config.py

class Config(object):
    """ Wrapper class for various (hyper)parameters. """
    def __init__(self):
        # about the model architecture
        self.cnn = 'rpn'               # 'vgg16' or 'resnet50'
        self.max_caption_length = 20
        self.dim_embedding = 384
        self.num_lstm_units = 384
        self.num_initalize_layers = 2    # 1 or 2
        self.dim_initalize_layer = 512
        self.num_attend_layers = 2       # 1 or 2
        self.dim_attend_layer = 512
        self.num_decode_layers = 2       # 1 or 2
        self.dim_decode_layer = 1024

        # about the weight initialization and regularization
        self.fc_kernel_initializer_scale = 0.08
        self.fc_kernel_regularizer_scale = 1e-4
        self.fc_activity_regularizer_scale = 0.0
        self.conv_kernel_regularizer_scale = 1e-4
        self.conv_activity_regularizer_scale = 0.0
        self.fc_drop_rate = 0.5
        self.lstm_drop_rate = 0.3
        self.attention_loss_factor = 0.01

        # about the optimization
        self.num_epochs = 7
        self.batch_size = 16
        self.optimizer = 'Adam'    # 'Adam', 'RMSProp', 'Momentum' or 'SGD'
        self.initial_learning_rate = 0.0001
        self.learning_rate_decay_factor = 1.0
        self.num_steps_per_decay = 100000
        self.clip_gradients = 5.0
        self.momentum = 0.0
        self.use_nesterov = True
        self.decay = 0.9
        self.centered = True
        self.beta1 = 0.9
        self.beta2 = 0.999
        self.epsilon = 1e-6

        # about the saver
        self.save_period = 1000
        self.save_dir = '/output/models/'
        self.summary_period = 100
        self.summary_dir = '/output/summary/'

        # about the vocabulary
        self.vocabulary_file = '/data/zisang0210/flickr8k-ann/vocabulary-flickr.csv'
        self.vocabulary_size = 5000

        # about the training
        self.dataset = 'flickr'
        self.caption_file =  '/output/Flickr8k_text/Flickr8k.token.txt'
        
        self.train_image_dir = '/output/Flickr8k_Dataset/Flicker8k_Dataset/'
        self.train_image_file = '/output/Flickr8k_text/Flickr_8k.trainImages.txt'
        self.temp_annotation_file = '/data/zisang0210/flickr8k-ann/anns_flickr.csv'
        self.temp_data_file = '/data/zisang0210/flickr8k-ann/data_flickr.npy'

        # about the evaluation
        self.eval_image_dir = '/output/Flickr8k_Dataset/Flicker8k_Dataset/'
        self.eval_image_file = '/output/Flickr8k_text/Flickr_8k.testImages.txt'
        self.eval_result_dir = './val/flickr-results/'
        self.eval_result_file = './val/flickr-results/results.json'
        self.save_eval_result_as_image = False

        # about the testing
        self.test_image_dir = './test/images/'
        self.test_result_dir = './test/results/'
        self.test_result_file = './test/results.csv'