Matheus Schmitz
LinkedIn
Github Portfolio
# File manipulation imports for Google Colab
from google.colab import drive
drive.mount('/content/drive')
import os
os.chdir("/content/drive/My Drive/Colab Notebooks/DSCI 558")
Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).
!pip install -q ampligraph
!pip install -q tensorflow==1.15
import requests
from ampligraph.datasets import load_from_csv
import numpy as np
url = 'https://ampligraph.s3-eu-west-1.amazonaws.com/datasets/GoT.csv'
with open('GoT.csv', 'wb') as f_in:
f_in.write(requests.get(url).content)
X = load_from_csv('.', 'GoT.csv', sep=',')
X[:5, ]
array([['Smithyton', 'SEAT_OF', 'House Shermer of Smithyton'],
['House Mormont of Bear Island', 'LED_BY', 'Maege Mormont'],
['Margaery Tyrell', 'SPOUSE', 'Joffrey Baratheon'],
['Maron Nymeros Martell', 'ALLIED_WITH',
'House Nymeros Martell of Sunspear'],
['House Gargalen of Salt Shore', 'IN_REGION', 'Dorne']],
dtype=object)
entities = np.unique(np.concatenate([X[:, 0], X[:, 2]]))
entities
array(['Abelar Hightower', 'Acorn Hall', 'Addam Frey', ..., 'the Antlers',
'the Paps', 'unnamed tower'], dtype=object)
relations = np.unique(X[:, 1])
relations
array(['ALLIED_WITH', 'BRANCH_OF', 'FOUNDED_BY', 'HEIR_TO', 'IN_REGION',
'LED_BY', 'PARENT_OF', 'SEAT_OF', 'SPOUSE', 'SWORN_TO'],
dtype=object)
from ampligraph.evaluation import train_test_split_no_unseen
X_train, X_test = train_test_split_no_unseen(X, test_size=100)
print('Train set size: ', X_train.shape)
print('Test set size: ', X_test.shape)
Train set size: (3075, 3) Test set size: (100, 3)
from ampligraph.latent_features import ComplEx
model = ComplEx(batches_count=100,
seed=0,
epochs=200,
k=150,
eta=5,
optimizer='adam',
optimizer_params={'lr':1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={'p':3, 'lambda':1e-5},
verbose=True)
positives_filter = X
import tensorflow as tf
tf.logging.set_verbosity(tf.logging.ERROR)
model.fit(X_train, early_stopping = False)
Average Loss: 0.017603: 100%|██████████| 200/200 [03:25<00:00, 1.03s/epoch]
from ampligraph.evaluation import evaluate_performance
ranks = evaluate_performance(X_test,
model=model,
filter_triples=positives_filter, # Corruption strategy filter defined above
use_default_protocol=True, # corrupt subj and obj separately while evaluating
verbose=True)
WARNING - DeprecationWarning: use_default_protocol will be removed in future. Please use corrupt_side argument instead.
100%|██████████| 100/100 [00:01<00:00, 62.33it/s]
# Dictionary to compare performances of different models
model_comparison = {}
from ampligraph.evaluation import mr_score, mrr_score, hits_at_n_score
model_comparison['ComplEx'] = {}
model_comparison['ComplEx']['MMR'] = mrr_score(ranks)
print("MRR: %.2f" % (model_comparison['ComplEx']['MMR']))
model_comparison['ComplEx']['Hits@10'] = hits_at_n_score(ranks, n=10)
print("Hits@10: %.2f" % (model_comparison['ComplEx']['Hits@10']))
model_comparison['ComplEx']['Hits@3'] = hits_at_n_score(ranks, n=3)
print("Hits@3: %.2f" % (model_comparison['ComplEx']['Hits@3']))
model_comparison['ComplEx']['Hits@1'] = hits_at_n_score(ranks, n=1)
print("Hits@1: %.2f" % (model_comparison['ComplEx']['Hits@1']))
MRR: 0.41 Hits@10: 0.55 Hits@3: 0.45 Hits@1: 0.34
X_unseen = np.array([
['Jorah Mormont', 'SPOUSE', 'Daenerys Targaryen'],
['Tyrion Lannister', 'SPOUSE', 'Missandei'],
["King's Landing", 'SEAT_OF', 'House Lannister of Casterly Rock'],
['Sansa Stark', 'SPOUSE', 'Petyr Baelish'],
['Daenerys Targaryen', 'SPOUSE', 'Jon Snow'],
['Daenerys Targaryen', 'SPOUSE', 'Craster'],
['House Stark of Winterfell', 'IN_REGION', 'The North'],
['House Stark of Winterfell', 'IN_REGION', 'Dorne'],
['House Tyrell of Highgarden', 'IN_REGION', 'Beyond the Wall'],
