Python chainer.training.extensions 模块，snapshot_object() 实例源码

def train(args):
    model = EmbeddingTagger(args.model, 50, 20, 30)
    model.setup_training(args.embed)
    if args.initmodel:
        print('Load model from', args.initmodel)
        chainer.serializers.load_npz(args.initmodel, model)

    train = CCGBankDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = CCGBankDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.SGD(lr=0.01)
    optimizer.setup(model)
    updater = training.StandardUpdater(train_iter, optimizer)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 5000, 'iteration'
    log_interval = 200, 'iteration'
    val_model = model.copy()

    trainer.extend(extensions.Evaluator(val_iter, val_model), trigger=val_interval)
    trainer.extend(extensions.dump_graph('main/loss'))
    trainer.extend(extensions.snapshot(), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration', 'main/loss', 'validation/main/loss',
        'main/accuracy', 'validation/main/accuracy',
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def pretrain_source_cnn(data, args, epochs=1000):
    print(":: pretraining source encoder")
    source_cnn = Loss(num_classes=10)
    if args.device >= 0:
        source_cnn.to_gpu()

    optimizer = chainer.optimizers.Adam()
    optimizer.setup(source_cnn)

    train_iterator, test_iterator = data2iterator(data, args.batchsize, multiprocess=False)

    # train_iterator = chainer.iterators.MultiprocessIterator(data, args.batchsize, n_processes=4)

    updater = chainer.training.StandardUpdater(iterator=train_iterator, optimizer=optimizer, device=args.device)
    trainer = chainer.training.Trainer(updater, (epochs, 'epoch') ,out=args.output)

    # learning rate decay
    # trainer.extend(extensions.ExponentialShift("alpha", rate=0.9, init=args.learning_rate, target=args.learning_rate*10E-5))

    trainer.extend(extensions.Evaluator(test_iterator, source_cnn, device=args.device))
    # trainer.extend(extensions.snapshot(filename='snapshot_epoch_{.updater.epoch}'), trigger=(10, "epoch"))
    trainer.extend(extensions.snapshot_object(optimizer.target, "source_model_epoch_{.updater.epoch}"), trigger=(epochs, "epoch"))

    trainer.extend(extensions.ProgressBar(update_interval=10))
    trainer.extend(extensions.LogReport(trigger=(1, "epoch")))
    trainer.extend(extensions.PrintReport(
        ['epoch', 'main/loss', 'validation/main/loss',
         'main/accuracy', 'validation/main/accuracy', 'elapsed_time']))

    trainer.run()

    return source_cnn

def train_target_cnn(source, target, source_cnn, target_cnn, args, epochs=10000):
    print(":: training encoder with target domain")
    discriminator = Discriminator()

    if args.device >= 0:
        source_cnn.to_gpu()
        target_cnn.to_gpu()
        discriminator.to_gpu()

    # target_optimizer = chainer.optimizers.Adam(alpha=1.0E-5, beta1=0.5)
    target_optimizer = chainer.optimizers.RMSprop(lr=args.lr)
    # target_optimizer = chainer.optimizers.MomentumSGD(lr=1.0E-4, momentum=0.99)
    target_optimizer.setup(target_cnn.encoder)
    target_optimizer.add_hook(chainer.optimizer.WeightDecay(args.weight_decay))

    # discriminator_optimizer = chainer.optimizers.Adam(alpha=1.0E-5, beta1=0.5)
    discriminator_optimizer = chainer.optimizers.RMSprop(lr=args.lr)
    # discriminator_optimizer = chainer.optimizers.MomentumSGD(lr=1.0E-4, momentum=0.99)
    discriminator_optimizer.setup(discriminator)
    discriminator_optimizer.add_hook(chainer.optimizer.WeightDecay(args.weight_decay))

    source_train_iterator, source_test_iterator = data2iterator(source, args.batchsize, multiprocess=False)
    target_train_iterator, target_test_iterator = data2iterator(target, args.batchsize, multiprocess=False)

