def load_model(embedding_size, vocab_size, num_steps, dataset):
reverse_filename = get_reverse_filename(embedding_size, num_steps,
vocab_size, dataset)
embeddings_filename = get_embeddin_filename(embedding_size, num_steps,
vocab_size, dataset)
reverse_dictionary = load_pickle_file(reverse_filename)
final_embeddings = load_pickle_file(embeddings_filename)
dict_filename = get_dict_filename(embedding_size, num_steps, vocab_size,
dataset)
dictionary = load_pickle_file(dict_filename)
return reverse_dictionary, final_embeddings, dictionary
def hist_plotter():
colors = ['#F95400', '#0C56A2', '#F9DC00', '#00A670', '#C60074']
seqgan_dict = load_pickle_file(
"/Users/markus/workspace/master/Master/seqgan_score_dict.p")
emb_dict = load_pickle_file(
"/Users/markus/workspace/master/Master/emb_score_dict.p")
color_seqgan = colors.pop(0)
color_emb = colors.pop(0)
seqgan_intervals = []
seqgan_intervals_uniq = []
for (s, (b, n)) in seqgan_dict.iteritems():
seqgan_intervals_uniq.append(b)
for _ in range(n):
seqgan_intervals.append(b)
emb_intervals = []
emb_intervals_uniq = []
for (s, (b, n)) in emb_dict.iteritems():
emb_intervals_uniq.append(b)
for _ in range(n):
emb_intervals.append(b)
num_bins = 10
fig, ax = plt.subplots()
plt.rc('font', family='Arial')
# the histogram of the data
data = np.vstack([seqgan_intervals, emb_intervals]).T
ax.hist(data,
num_bins,
color=[color_seqgan, color_emb],
label=["Baseline", "Our model"])
# ax.hist(data_uniq, num_bins)
# ax.hist(emb_intervals, num_bins, normed=1)
# add a 'best fit' line
ax.set_xlabel(u'β')
ax.set_ylabel('Count')
# Tweak spacing to prevent clipping of ylabel
fig.tight_layout()
plt.legend()
plt.show()
def get_word_embeddings(conf):
if conf.WORD_EMBEDDING_METHOD == 'glove':
embeddings_index = {}
f = open('data/embeddings/glove.6B.300d.txt')
count = 0
for line in f:
values = line.split()
word = values[0]
coefs = np.asarray(values[1:], dtype='float32')
embeddings_index[word] = coefs
count += 1
if count % 100 == 0:
print_progress(count,
400000,
prefix="Getting glove word embeddings")
f.close()
return embeddings_index
elif conf.WORD_EMBEDDING_METHOD == 'word2vec':
embedding_dict_name = "word2vec/saved_models/word2vec_%sd%svoc100001steps_dict_%s.pkl" % (
conf.EMBEDDING_DIMENSION, conf.NB_WORDS,
conf.DATASET if conf.DATASET != None else "flickr")
return load_pickle_file(embedding_dict_name)
print("WORD_EMBEDDING_METHOD not found")
return None
def preprocess_sentences(config, sentences):
sos_token = "<sos>"
eos_token = "<eos>"
pad_token = "<pad>"
if config[Conf.WORD_EMBEDDING] == WordEmbedding.GLOVE:
print "Loading Glove dictionary..."
word_embedding_dict = get_word_embeddings()
sos_token = "<"
eos_token = ">"
pad_token = "="
else:
filename = get_dict_filename(config[Conf.EMBEDDING_SIZE],
config[Conf.WORD2VEC_NUM_STEPS],
config[Conf.VOCAB_SIZE],
config[Conf.W2V_SET])
print "Loading Word2Vec dictionary (%s)..." % filename
# word_embedding_dict = load_pickle_file("word2vec/saved_models/word2vec_%sd%svoc%ssteps_dict.pkl" % (config[Conf.EMBEDDING_SIZE], config[Conf.VOCAB_SIZE], config[Conf.WORD2VEC_NUM_STEPS]))
word_embedding_dict = load_pickle_file(filename)
word_list_sentences = []
for sentence in sentences:
word_list = [sos_token]
for word in sentence.split(" "):
word_list.append(word.lower())
word_list.append(eos_token)
while len(word_list) < config[Conf.MAX_SEQ_LENGTH]:
word_list.append(pad_token)
word_list_sentences.append(word_list)
# word_list_sentences = [[word.lower() for word in sentence.split(" ")] for sentence in sentences]
return np.asarray(word_list_sentences), word_embedding_dict