['Brandon Stark', 'ALLIED_WITH', 'House Stark of Winterfell'],
['Brandon Stark', 'ALLIED_WITH', 'House Lannister of Casterly Rock'],
['Rhaegar Targaryen', 'PARENT_OF', 'Jon Snow'],
['House Hutcheson', 'SWORN_TO', 'House Tyrell of Highgarden'],
['Daenerys Targaryen', 'ALLIED_WITH', 'House Stark of Winterfell'],
['Daenerys Targaryen', 'ALLIED_WITH', 'House Lannister of Casterly Rock'],
['Jaime Lannister', 'PARENT_OF', 'Myrcella Baratheon'],
['Robert I Baratheon', 'PARENT_OF', 'Myrcella Baratheon'],
['Cersei Lannister', 'PARENT_OF', 'Myrcella Baratheon'],
['Cersei Lannister', 'PARENT_OF', 'Brandon Stark'],
["Tywin Lannister", 'PARENT_OF', 'Jaime Lannister'],
["Missandei", 'SPOUSE', 'Grey Worm'],
["Brienne of Tarth", 'SPOUSE', 'Jaime Lannister']
])
unseen_filter = np.array(list({tuple(i) for i in np.vstack((positives_filter, X_unseen))}))
ranks_unseen = evaluate_performance(
X_unseen,
model=model,
filter_triples=unseen_filter, # Corruption strategy filter defined above
corrupt_side = 's+o',
use_default_protocol=False, # corrupt subj and obj separately while evaluating
verbose=True
)
100%|██████████| 22/22 [00:00<00:00, 53.90it/s]
scores = model.predict(X_unseen)
from scipy.special import expit
probs = expit(scores)
import pandas as pd
pd.DataFrame(list(zip([' '.join(x) for x in X_unseen],
ranks_unseen,
np.squeeze(scores),
np.squeeze(probs))),
columns=['statement', 'rank', 'score', 'prob']).sort_values("score")
| statement | rank | score | prob | |
|---|---|---|---|---|
| 10 | Brandon Stark ALLIED_WITH House Lannister of C... | 4017 | -3.814217 | 0.021579 |
| 18 | Cersei Lannister PARENT_OF Brandon Stark | 4083 | -1.994114 | 0.119822 |
| 9 | Brandon Stark ALLIED_WITH House Stark of Winte... | 2995 | -0.747185 | 0.321435 |
| 1 | Tyrion Lannister SPOUSE Missandei | 3389 | -0.740518 | 0.322891 |
| 21 | Brienne of Tarth SPOUSE Jaime Lannister | 3493 | -0.702415 | 0.331277 |
| 5 | Daenerys Targaryen SPOUSE Craster | 3319 | -0.702258 | 0.331312 |
| 15 | Jaime Lannister PARENT_OF Myrcella Baratheon | 2943 | -0.206153 | 0.448643 |
| 0 | Jorah Mormont SPOUSE Daenerys Targaryen | 2450 | -0.201867 | 0.449704 |
| 8 | House Tyrell of Highgarden IN_REGION Beyond th... | 2155 | -0.131668 | 0.467130 |
| 2 | King's Landing SEAT_OF House Lannister of Cast... | 1724 | -0.014984 | 0.496254 |
| 11 | Rhaegar Targaryen PARENT_OF Jon Snow | 2176 | 0.039815 | 0.509952 |
| 4 | Daenerys Targaryen SPOUSE Jon Snow | 1371 | 0.164007 | 0.540910 |
| 14 | Daenerys Targaryen ALLIED_WITH House Lannister... | 838 | 0.555441 | 0.635397 |
| 17 | Cersei Lannister PARENT_OF Myrcella Baratheon | 491 | 0.634258 | 0.653454 |
| 19 | Tywin Lannister PARENT_OF Jaime Lannister | 290 | 0.813547 | 0.692865 |
| 7 | House Stark of Winterfell IN_REGION Dorne | 129 | 1.334614 | 0.791603 |
| 13 | Daenerys Targaryen ALLIED_WITH House Stark of ... | 224 | 1.390047 | 0.800600 |
| 16 | Robert I Baratheon PARENT_OF Myrcella Baratheon | 21 | 1.921836 | 0.872343 |
| 3 | Sansa Stark SPOUSE Petyr Baelish | 29 | 2.685783 | 0.936182 |
| 20 | Missandei SPOUSE Grey Worm | 78 | 2.945535 | 0.950052 |
| 6 | House Stark of Winterfell IN_REGION The North | 9 | 3.121475 | 0.957770 |
| 12 | House Hutcheson SWORN_TO House Tyrell of Highg... | 10 | 3.323205 | 0.965216 |
from ampligraph.utils import create_tensorboard_visualizations
create_tensorboard_visualizations(model, 'GoT_embeddings')
#%reload_ext tensorboard
#%tensorboard --logdir=./GoT_embeddings
# Control TensorBoard display. If no port is provided,
# the most recently launched TensorBoard is used
from tensorboard import notebook
notebook.list() # View open TensorBoard instances
notebook.display(port=6006, height=1000)
No known TensorBoard instances running.