    updater = ADDAUpdater(source_train_iterator, target_train_iterator, source_cnn, target_optimizer, discriminator_optimizer, args)

    trainer = chainer.training.Trainer(updater, (epochs, 'epoch'), out=args.output)

    trainer.extend(extensions.Evaluator(target_test_iterator, target_cnn, device=args.device))
    # trainer.extend(extensions.snapshot(filename='snapshot_epoch_{.updater.epoch}'), trigger=(10, "epoch"))
    trainer.extend(extensions.snapshot_object(target_cnn, "target_model_epoch_{.updater.epoch}"), trigger=(epochs, "epoch"))

    trainer.extend(extensions.ProgressBar(update_interval=10))
    trainer.extend(extensions.LogReport(trigger=(1, "epoch")))
    trainer.extend(extensions.PrintReport(
        ["epoch", "loss/discrim", "loss/encoder",
         "validation/main/loss", "validation/main/accuracy", "elapsed_time"]))

    trainer.run()

def train_main(args):
    """
    trains model specfied in args.
    main method for train subcommand.
    """
    # load text
    with open(args.text_path) as f:
        text = f.read()
    logger.info("corpus length: %s.", len(text))

    # data iterator
    data_iter = DataIterator(text, args.batch_size, args.seq_len)

    # load or build model
    if args.restore:
        logger.info("restoring model.")
        load_path = args.checkpoint_path if args.restore is True else args.restore
        model = load_model(load_path)
    else:
        net = Network(vocab_size=VOCAB_SIZE,
                      embedding_size=args.embedding_size,
                      rnn_size=args.rnn_size,
                      num_layers=args.num_layers,
                      drop_rate=args.drop_rate)
        model = L.Classifier(net)

    # make checkpoint directory
    log_dir = make_dirs(args.checkpoint_path)
    with open("{}.json".format(args.checkpoint_path), "w") as f:
        json.dump(model.predictor.args, f, indent=2)
    chainer.serializers.save_npz(args.checkpoint_path, model)
    logger.info("model saved: %s.", args.checkpoint_path)

    # optimizer
    optimizer = chainer.optimizers.Adam(alpha=args.learning_rate)
    optimizer.setup(model)
    # clip gradient norm
    optimizer.add_hook(chainer.optimizer.GradientClipping(args.clip_norm))

    # trainer
    updater = BpttUpdater(data_iter, optimizer)
    trainer = chainer.training.Trainer(updater, (args.num_epochs, 'epoch'), out=log_dir)
    trainer.extend(extensions.snapshot_object(model, filename=os.path.basename(args.checkpoint_path)))
    trainer.extend(extensions.ProgressBar(update_interval=1))
    trainer.extend(extensions.LogReport())
    trainer.extend(extensions.PlotReport(y_keys=["main/loss"]))
    trainer.extend(LoggerExtension(text))

    # training start
    model.predictor.reset_state()
    logger.info("start of training.")
    time_train = time.time()
    trainer.run()

    # training end
    duration_train = time.time() - time_train
    logger.info("end of training, duration: %ds.", duration_train)
    # generate text
    seed = generate_seed(text)
    generate_text(model, seed, 1024, 3)
    return model

def create_trainer(
        config,
        project_path,
        updater,
        model,
        eval_func,
        iterator_test,
        iterator_train_varidation,
        loss_names,
        converter=chainer.dataset.convert.concat_examples,
):
    # type: (TrainConfig, str, any, typing.Dict, any, any, any, any, any) -> any
    def _make_evaluator(iterator):
        return utility.chainer_utility.NoVariableEvaluator(
            iterator,
            target=model,
            converter=converter,
            eval_func=eval_func,
            device=config.gpu,
        )

    trainer = chainer.training.Trainer(updater, out=project_path)

    log_trigger = (config.log_iteration, 'iteration')
    save_trigger = (config.save_iteration, 'iteration')

