def save_trained_embeddings():
model = get_prediction_model()
start_time = time.time()
count = 0
tot = len(
glob.glob("./preprocessing/stored_image_embeddings_train/*.pickle"))
for file in glob.glob(
"./preprocessing/stored_image_embeddings_train/*.pickle"):
store_path = "./preprocessing/trained_image_embeddings/" + file.split(
'/')[-1]
trained_image_embeddings = {}
if not os.path.isfile(store_path):
image_dict = load_pickle_file(file)
for image_filepath in image_dict:
trained_image_embeddings[image_filepath] = model.predict(
image_dict[image_filepath])
save_pickle_file(trained_image_embeddings, store_path)
print_progress(count,
tot,
prefix="Saving trained image embeddings")
else:
print("Skipping already created file", store_path)
count += 1
print("Time to save trained_embeddings: ", time.time() - start_time)
def run_word_preprocessing():
train_labels = generate_dict_from_directory()
if not os.path.isfile("preprocessing/labels_embedding.pickle"):
convert_sentences(train_labels, settings.WORD_EMBEDDING_DIMENSION)
labels_embedding = load_pickle_file(
"preprocessing/labels_embedding.pickle")
return labels_embedding
def create_cluster(vectors, cluster_path):
saved_cluster = load_pickle_file(cluster_path)
if saved_cluster is not None:
print("Loading saved cluster...")
return saved_cluster
print("Creating cluster...")
vectors = np.asarray(vectors)
cluster = MiniBatchKMeans(n_clusters=100, random_state=0, init_size=3000)
# cluster = DBSCAN(metric=cosine_similarity)
cluster.fit(vectors)
save_pickle_file(cluster, cluster_path)
return cluster
def prepare_training_data(labels_dictionary, location="./train/"):
"""
:param labels_dictionary: dictionary of labels (key: filename without .jpg, value: 300dim averaged label vector)
:param location: ./train/ ./test/ ./validation/
:return: All data to be trained on
"""
image_vectors = []
label_vectors = []
data_type = location.split("/")[1]
for folder_path in glob.glob("./preprocessing/stored_image_embeddings_" +
data_type + "/*.pickle"):
image_dictionary = load_pickle_file(folder_path)
for image in image_dictionary:
if image in labels_dictionary:
image_vectors.append(image_dictionary[image][0])
label_vectors.append(labels_dictionary[image])
return [image_vectors, label_vectors]
def get_cluster():
cluster_path = "preprocessing/image_vector_cluster.pickle"
saved_cluster = load_pickle_file(cluster_path)
if saved_cluster is not None:
return saved_cluster
return None
cluster = get_cluster()
model = get_prediction_model()
for query in queries:
# This is the image. Just opening if here for the fun of it; not used later
# query_image = Image.open(location + '/pics/' + query + '.jpg')
# query_image.show()
# Generate a random list of 50 entries
# cluster = [training_labels[random.randint(0, len(training_labels) - 1)] for idx in range(50)]
image_embedding = embed_image(location + '/pics/' + query + '.jpg')
trained_image_embedding = predict_vector_on_model(
image_embedding, model)
cluster_filenames = compare_to_cluster(trained_image_embedding,
cluster, 50)
my_return_dict[query] = cluster_filenames
print_progress(count, tot, prefix="Predicting images")
count += 1
return my_return_dict
if __name__ == "__main__":
start_time = time.time()
# train()
labels_dict = load_pickle_file(
"./validate/pickle/descriptions000000000.pickle")
predicted_images_dict = test(
[(f.split("pics/")[-1]).split(".jpg")[0]
for f in glob.glob("./validate/pics/000000000/*.jpg")], "./validate")
print("Time: ", time.time() - start_time)