show-attend-and-tell-tensorflow源码解读：preprocess.py

from scipy import ndimage
# scipy.ndimage: Multi-dimentional image processing(多维图像处理包） 更强大的图像处理库包括：opencv, scikit-image等
from collections import Counter
# collections模块包含多种集合类，Counter是其中之一，它是一个简单的计数器，统计字符出现的个数，是dict的一个子类
from core.vggnet import Vgg19
# core.vggnet.Vgg19类从imagenet-vgg-verydeep-19.mat中获取了预训练参数，用这些预训练参数构造了vgg19网络的计算模型。
from core.utils import *
import tensorflow as tf
import numpy as np
import pandas as pd
import hickle
# hickle与pickle都是常用的序列化/反序列化模块，用来保存程序运行结果或者加载包含程序所需信息的文件。
import os
# os模块用于程序与操作系统交互，访问文件夹
import json
# json和hickle，pickle作用类似

以上是preprocess.py的import信息。因为该python文件包含多个python函数，内容较多，需要理清头绪，找出程序入口，程序入口即main()函数。

def main():
    # batch size for extracting feature vectors from vggnet
    batch_size = 100    # 一次提取100幅图像的feature vectors
    # maximum length of caption (number of word). if caption is longer than max_length, deleted.
    max_length = 15    # 标签语句最长15个单词，超过15个单词的语句删掉
    # if word occurs less than word_count_threshold in training dataset, the word index is special unknown token.
    word_count_threshold = 1 # 如果训练集中某个单词出现次数小于1，那就设为null（一个特殊的token）
    # vgg model path
    vgg_model_path = './data/imagenet-vgg-verydeep-19.mat'
    
    # about 80000 images and 400000 captions for train dataset
    train_dataset = _process_caption_data(caption_file='data/annotations/captions_train2014.json, image_dir='image/train2014_resized', max_length=max_length)
    # 有图像文件夹image_dir，有包含标签语句和图像与标签的连接信息的caption_file，这个函数（后面详细介绍）事实上构建了训练集变量，另外一点：./data/ == data/

    # about 40000 images and 200000 captions
    val_dataset = _process_caption_data(caption_file='data/annotations/captions_val2014.json', image_dir='image/val_resized', max_length=max_length)
    # 这里构建了验证集变量

    # about 4000 images and 20000 captions for val / test dataset
    val_cutoff = int(0.1 * len(val_dataset))
    test_cutoff = int(0.2 * len(val_dataset))
    print('Finished processing caption data')

    save_pickle(train_dataset, 'data/train/train.annotations.pkl')
    save_pickle(val_dataset[:val_cutoff], 'data/val/val.annotations.pkl')
    save_pickle(val_dataset[val_cutoff:test_cutoff].reset_index(drop=True), 'data/test/test.annotations.pkl')
    """
        这里save_pickle()函数与pickle模块有关，pickle模块保存的文件后缀名都是pkl，save_pickle()是对pickle.dump()函数的扩展，它的定义在core.utils模块中(前面导入模块中已经写了)。
        reset_index()方法的全称是pandas.DataFrame.reset_index()，用来防止原索引变成数据列。可见_process_caption_data返回的结果是pd.DataFrame类的实例，但疑点是为什么前两个没用该方法？
        从这儿开始，对上面得到的train, val, test三个文件，逐个执行相关操作。
    """
    
    for split in ['train', 'val', 'test']:
        annotations = load_pickle('./data/%s/%s.annotations.pkl' % (split, split))
        # load_pickle()与save_pickle()情形相似，都位于core.utils模块中(core/utils.py文件中)，都是对pickle模块中的函数进行扩展，不同之处在于load_pickle()扩展的是pickle.load()
        
        if split == 'train':
            word_to_idx = _build_vocab(annotations=annotations, threshold=word_count_threshold)
            # 在training阶段，制作词汇表，方便后续的one-hot词编码和词嵌入。
            save_pickle(word_to_idx, '.data/%s/word_to_idx.pkl' % split)    # 把词汇表保存起来
        captions = _build_caption_vector(annotations=annotations, word_to_idx=word_to_idx, max_length=max_length)
        # 制作好词汇表后，对整个句子进行编码
        save_pickle(captions, './data/%s/%s.captions.pkl' % (split, split))    # 对句子编码向量保存起来。
        file_names, id_to_idx = _build_file_names(annotations)
        save_pickle(file_names, './data/%s/%s.file.names.pkl' % (split, split))
        image_idxs = _build_image_idxs(annotations, id_to_idx)
        save_pickle(image_idxs, './data/%s/%s.image.idxs.pkl' % (split, split))
        """这四句暂时不清楚具体干了啥，但应该是提取了图片文件名，图片id，标签语句，标签编码之间的关系"""
        # prepare reference captions to compute bleu scores later
        image_ids = {}
        feature_to_captions = {}
        i = -1
        for caption, image_id in zip(annotations['caption'], annotations['image_id']):
            if not image_id in image_ids:
                image_ids[image_id] = 0
                i += 1
                feature_to_captions[i] = []
            feature_to_captions[i].append(caption.lower() + ' .')
        save_pickle(feature_to_captions, './data/%s/%s.references.pkl' % (split, split))
        print("finished building %s caption dataset" % split)

    # extract conv5_3 feature vectors
    vggnet = Vgg19(vgg_model_path)
    # 加载预训练的模型参数
    vggnet.build()
    # 加载后构建vgg19模型，得到完整的计算流程
    with tf.Session() as sess:
        tf.initialize_all_variables().run()
        for split in ['train', 'val', 'test']:
            anno_path = './data/%s/%s.annotations.pkl' % (split, split)
            save_path = './data/%s/%s.features.hkl' % (split, split)
            annotations = load_pickle(anno_path)
            image_path = list(annotations['file_name'].unique())
            n_example = len(image_path)
            
            all_feats = np.ndarray([n_example, 196, 512], dtype=np.float32)
            
            for start, end in zip(range(0, n_example, batch_size), range(batch_size, n_example+batch_size, batch_size)):
                image_batch_file = image_path[staart:end]
                image_batch = np.array(map(lambda x: ndimage.imread(x, mode='RGB'), image_batch_file)).astype(np.float32)
                feats = sess.run(vggnet.features, feed_dict={vggnet.images:image_batch})
                all_feats[start:end, :] = feats
                print("Processed %d %s features.." % (end, split))
            
            # use hickle to save huge feature vectors
            hickle.dump(all_feats, save_path)
            print("Saved %s.." % (save_path))