    eval_test_name = 'eval/test'
    eval_train_name = 'eval/train'

    snapshot = extensions.snapshot_object(model['main'], '{.updater.iteration}.model')
    trainer.extend(snapshot, trigger=save_trigger)

    trainer.extend(extensions.dump_graph('main/' + loss_names[0], out_name='main.dot'))

    trainer.extend(_make_evaluator(iterator_test), name=eval_test_name, trigger=log_trigger)
    trainer.extend(_make_evaluator(iterator_train_varidation), name=eval_train_name, trigger=log_trigger)

    report_target = []
    for evaluator_name in ['', eval_test_name + '/', eval_train_name + '/']:
        for model_name in ['main/']:
            for loss_name in loss_names:
                report_target.append(evaluator_name + model_name + loss_name)

    trainer.extend(extensions.LogReport(trigger=log_trigger, log_name="log.txt"))
    trainer.extend(extensions.PrintReport(report_target))

    return trainer

def train(args):
    model = LSTMParser(args.model, args.word_emb_size, args.afix_emb_size, args.nlayers,
            args.hidden_dim, args.elu_dim, args.dep_dim, args.dropout_ratio)
    with open(args.model + "/params", "w") as f: log(args, f)

    if args.initmodel:
        print 'Load model from', args.initmodel
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print 'Load pretrained word embeddings from', args.pretrained
        model.load_pretrained_embeddings(args.pretrained)

    if args.gpu >= 0:
        chainer.cuda.get_device(args.gpu).use()
        model.to_gpu()

    train = LSTMParserDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = LSTMParserDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.Adam(beta2=0.9)
    # optimizer = chainer.optimizers.MomentumSGD(momentum=0.7)
    optimizer.setup(model)
    optimizer.add_hook(WeightDecay(1e-6))
    # optimizer.add_hook(GradientClipping(5.))
    updater = training.StandardUpdater(train_iter, optimizer,
            device=args.gpu, converter=converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 1000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.Evaluator(val_iter, eval_model,
                    converter, device=args.gpu), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration',
        'main/tagging_accuracy', 'main/tagging_loss',
        'main/parsing_accuracy', 'main/parsing_loss',
        'validation/main/tagging_accuracy',
        'validation/main/parsing_accuracy'
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def train(args):
    model = LSTMTagger(args.model, args.word_emb_size, args.afix_emb_size,
            args.nlayers, args.hidden_dim, args.relu_dim, args.dropout_ratio)
    with open(args.model + "/params", "w") as f:
            log(args, f)
    if args.initmodel:
        print('Load model from', args.initmodel)
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print('Load pretrained word embeddings from', args.pretrained)
        model.load_pretrained_embeddings(args.pretrained)

    if args.gpu >= 0:
        chainer.cuda.get_device(args.gpu).use()
        model.to_gpu()

    train = LSTMTaggerDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = LSTMTaggerDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.MomentumSGD(momentum=0.7)
    optimizer.setup(model)
    optimizer.add_hook(WeightDecay(1e-6))
    optimizer.add_hook(GradientClipping(5.))
    updater = training.StandardUpdater(train_iter, optimizer,
            device=args.gpu, converter=converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 2000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.Evaluator(
        val_iter, eval_model, converter, device=args.gpu), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration', 'main/loss', 'validation/main/loss',
        'main/accuracy', 'validation/main/accuracy',
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def train(args):
    model = BiaffineJaLSTMParser(args.model, args.word_emb_size, args.char_emb_size,
            args.nlayers, args.hidden_dim, args.dep_dim, args.dropout_ratio)
    with open(args.model + "/params", "w") as f: log(args, f)

    if args.initmodel:
        print('Load model from', args.initmodel)
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print('Load pretrained word embeddings from', args.pretrained)
        model.load_pretrained_embeddings(args.pretrained)

    if args.gpu >= 0:
        chainer.cuda.get_device(args.gpu).use()
        model.to_gpu()