def get_word_embedding_matrix(word_to_id, embedding_dim):
embeddings_dict = load_pickle_file(
'word2vec/saved_models/word2vec_50d1000voc100001steps_dict_flowers.pkl'
)
embedding_matrix = numpy.zeros((len(word_to_id) + 1, embedding_dim))
for word, i in word_to_id.items():
if word in embeddings_dict:
embedding_matrix[i] = embeddings_dict[word]
def generate_input_noise(config):
if config[Conf.PREINIT] == PreInit.ENCODER_DECODER:
if config[Conf.WORD_EMBEDDING] == WordEmbedding.ONE_HOT:
noise_size = config[Conf.VOCAB_SIZE]
else:
noise_size = config[Conf.EMBEDDING_SIZE]
else:
noise_size = config[Conf.NOISE_SIZE]
if config[Conf.NOISE_MODE] == NoiseMode.REPEAT:
noise_matrix = np.zeros(
(config[Conf.BATCH_SIZE], config[Conf.MAX_SEQ_LENGTH], noise_size))
for batch_index in range(config[Conf.BATCH_SIZE]):
word_noise = np.random.normal(size=noise_size)
for word_index in range(config[Conf.MAX_SEQ_LENGTH]):
noise_matrix[batch_index][word_index] = word_noise
return noise_matrix
elif config[Conf.NOISE_MODE] == NoiseMode.REPEAT_SINGLE:
noise_matrix = np.zeros((config[Conf.BATCH_SIZE], noise_size))
for batch_index in range(config[Conf.BATCH_SIZE]):
noise_matrix[batch_index] = np.random.normal(size=noise_size)
return noise_matrix
elif config[Conf.NOISE_MODE] == NoiseMode.NEW:
return np.random.rand(config[Conf.BATCH_SIZE],
config[Conf.MAX_SEQ_LENGTH], noise_size)
elif config[Conf.NOISE_MODE] == NoiseMode.FIRST_ONLY:
noise_matrix = np.zeros(
(config[Conf.BATCH_SIZE], config[Conf.MAX_SEQ_LENGTH], noise_size))
for batch_index in range(config[Conf.BATCH_SIZE]):
word_noise = np.random.normal(size=noise_size)
for word_index in range(config[Conf.MAX_SEQ_LENGTH]):
noise_matrix[batch_index][word_index] = word_noise
for batch_index in range(config[Conf.BATCH_SIZE]):
if random.random() < 0.5:
word_noise = np.zeros(noise_size)
else:
word_noise = np.ones(noise_size)
noise_matrix[batch_index][0] = word_noise
return noise_matrix
elif config[Conf.NOISE_MODE] == NoiseMode.ONES:
return np.ones(
(config[Conf.BATCH_SIZE], config[Conf.MAX_SEQ_LENGTH], noise_size))
elif config[Conf.NOISE_MODE] == NoiseMode.ENCODING:
embedded_data = load_pickle_file(
"sequence_to_sequence/logs/S2S_2EMB_2017-04-04_VS2+1000_BS128_HD30_DHL1_ED50_SEQ5_WEMword2vec/encoded_data.pkl"
)
random_distribution_of_embedded_data = []
for i in range(config[Conf.BATCH_SIZE]):
# random_distribution_of_embedded_data.append(embedded_data[np.random.randint(0, len(embedded_data))])
random_distribution_of_embedded_data.append(embedded_data[i])
return np.asarray(random_distribution_of_embedded_data)
def compare_distributions():
perplexity = 15
data = "encoded"
tsne = TSNE(perplexity=perplexity, n_components=2, init='pca', n_iter=5000)
suffix = "_lambda"
# suffix = ""
if data == "encoded":
embs = load_pickle_file(
"sequence_to_sequence/logs/NORM_S2S_2EMB_2017-04-07_VS2+1000_BS128_HD40_DHL1_ED50_SEQ5_WEMword2vec/encoded_data_lambda.pkl")
embs = embs[:1000, 0, :]
elif data == "word2vec":
embs = load_pickle_file("word2vec/saved_models/word2vec_50d1000voc100001steps_embs.pkl")
embs = embs[:1000, :]
random_uniform = numpy.random.normal(size=embs.shape)
# embs = normalize(embs, norm="l2")
append = numpy.append(random_uniform, embs, axis=0)
embs_pca = tsne.fit_transform(append)
plot_collections([embs_pca[:1000], embs_pca[1000:]], ["gaussian-yellow", data + "-blue"], perplexity, suffix)
def gen_class_embs():
# create_common_words_pickle()
print "Generating classes"
common_words = load_pickle_file("common_words.p")
print "Loading captions..."
filename_caption_text_tuples = fetch_all_caption_text_tuples()[:5000]
print "Loading word embeddings..."