from ampligraph.latent_features import TransE
model = TransE(batches_count=100,
seed=0,
epochs=200,
k=150,
eta=5,
optimizer='adam',
optimizer_params={'lr':1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={'p':3, 'lambda':1e-5},
verbose=True)
positives_filter = X
tf.logging.set_verbosity(tf.logging.ERROR)
model.fit(X_train, early_stopping = False)
ranks = evaluate_performance(X_test,
model=model,
filter_triples=positives_filter, # Corruption strategy filter defined above
use_default_protocol=True, # corrupt subj and obj separately while evaluating
verbose=True)
model_comparison['TransE'] = {}
model_comparison['TransE']['MMR'] = mrr_score(ranks)
print("MRR: %.2f" % (model_comparison['TransE']['MMR']))
model_comparison['TransE']['Hits@10'] = hits_at_n_score(ranks, n=10)
print("Hits@10: %.2f" % (model_comparison['TransE']['Hits@10']))
model_comparison['TransE']['Hits@3'] = hits_at_n_score(ranks, n=3)
print("Hits@3: %.2f" % (model_comparison['TransE']['Hits@3']))
model_comparison['TransE']['Hits@1'] = hits_at_n_score(ranks, n=1)
print("Hits@1: %.2f" % (model_comparison['TransE']['Hits@1']))
Average Loss: 0.017881: 100%|██████████| 200/200 [01:49<00:00, 1.82epoch/s]
WARNING - DeprecationWarning: use_default_protocol will be removed in future. Please use corrupt_side argument instead.
100%|██████████| 100/100 [00:00<00:00, 178.01it/s]
MRR: 0.20 Hits@10: 0.36 Hits@3: 0.24 Hits@1: 0.12
from ampligraph.latent_features import DistMult
model = DistMult(batches_count=100,
seed=0,
epochs=200,
k=150,
eta=5,
optimizer='adam',
optimizer_params={'lr':1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={'p':3, 'lambda':1e-5},
verbose=True)
positives_filter = X
tf.logging.set_verbosity(tf.logging.ERROR)
model.fit(X_train, early_stopping = False)
ranks = evaluate_performance(X_test,
model=model,
filter_triples=positives_filter, # Corruption strategy filter defined above
use_default_protocol=True, # corrupt subj and obj separately while evaluating
verbose=True)
model_comparison['DistMult'] = {}
model_comparison['DistMult']['MMR'] = mrr_score(ranks)
print("MRR: %.2f" % (model_comparison['DistMult']['MMR']))
model_comparison['DistMult']['Hits@10'] = hits_at_n_score(ranks, n=10)
print("Hits@10: %.2f" % (model_comparison['DistMult']['Hits@10']))
model_comparison['DistMult']['Hits@3'] = hits_at_n_score(ranks, n=3)
print("Hits@3: %.2f" % (model_comparison['DistMult']['Hits@3']))
model_comparison['DistMult']['Hits@1'] = hits_at_n_score(ranks, n=1)
print("Hits@1: %.2f" % (model_comparison['DistMult']['Hits@1']))
Average Loss: 0.016550: 100%|██████████| 200/200 [01:54<00:00, 1.74epoch/s]
WARNING - DeprecationWarning: use_default_protocol will be removed in future. Please use corrupt_side argument instead.
100%|██████████| 100/100 [00:00<00:00, 172.32it/s]
MRR: 0.39 Hits@10: 0.54 Hits@3: 0.43 Hits@1: 0.31
import pandas as pd
results = pd.DataFrame(model_comparison)
results
| ComplEx | TransE | DistMult | |
|---|---|---|---|
| MMR | 0.413934 | 0.199833 | 0.393764 |
| Hits@10 | 0.550000 | 0.365000 | 0.540000 |
| Hits@3 | 0.445000 | 0.240000 | 0.435000 |
| Hits@1 | 0.340000 | 0.120000 | 0.310000 |
Matheus Schmitz
LinkedIn
Github Portfolio