_name], op=paddle.distributed.ReduceOp.
                SUM) / paddle.distributed.get_world_size()

    return fetchs


def create_fetchs(feeds, net, config, mode="train"):
    """
    Create fetchs as model outputs(included loss and measures),
    will call create_loss and create_metric(if use_mix).

    Args:
        out(variable): model output variable
        feeds(dict): dict of model input variables.
            If use mix_up, it will not include label.
        architecture(dict): architecture information,
            name(such as ResNet50) is needed
        topk(int): usually top5
        classes_num(int): num of classes
        epsilon(float): parameter for label smoothing, 0.0 <= epsilon <= 1.0
        use_mix(bool): whether to use mix(include mixup, cutmix, fmix)

    Returns:
        fetchs(dict): dict of model outputs(included loss and measures)
    """
    architecture = config.ARCHITECTURE
    topk = config.topk
    classes_num = config.classes_num
    epsilon = config.get('ls_epsilon')
    use_mix = config.get('use_mix') and mode == 'train'
    use_distillation = config.get('use_distillation')
    multilabel = config.get('multilabel', False)

    out = net(feeds["image"])

    fetchs = OrderedDict()
    fetchs['loss'] = create_loss(feeds, out, architecture, classes_num,
                                 epsilon, use_mix, use_distillation,
                                 multilabel)
    if not use_mix:
        metric = create_metric(
            out,
            feeds["label"],
            architecture,
            topk,
            classes_num,
            use_distillation,
            multilabel=multilabel,
            mode=mode)
        fetchs.update(metric)

    return fetchs


def create_optimizer(config, parameter_list=None):
    """
    Create an optimizer using config, usually including
    learning rate and regularization.

    Args:
        config(dict):  such as
        {
            'LEARNING_RATE':
                {'function': 'Cosine',
                 'params': {'lr': 0.1}
                },
            'OPTIMIZER':
                {'function': 'Momentum',
                 'params':{'momentum': 0.9},
                 'regularizer':
                    {'function': 'L2', 'factor': 0.0001}
                }
        }

    Returns:
        an optimizer instance
    """
    # create learning_rate instance
    lr_config = config['LEARNING_RATE']
    lr_config['params'].update({
        'epochs': config['epochs'],
        'step_each_epoch':
        config['total_images'] // config['TRAIN']['batch_size'],
    })
    lr = LearningRateBuilder(**lr_config)()

    # create optimizer instance
    opt_config = config['OPTIMIZER']
    opt = OptimizerBuilder(**opt_config)
    return opt(lr, parameter_list), lr


def create_feeds(batch, use_mix, num_classes, multilabel=False):
    image = batch[0]
    if use_mix:
        y_a = to_tensor(batch[1].numpy().astype("int64").reshape(-1, 1))
        y_b = to_tensor(batch[2].numpy().astype("int64").reshape(-1, 1))
        lam = to_tensor(batch[3].numpy().astype("float32").reshape(-1, 1))
        feeds = {"image": image, "y_a": y_a, "y_b": y_b, "lam": lam}
    else:
        if not multilabel:
            label = to_tensor(batch[1].numpy().astype("int64").reshape(-1, 1))
        else:
            label = to_tensor(batch[1].numpy().astype('float32').reshape(-1, num_classes))
        feeds = {"image": image, "label": label}
    return feeds


total_step = 0


def run(dataloader,
        config,
        net,
        optimizer=None,
        lr_scheduler=None,
        epoch=0,
        mode='train',
        vdl_writer=None):
    """
    Feed data to the model and fetch the measures and loss

    Args:
        dataloader(paddle dataloader):
        exe():
        program():
        fetchs(dict): dict of measures and the loss
        epoch(int): epoch of training or validation
        model(str): log only

