Hands-On Transfer Learning with Python

Implementing a beam search-based caption generator

We will now be implementing a basic beam search-based algorithm to generate caption sequences:

from keras.preprocessing import image, sequence 
 
def get_raw_caption_sequences(model, word_to_index, image_features,  max_caption_size, beam_size=1): start = [word_to_index['<START>']] caption_seqs = [[start, 0.0]] while len(caption_seqs[0][0]) < max_caption_size: temp_caption_seqs = [] for caption_seq in caption_seqs: partial_caption_seq = sequence.pad_sequences( [caption_seq[0]], maxlen=max_caption_size, padding='post') next_words_pred = model.predict( [np.asarray([image_features]), np.asarray(partial_caption_seq)])[0] next_words = np.argsort(next_words_pred)[-beam_size:] for word in next_words: ...