    train = LSTMParserDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = LSTMParserDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.Adam(beta2=0.9)
    # optimizer = chainer.optimizers.MomentumSGD(momentum=0.7)
    optimizer.setup(model)
    optimizer.add_hook(WeightDecay(2e-6))
    # optimizer.add_hook(GradientClipping(5.))
    updater = training.StandardUpdater(train_iter, optimizer,
            device=args.gpu, converter=converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 1000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.ExponentialShift(
                    "eps", .75, 2e-3), trigger=(2500, 'iteration'))
    trainer.extend(extensions.Evaluator(
        val_iter, eval_model, converter, device=args.gpu), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration',
        'main/tagging_accuracy', 'main/tagging_loss',
        'main/parsing_accuracy', 'main/parsing_loss',
        'validation/main/tagging_accuracy',
        'validation/main/parsing_accuracy'
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def train(args):
    model = JaLSTMParser(args.model, args.word_emb_size, args.char_emb_size,
            args.nlayers, args.hidden_dim, args.relu_dim, args.dep_dim, args.dropout_ratio)
    with open(args.model + "/params", "w") as f: log(args, f)

    if args.initmodel:
        print('Load model from', args.initmodel)
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print('Load pretrained word embeddings from', args.pretrained)
        model.load_pretrained_embeddings(args.pretrained)

    train = LSTMParserDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = LSTMParserDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.Adam(beta2=0.9)
    # optimizer = chainer.optimizers.MomentumSGD(momentum=0.7)
    optimizer.setup(model)
    optimizer.add_hook(WeightDecay(1e-6))
    # optimizer.add_hook(GradientClipping(5.))
    updater = training.StandardUpdater(train_iter, optimizer, converter=converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 1000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.Evaluator(
        val_iter, eval_model, converter), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration', 'main/tagging_loss',
        'main/tagging_accuracy', 'main/tagging_loss',
        'main/parsing_accuracy', 'main/parsing_loss',
        'validation/main/tagging_loss', 'validation/main/tagging_accuracy',
        'validation/main/parsing_loss', 'validation/main/parsing_accuracy'
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def train(args):
    model = LSTMTagger(args.model, args.word_emb_size, args.char_emb_size,
            args.nlayers, args.hidden_dim, args.relu_dim, args.dropout_ratio)
    with open(args.model + "/params", "w") as f:
            log(args, f)
    if args.initmodel:
        print('Load model from', args.initmodel)
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print('Load pretrained word embeddings from', args.pretrained)
        model.load_pretrained_embeddings(args.pretrained)

    train = LSTMTaggerDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = LSTMTaggerDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.MomentumSGD(momentum=0.7)
    optimizer.setup(model)
    optimizer.add_hook(WeightDecay(1e-6))
    # optimizer.add_hook(GradientClipping(5.))
    updater = training.StandardUpdater(train_iter, optimizer, converter=converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 1000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.Evaluator(
        val_iter, eval_model, converter), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration', 'main/loss', 'validation/main/loss',
        'main/accuracy', 'validation/main/accuracy',
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def train(args):
    model = LSTMTagger(args.model, args.word_emb_size, args.afix_emb_size,
            args.nlayers, args.hidden_dim, args.relu_dim, args.dropout_ratio)
    with open(args.model + "/params", "w") as f:
            log(args, f)
    if args.initmodel:
        print 'Load model from', args.initmodel
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print 'Load pretrained word embeddings from', args.pretrained
        model.load_pretrained_embeddings(args.pretrained)

    if args.gpu >= 0:
        chainer.cuda.get_device(args.gpu).use()
        model.to_gpu()