word_embedding_dict = dict(fetch_all_word_vectors())
filname_text_vector_tuples = []
tot = len(filename_caption_text_tuples)
counter = 1
print_progress(counter, tot, prefix="Converting classes to embs")
for filename, caption in filename_caption_text_tuples:
classes = get_classes(caption, common_words)
filname_text_vector_tuples.extend([(filename, c, word_embedding_dict[c]) for c in classes if c in word_embedding_dict.keys()])
counter += 1
print_progress(counter, tot, prefix="Converting classes to embs")
save_class_vector_list(filname_text_vector_tuples)
def wmd_retrieval(pred_strindgs, dataset_string_list_sentences):
filename = get_dict_filename(config[Conf.EMBEDDING_SIZE],
config[Conf.WORD2VEC_NUM_STEPS],
config[Conf.VOCAB_SIZE], config[Conf.W2V_SET])
word_embedding_dict = load_pickle_file(filename)
best_sentence_lists = []
for pred_string in pred_strings:
score_tuples = []
for dataset_string_list_sentence in dataset_string_list_sentences:
dataset_string = " ".join(dataset_string_list_sentence)
score = get_wmd_distance(pred_string, dataset_string,
word_embedding_dict)
score_tuples.append((dataset_string, score))
score_tuples = sorted(score_tuples, key=lambda x: x[1], reverse=False)
result = [x[0] for x in score_tuples[:5]]
best_sentence_lists.append(result)
return best_sentence_lists
def background_wmd_retrieval(pred_strings, dataset_string_list_sentences):
filename = get_dict_filename(config[Conf.EMBEDDING_SIZE],
config[Conf.WORD2VEC_NUM_STEPS],
config[Conf.VOCAB_SIZE], config[Conf.W2V_SET])
word_embedding_dict = load_pickle_file(filename)
counter = Value('i', 0)
sentence_count = len(pred_strings)
cpu_count = multiprocessing.cpu_count()
print "CPUs:", cpu_count
if cpu_count > 8 and cpu_count < 15:
cpu_count = 10
elif cpu_count > 40:
cpu_count = 40
print "Starting pool with %s processes" % cpu_count
pool = Pool(cpu_count, initializer=init, initargs=(counter, ))
tuple_array = [(pred_string, dataset_string_list_sentences,
word_embedding_dict, sentence_count)
for pred_string in pred_strings]
best_sentence_lists = pool.map(background_wmd, tuple_array, chunksize=1)
pool.close()
pool.join()
return best_sentence_lists
def plotter():
colors = ['#F95400', '#0C56A2', '#F9DC00', '#00A670', '#C60074']
seqgan_dict = load_pickle_file(
"/Users/markus/workspace/master/Master/seqgan_score_dict.p")
emb_dict = load_pickle_file(
"/Users/markus/workspace/master/Master/emb_score_dict.p")
color_seqgan = colors.pop(0)
colors.pop(0)
color_emb = colors.pop(0)
plt.rc('font', family='Arial')
buckets = 10
plt.rcParams.update({'font.size': 20})
seqgan_count = [0 for _ in range(buckets)]
seqgan_count_uniq = [0 for _ in range(buckets)]
emb_count = [0 for _ in range(buckets)]
emb_count_uniq = [0 for _ in range(buckets)]
for (n, (c, u)) in seqgan_dict.iteritems():
seqgan_count[int(c * 10 - 1)] += u
seqgan_count_uniq[int(c * 10 - 1)] += 1
for (n, (c, u)) in emb_dict.iteritems():
emb_count[int(c * 10 - 1)] += u
emb_count_uniq[int(c * 10 - 1)] += 1
ind = np.arange(buckets) # the x locations for the groups
# width = 0.35 # the width of the bars
width = 0.49 # the width of the bars
alpha_uniq = 0.4
fig, ax = plt.subplots()
# axes.set_xlim([0.5, 1.0])
# axes.set_ylim([ymin,ymax])
seqgan_bars = ax.bar(ind, seqgan_count, width, color=color_seqgan)
seqgan_bars_uniq = ax.bar(ind,
seqgan_count_uniq,
width,
color='black',
alpha=alpha_uniq)
emb_bars = ax.bar(ind + width, emb_count, width, color=color_emb)
emb_bars_uniq = ax.bar(ind + width,
emb_count_uniq,
width,
color='black',
alpha=alpha_uniq)
# add some text for labels, title and axes ticks
ax.set_ylabel('Count')
ax.set_xlabel(u'β')
ax.set_xticks(ind + width / 2)
x_tick_labels = []
for i in range(buckets):
x_tick_labels.append("%.1f-%.1f" %
(float(i) / buckets, float(i + 1) / buckets))
# x_tick_labels.append("%.1f" % (float(i + 1) / buckets))
ax.set_xticklabels(x_tick_labels)
# ax.set_xlabel(x_tick_labels)
plt.tick_params(
axis='x',
which='both',
bottom='off',
)
ax.legend((seqgan_bars[0], emb_bars[0]),
('SeqGan', 'Word Embedding Model'),
fontsize=20)
autolabel(seqgan_bars_uniq, ax, seqgan_count)
autolabel(emb_bars_uniq, ax, emb_count)
plt.show()
