data.py

#from transformers import AlbertTokenizer, AlbertModel
from sklearn.metrics.pairwise import cosine_similarity
from sentence_transformers import SentenceTransformer

#This is a quick evaluation on a few cases

# base
# large
#tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
#model = AlbertModel.from_pretrained("albert-base-v2")
#'sentence-transformers/paraphrase-albert-base-v2'
model_name = 'output/training_OnlineConstrativeLoss-2023-03-09_23-55-34'
model_sbert = SentenceTransformer(model_name)

def get_sbert_embedding(input_text):
    embedding = model_sbert.encode(input_text)
    return embedding.tolist()

a1 = "65 Mountain Blvd Ext, Warren, NJ 07059"
a2 = "112 Mountain Blvd Ext, Warren, NJ 07059"
a3 = "1677 NJ-27 #2, Edison, NJ 08817"
a4 = "5078 S Maryland Pkwy, Las Vegas, NV 89119"
a5 = "65 Mountain Boulevard Ext, Warren, NJ 07059"
a6 = "123 Broad St, New York, NY, 10304-2345"
a7 = "440 TECHNOLOGY CENTER DRIVE, Boston, MA 10034"
a8 = "200 Technology Center Drive, Boston, MA 10034"
a8x= "87 Technology Center Drive, Boston, MA 10034"
a9 = "440 Technology Center Dr., Boston, MA 10034-0345"
a10 = "440 Technology Center Dr., Boston, MA 10034"

#def get_embedding(input_text):
#    encoded_input = tokenizer(input_text, return_tensors='pt')
#    input_ids = encoded_input.input_ids
#    input_num_tokens = input_ids.shape[1]
#
#    print( "Number of input tokens: " + str(input_num_tokens))
#    print("Length of input: " + str(len(input_text)))
#
#    list_of_tokens = tokenizer.convert_ids_to_tokens(input_ids.view(-1).tolist())
#
#    print( "Tokens : " + ' '.join(list_of_tokens))
#    with torch.no_grad():
#
#        outputs = model(**encoded_input)
#        last_hidden_states = outputs[0]
#        sentence_embedding = torch.mean(last_hidden_states[0], dim=0)
#        #sentence_embedding = output.last_hidden_state[0][0]
#        return sentence_embedding.tolist()

e1 = get_sbert_embedding(a1)
e2 = get_sbert_embedding(a2)
#e3 = get_sbert_embedding(a3)
e4 = get_sbert_embedding(a4)
e5 = get_sbert_embedding(a5)
e6 = get_sbert_embedding(a6)
e7 = get_sbert_embedding(a7)
e8 = get_sbert_embedding(a8)
e8x = get_sbert_embedding(a8x)
e9 = get_sbert_embedding(a9)
e10 = get_sbert_embedding(a10)
print(f"a1 \"{a1}\" to \"{a2}\" a2")
print(cosine_similarity([e1], [e2]))
print(f"a1 \"{a1}\" to \"{a4}\" a4")
print(cosine_similarity([e1], [e4]))
print(f"a1 \"{a1}\" to \"{a5}\" a5")
print(cosine_similarity([e1], [e5]))

print(f"a7 \"{a7}\" to \"{a8}\" a8")
print(cosine_similarity([e7], [e8]))
print(f"a7 \"{a7}\" to \"{a8x}\" a8x")
print(cosine_similarity([e7], [e8x]))

print(f"a7 \"{a7}\" to \"{a9}\" a9")
print(cosine_similarity([e7], [e9]))

print(f"a7 \"{a7}\" to \"{a10}\" a10")
print(cosine_similarity([e7], [e10]))
# with base
#a1 to a2
#[[0.99512167]]
#a1 to a4
#[[0.94850088]]
#a1 to a5
#[[0.99636901]]

# with large
#a1 to a2
#[[0.99682108]]
#a1 to a4
#[[0.94006972]]
#a1 to a5
#[[0.99503919]]