    train = LSTMTaggerDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = LSTMTaggerDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.MomentumSGD(momentum=0.7)
    optimizer.setup(model)
    optimizer.add_hook(WeightDecay(1e-6))
    optimizer.add_hook(GradientClipping(5.))
    updater = training.StandardUpdater(train_iter, optimizer,
            device=args.gpu, converter=converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 2000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.Evaluator(
        val_iter, eval_model, converter, device=args.gpu), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration', 'main/loss', 'validation/main/loss',
        'main/accuracy', 'validation/main/accuracy',
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def train(args):
    model = LSTMParser(args.model, args.word_emb_size, args.afix_emb_size, args.nlayers,
            args.hidden_dim, args.elu_dim, args.dep_dim, args.dropout_ratio)
    with open(args.model + "/params", "w") as f: log(args, f)

    if args.initmodel:
        print('Load model from', args.initmodel)
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print('Load pretrained word embeddings from', args.pretrained)
        model.load_pretrained_embeddings(args.pretrained)

    if args.gpu >= 0:
        chainer.cuda.get_device(args.gpu).use()
        model.to_gpu()

    train = LSTMParserDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = LSTMParserDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.Adam(beta2=0.9)
    # optimizer = chainer.optimizers.MomentumSGD(momentum=0.7)
    optimizer.setup(model)
    optimizer.add_hook(WeightDecay(1e-6))
    # optimizer.add_hook(GradientClipping(5.))
    updater = training.StandardUpdater(train_iter, optimizer,
            device=args.gpu, converter=converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 1000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.Evaluator(val_iter, eval_model,
                    converter, device=args.gpu), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration',
        'main/tagging_accuracy', 'main/tagging_loss',
        'main/parsing_accuracy', 'main/parsing_loss',
        'validation/main/tagging_accuracy',
        'validation/main/parsing_accuracy'
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def train(args):
    model = JaCCGEmbeddingTagger(args.model,
                args.word_emb_size, args.char_emb_size)
    if args.initmodel:
        print('Load model from', args.initmodel)
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print('Load pretrained word embeddings from', args.pretrained)
        model.load_pretrained_embeddings(args.pretrained)

    train = JaCCGTaggerDataset(args.model, args.train)
    train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
    val = JaCCGTaggerDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.AdaGrad()
    optimizer.setup(model)
    # optimizer.add_hook(WeightDecay(1e-8))
    my_converter = lambda x, dev: convert.concat_examples(x, dev, (None,-1,None,None))
    updater = training.StandardUpdater(train_iter, optimizer, converter=my_converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 1000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.Evaluator(
        val_iter, eval_model, my_converter), trigger=val_interval)
    trainer.extend(extensions.dump_graph('main/loss'))
    trainer.extend(extensions.snapshot(), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration', 'main/loss', 'validation/main/loss',
        'main/accuracy', 'validation/main/accuracy',
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def train(args):
    model = PeepHoleLSTMTagger(args.model, args.word_emb_size, args.afix_emb_size,
            args.nlayers, args.hidden_dim, args.relu_dim, args.dropout_ratio)
    with open(args.model + "/params", "w") as f:
            log(args, f)
    if args.initmodel:
        print('Load model from', args.initmodel)
        chainer.serializers.load_npz(args.initmodel, model)

    if args.pretrained:
        print('Load pretrained word embeddings from', args.pretrained)
        model.load_pretrained_embeddings(args.pretrained)

    if args.gpu >= 0:
        chainer.cuda.get_device(args.gpu).use()
        model.to_gpu()

    converter = lambda x, device: \
            concat_examples(x, device=device, padding=-1)

    train = LSTMTaggerDataset(args.model, args.train)
    train_iter = SerialIterator(train, args.batchsize)
    val = LSTMTaggerDataset(args.model, args.val)
    val_iter = chainer.iterators.SerialIterator(
            val, args.batchsize, repeat=False, shuffle=False)
    optimizer = chainer.optimizers.MomentumSGD(momentum=0.7)
    optimizer.setup(model)
    optimizer.add_hook(WeightDecay(1e-6))
    optimizer.add_hook(GradientClipping(5.))
    updater = training.StandardUpdater(train_iter, optimizer,
            device=args.gpu, converter=converter)
    trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.model)

