当前位置: 首页>>代码示例>>Python>>正文


Python Preprocessor.load_sample方法代码示例

本文整理汇总了Python中preprocessor.Preprocessor.load_sample方法的典型用法代码示例。如果您正苦于以下问题:Python Preprocessor.load_sample方法的具体用法?Python Preprocessor.load_sample怎么用?Python Preprocessor.load_sample使用的例子?那么, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在preprocessor.Preprocessor的用法示例。


在下文中一共展示了Preprocessor.load_sample方法的1个代码示例,这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞,您的评价将有助于系统推荐出更棒的Python代码示例。

示例1: Bird

# 需要导入模块: from preprocessor import Preprocessor [as 别名]
# 或者: from preprocessor.Preprocessor import load_sample [as 别名]

#.........这里部分代码省略.........
        
        
    # loads and randomizes data
    def load_data(self):
        (paths, labels) = self.load_labels(self.label_path,self.label_bg_path)
        self.load_meta_data(self.meta_path)
        
        nr_files = len(paths)

        mask = np.arange(nr_files)

        np.random.shuffle(mask)

        train_size = int(nr_files * (1 - self.train_val_ratio))

        paths = np.array(paths)[mask]
        labels = np.array(labels)[mask]

        self.class_weights = {}
        for i in range(self.nb_species):
            weight_mask = labels == str(i)  #  np.equal(labels, i*np.ones(labels.shape))
            nb_class = np.sum(weight_mask)
            if nb_class == 0:
                print("No data for class", str(i))
                continue
            self.class_weights[i] = nr_files/np.sum(weight_mask)

        self.paths = paths[:train_size]
        self.labels = labels[:train_size]
        self.nr_files = train_size

        self.val_paths = paths[train_size:]
        self.val_labels = labels[train_size:]
        self.nr_val_files = (nr_files - train_size) // self.batch_size * self.batch_size


    def train_data_generator(self):
        while True:
            specs = []
            labels = []
            for i in range(self.batch_size):
                (spec, label) = self.get_random_training_sample()
                specs.append(np.array([spec]).transpose((1, 2, 0)))
                labels.append(np.array([label]))

            yield (np.array(specs), np.array(labels))


    def val_data_generator(self):
        specs = []
        labels = []
        for val_path, val_label in zip(self.val_paths, self.val_labels):
            sample = self.preprocessor.load_sample(val_path)
            if np.max(sample[0]) <= 0:
                continue
            spec = self.preprocessor.preprocess(sample)
            # spec = self.augmenter.augment_transform(spec, val_label)
            specs.append(np.array([spec[0]]).transpose((1, 2, 0)))
            labels.append(np.array([val_label]))
            if len(specs) == self.batch_size:
                yield (np.array(specs), np.array(labels))
                specs = []
                labels = []

        if len(specs) > 0:
            yield (np.array(specs), np.array(labels))


    # loads a single new training sample from disc. 
    # preprocesses and augments the training sample.
    def get_random_training_sample(self):
        r = random.randint(0, self.nr_files - 1)
        path = self.paths[r]
        label = self.labels[r]
        sample = self.preprocessor.load_sample(path)
        if np.max(sample[0]) <= 0:
            return self.get_random_training_sample()
        spec = self.preprocessor.preprocess(sample)
        spec = self.augmenter.augment_transform(spec, label)
        return (spec[0], label)

    
    # start training process
    def train(self):
        self.load_data()
        self.model = models.model_fg_bg(self.nb_species,
                                        (self.nb_f_steps, self.nb_t_steps))
        sgd = SGD(lr=0.01, decay=0.0, momentum=0.9, nesterov=True)
        self.model.compile(loss='sparse_categorical_crossentropy',optimizer=sgd,metrics=['accuracy'])
        
        self.model.summary()

        modelCheckpoint = keras.callbacks.ModelCheckpoint("/" + self.training_description + "/{epoch:02d}-{val_loss:.2f}.hdf5")
        reduce_lr = keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=2, verbose=1,
                                                      min_lr=0.0001, epsilon=1e-5)

        history = self.model.fit_generator(self.train_data_generator(), samples_per_epoch=self.nr_files,
                                           nb_epoch=self.nr_epoch, verbose=1, max_q_size=self.batch_size,
                                           validation_data=self.val_data_generator(), nb_val_samples=self.nr_val_files,
                                           nb_worker=4, pickle_safe=True, callbacks=[modelCheckpoint, reduce_lr])
开发者ID:FluxB,项目名称:BirdSongClassifier,代码行数:104,代码来源:train.py


注:本文中的preprocessor.Preprocessor.load_sample方法示例由纯净天空整理自Github/MSDocs等开源代码及文档管理平台,相关代码片段筛选自各路编程大神贡献的开源项目,源码版权归原作者所有,传播和使用请参考对应项目的License;未经允许,请勿转载。