Tokenizer.word_index did not contain "START" or "END", rather contained "start" and "end"

I was trying to make an Image Captioning model in a similar fashion as in here I used ResNet50 instead off VGG16 and also had to use progressive loading via model.fit_generator() method. I used ResNet50 from here and when I imported it by setting include_top = False, It gave me features of photo in shape of {'key': [[[[value1, value2, .... value 2048]]]]}, where "key" is the image id. Here's my code of captionGenerator function:-

def createCaptions(tokenizer, photoData, MaxLength, model):
    for key, feature in photoData.items():
        inSeq = "START"
        for i in range(MaxLength):
            sequence = tokenizer.texts_to_sequences([inSeq])[0]
            sequence = pad_sequences([sequence], maxlen = MaxLength)
            ID = model.predict([np.array(feature[0][0][0]), sequence])
            ID = np.argmax(ID)
            ID = word_for_id(ID)
            if ID is None:
                break
            inSeq += " " + ID
            if ID == "END":
                break
        print(inSeq)

The function word_for_id is :-

def word_for_id(integer, tokenizer):
    for word, index in tokenizer.word_index.items():
        if index == integer:
            return word
    return None

I had generated photoData via:-

features = {}
for images in os.listdir(args["image"]):
    filename = args["image"] + '/' + images
    image = load_img(filename, target_size = inputShape)
    image = img_to_array(image)
    image = np.expand_dims(image, axis = 0)
    image = preprocess(image)
    pred = resnet.predict(image)
    image_id = images.split('.')[0]
    features[image_id] = pred
    print('>{}'.format(images))

features is my photoData dictionary.

The problem is, in training data photos descriptions which I generate through:-

def train_test_data(filename):
    DataFile = open(filename, 'r')
    Data = DataFile.read()
    DataFile.close()

    ImageID = []

    textDataFile = pickle.load(open('descriptions.pkl', 'rb'))

    for line in Data.split('\n'):
        if len(line) < 1:
            continue
        ImageID.append(line.split('.')[0])

    Data = {}

    for key in textDataFile:
        if key in ImageID:
            Data[key] = textDataFile[key]

    for ID in Data:
        for i in range(len(Data[ID])):
            l = Data[ID][i]
            l = "START " + " ".join(l) + " END"
            Data[ID][i] = l

    return Data

Here, I added "START" and "END" at the begginning and end of each sentences of description respectively. But in tokenizer.word_index, "START" and "END" are not found as keys. That is:-

k = pickle.load(open('word_index.pkl', 'rb'))
print("START" in k)

This gives result as False. Please explain to me why this is happening. If I do:-

k = pickle.load(open('word_index.pkl', 'rb'))
print("start" in k)

The answer comes out True.

That is because by default the Tokenizer lowers the words when fitting based on the lower=True parameter. You can either use the lower case or pass lower=False when creating the tokenizer, documentation.