    val_interval = 1000, 'iteration'
    log_interval = 200, 'iteration'

    eval_model = model.copy()
    eval_model.train = False

    trainer.extend(extensions.Evaluator(
        val_iter, eval_model, converter, device=args.gpu), trigger=val_interval)
    trainer.extend(extensions.snapshot_object(
        model, 'model_iter_{.updater.iteration}'), trigger=val_interval)
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.PrintReport([
        'epoch', 'iteration', 'main/loss', 'validation/main/loss',
        'main/accuracy', 'validation/main/accuracy',
    ]), trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.run()

def start(self):
        """ Train pose net. """
        # set random seed.
        if self.seed is not None:
            random.seed(self.seed)
            np.random.seed(self.seed)
            if self.gpu >= 0:
                chainer.cuda.cupy.random.seed(self.seed)
        # initialize model to train.
        model = AlexNet(self.Nj, self.use_visibility)
        if self.resume_model:
            serializers.load_npz(self.resume_model, model)
        # prepare gpu.
        if self.gpu >= 0:
            chainer.cuda.get_device(self.gpu).use()
            model.to_gpu()
        # load the datasets.
        train = PoseDataset(self.train, data_augmentation=self.data_augmentation)
        val = PoseDataset(self.val, data_augmentation=False)
        # training/validation iterators.
        train_iter = chainer.iterators.MultiprocessIterator(
            train, self.batchsize)
        val_iter = chainer.iterators.MultiprocessIterator(
            val, self.batchsize, repeat=False, shuffle=False)
        # set up an optimizer.
        optimizer = self._get_optimizer()
        optimizer.setup(model)
        if self.resume_opt:
            chainer.serializers.load_npz(self.resume_opt, optimizer)
        # set up a trainer.
        updater = training.StandardUpdater(train_iter, optimizer, device=self.gpu)
        trainer = training.Trainer(
            updater, (self.epoch, 'epoch'), os.path.join(self.out, 'chainer'))
        # standard trainer settings
        trainer.extend(extensions.dump_graph('main/loss'))
        val_interval = (10, 'epoch')
        trainer.extend(TestModeEvaluator(val_iter, model, device=self.gpu), trigger=val_interval)
        # save parameters and optimization state per validation step
        resume_interval = (self.epoch/10, 'epoch')
        trainer.extend(extensions.snapshot_object(
            model, "epoch-{.updater.epoch}.model"), trigger=resume_interval)
        trainer.extend(extensions.snapshot_object(
            optimizer, "epoch-{.updater.epoch}.state"), trigger=resume_interval)
        trainer.extend(extensions.snapshot(
            filename="epoch-{.updater.epoch}.iter"), trigger=resume_interval)
        # show log
        log_interval = (10, "iteration")
        trainer.extend(extensions.LogReport(trigger=log_interval))
        trainer.extend(extensions.observe_lr(), trigger=log_interval)
        trainer.extend(extensions.PrintReport(
            ['epoch', 'main/loss', 'validation/main/loss', 'lr']), trigger=log_interval)
        trainer.extend(extensions.ProgressBar(update_interval=10))
        # start training
        if self.resume:
            chainer.serializers.load_npz(self.resume, trainer)
        trainer.run()

def __init__(self, folder, chain, train, test, batchsize=500, resume=True, gpu=0, nepoch=1, reports=[]):
        self.reports = reports
        self.nepoch = nepoch
        self.folder = folder
        self.chain = chain
        self.gpu = gpu

        if self.gpu >= 0:
            chainer.cuda.get_device(gpu).use()
            chain.to_gpu(gpu)
        self.eval_chain = eval_chain = chain.copy()
        self.chain.test = False
        self.eval_chain.test = True
        self.testset = test

        if not os.path.exists(folder):
            os.makedirs(folder)

        train_iter = chainer.iterators.SerialIterator(train, batchsize, shuffle=True)
        test_iter = chainer.iterators.SerialIterator(test, batchsize,
                                                     repeat=False, shuffle=False)