    Returns:
    """
    print_interval = config.get("print_interval", 10)
    use_mix = config.get("use_mix", False) and mode == "train"
    multilabel = config.get("multilabel", False)
    classes_num = config.get("classes_num")

    metric_list = [
        ("loss", AverageMeter(
            'loss', '7.5f', postfix=",")),
        ("lr", AverageMeter(
            'lr', 'f', postfix=",", need_avg=False)),
        ("batch_time", AverageMeter(
            'batch_cost', '.5f', postfix=" s,")),
        ("reader_time", AverageMeter(
            'reader_cost', '.5f', postfix=" s,")),
    ]
    if not use_mix:
        if not multilabel:
            topk_name = 'top{}'.format(config.topk)
            metric_list.insert(
                0, (topk_name, AverageMeter(
                    topk_name, '.5f', postfix=",")))
            metric_list.insert(
                0, ("top1", AverageMeter(
                    "top1", '.5f', postfix=",")))
        else:
            metric_list.insert(0, ("multilabel_accuracy", AverageMeter(
                                   "multilabel_accuracy", '.5f', postfix=",")))
            metric_list.insert(0, ("hamming_distance", AverageMeter(
                                   "hamming_distance", '.5f', postfix=",")))

    metric_list = OrderedDict(metric_list)

    tic = time.time()
    for idx, batch in enumerate(dataloader()):
        # avoid statistics from warmup time
        if idx == 10:
            metric_list["batch_time"].reset()
            metric_list["reader_time"].reset()

        metric_list['reader_time'].update(time.time() - tic)
        batch_size = len(batch[0])
        feeds = create_feeds(batch, use_mix, classes_num, multilabel)
        fetchs = create_fetchs(feeds, net, config, mode)
        if mode == 'train':
            avg_loss = fetchs['loss']
            avg_loss.backward()

            optimizer.step()
            optimizer.clear_grad()
            lr_value = optimizer._global_learning_rate().numpy()[0]
            metric_list['lr'].update(lr_value, batch_size)

            if lr_scheduler is not None:
                if lr_scheduler.update_specified:
                    curr_global_counter = lr_scheduler.step_each_epoch * epoch + idx
                    update = max(
                        0, curr_global_counter - lr_scheduler.update_start_step
                    ) % lr_scheduler.update_step_interval == 0
                    if update:
                        lr_scheduler.step()
                else:
                    lr_scheduler.step()

        for name, fetch in fetchs.items():
            metric_list[name].update(fetch.numpy()[0], batch_size)
        metric_list["batch_time"].update(time.time() - tic)
        tic = time.time()

        if vdl_writer and mode == "train":
            global total_step
            logger.scaler(
                name="lr", value=lr_value, step=total_step, writer=vdl_writer)
            for name, fetch in fetchs.items():
                logger.scaler(
                    name="train_{}".format(name),
                    value=fetch.numpy()[0],
                    step=total_step,
                    writer=vdl_writer)
            total_step += 1

        fetchs_str = ' '.join([
            str(metric_list[key].mean)
            if "time" in key else str(metric_list[key].value)
            for key in metric_list
        ])

        if idx % print_interval == 0:
            ips_info = "ips: {:.5f} images/sec".format(
                batch_size / metric_list["batch_time"].avg)

            if mode == "train":
                epoch_str = "epoch:{:<3d}".format(epoch)
                step_str = "{:s} step:{:<4d}".format(mode, idx)
                eta_sec = ((config["epochs"] - epoch) * len(dataloader) - idx
                           ) * metric_list["batch_time"].avg
                eta_str = "eta: {:s}".format(
                    str(datetime.timedelta(seconds=int(eta_sec))))
                logger.info("{:s}, {:s}, {:s} {:s}, {:s}".format(
                    epoch_str, step_str, fetchs_str, ips_info, eta_str))
            else:
                logger.info("{:s} step:{:<4d}, {:s} {:s}".format(
                    mode, idx, fetchs_str, ips_info))

    end_str = ' '.join([str(m.mean) for m in metric_list.values()] +
                       [metric_list['batch_time'].total])
    ips_info = "ips: {:.5f} images/sec.".format(
        batch_size * metric_list["batch_time"].count /
        metric_list["batch_time"].sum)

    if mode == 'eval':
        logger.info("END {:s} {:s} {:s}".format(mode, end_str, ips_info))
    else:
        end_epoch_str = "END epoch:{:<3d}".format(epoch)
        logger.info("{:s} {:s} {:s} {:s}".format(end_epoch_str, mode, end_str,
                                                 ips_info))

    # return top1_acc in order to save the best model
    if mode == 'valid':
        if multilabel:
            return metric_list['multilabel_accuracy'].avg
        else:
            return metric_list['top1'].avg