def __init__(self, folder='../data/downloaded_images/*.jpg', name = "images", username = 'demo', password = 'demo', hostname='localhost', port='1972', namespace='USER', recalculate=False):

self.name = name

self.username = username

self.password = password

self.hostname = hostname

self.port = port

self.recalculate = recalculate

self.namespace = namespace

self.engine = None

self.model = models.resnet152(pretrained=True)

self.model.eval()

self.paths = glob.glob(folder)

self.embeddings = False

self.connect_to_database()

self.create_images_table()

if self.embeddings == False:

self.generate_embeddings()

self.insert_data_into_database()

def connect_to_database(self):

CONNECTION_STRING = f"iris://{self.username}:{self.password}@{self.hostname}:{self.port}/{self.namespace}"

self.engine = create_engine(CONNECTION_STRING)

def create_images_table(self):

s_values = {

"float64": "DOUBLE",

"object": "VARCHAR(20000)",

"O": "VARCHAR(20000)",

"int32": "INT",

"int64": "INT"

}

with self.engine.connect() as conn:

with conn.begin():

try:

sql = f"CREATE TABLE {self.name} (\n"

sql += " monument_name VARCHAR(20000), \n description_vector VECTOR(DOUBLE, 1000)\n)"

conn.execute(text(sql))

except:

if self.recalculate:

sql = f"DROP TABLE {self.name}"

conn.execute(text(sql))

sql = f"CREATE TABLE {self.name} (\n"

sql += " monument_name VARCHAR(20000), \n description_vector VECTOR(DOUBLE, 1000)\n)"

conn.execute(text(sql))

else:

self.embeddings = True

def load_image(self, image_path):

transform = transforms.Compose([

transforms.Resize((224, 224)),

transforms.ToTensor(),

transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),

])

image = Image.open(image_path).convert('RGB') # Convertir la imagen a RGB

image = transform(image).unsqueeze(0) # Añade una dimensión al principio

return image

def clean_image_name(self, file_path):

base_name = os.path.basename(file_path).replace('.jpg', '')

# Delete any numbers and underscores at the end of the file name

clean_name = re.sub(r'_[0-9]+$', '', base_name)

# Replace any remaining underscores with spaces

clean_name = clean_name.replace('_', ' ')

return clean_name

def get_embedding(self, image_tensor):

with torch.no_grad():

embedding = self.model(image_tensor)

return embedding

def generate_embeddings(self):

self.image_embeddings = [self.get_embedding(self.load_image(img_path)) for img_path in self.paths]

def insert_data_into_database(self):

with self.engine.connect() as conn:

with conn.begin():

for index, row in enumerate(self.paths):

sql = text(f"""

INSERT INTO {self.name}

(monument_name, description_vector)

VALUES (:monument_name, TO_VECTOR(:description_vector))

""")

to_execute = {}

to_execute["monument_name"] = row

to_execute['description_vector'] = str(self.image_embeddings[index].tolist()[0])

conn.execute(sql, to_execute)

def search_similars(self, image_path, condition = "", number= 10):

search_vector = self.get_embedding(self.load_image(image_path)).tolist()[0]

with self.engine.connect() as conn:

with conn.begin():

sql = text(f"""

SELECT TOP {number} * FROM {self.name}

{condition}

ORDER BY VECTOR_COSINE(description_vector, TO_VECTOR(:search_vector)) DESC

""")

results = conn.execute(sql, {'search_vector': str(search_vector)}).fetchall()

results_df = pd.DataFrame(results, columns=["monument_name", "description_vector"])

results_df["monument_name"] = [self.clean_image_name(e) for e in results_df["monument_name"]]

pd.set_option('display.max_colwidth', None) # Easier to read description