        updater = training.StandardUpdater(train_iter, chain.optimizer, device=gpu)
        trainer = training.Trainer(updater, (nepoch, 'epoch'), out=folder)
        # trainer.extend(TrainingModeSwitch(chain))
        trainer.extend(extensions.dump_graph('main/loss'))
        trainer.extend(extensions.Evaluator(test_iter, eval_chain, device=gpu), trigger=(1,'epoch'))
        trainer.extend(extensions.snapshot_object(
            chain, 'chain_snapshot_epoch_{.updater.epoch:06}'), trigger=(1,'epoch'))
        trainer.extend(extensions.snapshot(
            filename='snapshot_epoch_{.updater.epoch:06}'), trigger=(1,'epoch'))
        trainer.extend(extensions.LogReport(trigger=(1,'epoch')), trigger=(1,'iteration'))
        trainer.extend(extensions.PrintReport(
            ['epoch']+reports), trigger=IntervalTrigger(1,'epoch'))

        self.trainer = trainer

        if resume:
            #if resumeFrom is not None:
            #    trainerFile = os.path.join(resumeFrom[0],'snapshot_epoch_{:06}'.format(resumeFrom[1]))
            #    S.load_npz(trainerFile, trainer)
            i = 1
            trainerFile = os.path.join(folder,'snapshot_epoch_{:06}'.format(i))
            while i <= nepoch and os.path.isfile(trainerFile):
                i = i + 1
                trainerFile = os.path.join(folder,'snapshot_epoch_{:06}'.format(i))
            i = i - 1
            trainerFile = os.path.join(folder,'snapshot_epoch_{:06}'.format(i))
            if i >= 0 and os.path.isfile(trainerFile):
                S.load_npz(trainerFile, trainer)

def create_trainer(
        config: TrainConfig,
        project_path: str,
        updater,
        model: typing.Dict,
        eval_func,
        iterator_test,
        iterator_train_eval,
        loss_names,
        converter=chainer.dataset.convert.concat_examples,
        log_name='log.txt',
):
    trainer = chainer.training.Trainer(updater, out=project_path)

    log_trigger = (config.log_iteration, 'iteration')
    save_trigger = (config.save_iteration, 'iteration')

    eval_test_name = 'eval/test'
    eval_train_name = 'eval/train'

    snapshot = extensions.snapshot_object(model['encoder'], 'encoder{.updater.iteration}.model')
    trainer.extend(snapshot, trigger=save_trigger)
    snapshot = extensions.snapshot_object(model['generator'], 'generator{.updater.iteration}.model')
    trainer.extend(snapshot, trigger=save_trigger)
    snapshot = extensions.snapshot_object(model['mismatch_discriminator'], 'mismatch_discriminator{.updater.iteration}.model')
    trainer.extend(snapshot, trigger=save_trigger)

    trainer.extend(utility.chainer.dump_graph([
        'encoder/' + loss_names[0],
        'generator/' + loss_names[0],
        'mismatch_discriminator/' + loss_names[0],
    ], out_name='main.dot'))

    def _make_evaluator(iterator):
        return utility.chainer.NoVariableEvaluator(
            iterator,
            target=model,
            converter=converter,
            eval_func=eval_func,
            device=config.gpu,
        )

    trainer.extend(_make_evaluator(iterator_test), name=eval_test_name, trigger=log_trigger)
    trainer.extend(_make_evaluator(iterator_train_eval), name=eval_train_name, trigger=log_trigger)

    report_target = []
    for evaluator_name in ['', eval_test_name + '/', eval_train_name + '/']:
        for model_name in [s + '/' for s in model.keys()]:
            for loss_name in set(loss_names):
                report_target.append(evaluator_name + model_name + loss_name)

    trainer.extend(extensions.LogReport(trigger=log_trigger, log_name=log_name))
    trainer.extend(extensions.PrintReport(report_target))

    return trainer