def load_model(self):
self.movielens = init_movielens(self.dataset,
min_rating=self.min_rating,
k=self.k,
# item_features=None,
item_features=['clusters'],
model='resnet50'
)
self.train_mapping = self.movielens['train-mapping']
train = self.movielens['train']
item_features = self.movielens['item_features']
king_rec = KingRec(no_components=self.optimal_no_components,
learning_rate=self.optimal_learning_rate,
alpha=self.optimal_alpha,
scale=self.optimal_scaling,
loss='warp')
self.model = king_rec.model
self.model.fit_partial(interactions=train, item_features=item_features,
epochs=self.optimal_epochs, verbose=True, num_threads=self.threads)
return self.model
optimal_learning_rate = 0.0570326091236193
optimal_no_components = 68
optimal_alpha = 0.0029503539747277366
optimal_scaling = 0.02563602355611453
return optimal_epochs, optimal_learning_rate, optimal_no_components, optimal_alpha, optimal_scaling
k = 5
threads = 16
dataset = '../king-rec-dataset/ml-latest-small'
epochs, learning_rate, no_components, alpha, scaling = load_params(
optimized_for='auc_clusters')
movielens = init_movielens(dataset,
min_rating=3.5,
k=k,
item_features=['clusters'])
train = movielens['train']
test = movielens['test']
item_features = movielens['item_features']
king_rec = KingRec(no_components=no_components,
learning_rate=learning_rate,
alpha=alpha,
scale=scaling,
loss='warp')
model = king_rec.model
model.fit_partial(interactions=train,
item_features=item_features,
mapatk = np.mean(patks)
# Make negative because we want to minimize objective
out = -mapatk
# Handle some weird numerical shit going on
if np.abs(out + 1) < 0.01 or out < -1.0:
return 0.0
else:
return out
results = open('../../optimized-params-' + str(time.time()), 'a')
for model in ['vgg19', 'inception_v3', 'resnet50']:
for features in [None, ['genres'], ['clusters'], ['genres', 'clusters']]:
movielens = init_movielens(dataset, min_rating=3.5, k=k, item_features=features, model=model)
train = movielens['train']
test = movielens['test']
item_features = movielens['item_features']
for loss in ['warp', 'bpr']: # 'warp-kos'
for function_to_optimize in [precision_at_k, auc_score]:
space = [(1, 250), # epochs
(10 ** -4, 1.0, 'log-uniform'), # learning_rate
(20, 200), # no_components
(10 ** -6, 10 ** -1, 'log-uniform'), # item_alpha
(0.001, 1., 'log-uniform'), # user_scaling
# (1, 5), # k for k-OS training
]
current_fixed_params = '\n==> Optimize with features: ' + str(features) + ' loss: ' + loss + \
def clusters_models(params_to_load_clusters=load_auc_params_clusters):
for model in ['vgg19', 'inception_v3', 'resnet50']:
# features = ['clusters']
features = ['genres', 'clusters']
movielens = init_movielens(dataset, min_rating=3.5, k=k, item_features=features, model=model)
if len(features) == 1:
epochs, learning_rate, no_components, alpha, scaling = params_to_load_clusters(optimized_for=features[0], model=model)
elif len(features) == 2:
epochs, learning_rate, no_components, alpha, scaling = params_to_load_clusters(optimized_for=features[0] + '_' + features[1], model=model)
train = movielens['train']
test = movielens['test']
item_features = movielens['item_features']
king_rec = KingRec(no_components=no_components, learning_rate=learning_rate, alpha=alpha, scale=scaling, loss='warp')
model = king_rec.model
train_auc_scores = []
test_auc_scores = []
train_precision_scores = []
test_precision_scores = []
for epoch in range(epochs):
print('Epoch:', epoch)
model.fit_partial(train, item_features=item_features, epochs=1)
train_precision = precision_at_k(model, train, item_features=item_features, k=k, num_threads=threads).mean()
test_precision = precision_at_k(model, test, item_features=item_features, k=k, num_threads=threads).mean()
train_precision_scores.append(train_precision)
test_precision_scores.append(test_precision)
# train_auc = auc_score(model, train, item_features=item_features, num_threads=threads).mean()
# test_auc = auc_score(model, test, item_features=item_features, num_threads=threads).mean()
# train_auc_scores.append(train_auc)
# test_auc_scores.append(test_auc)
# final_test_auc.append(test_auc_scores)
final_test_precision.append(test_precision_scores)
# plot results
plt.figure()
for auc_scores in final_test_auc:
x = np.arange(len(auc_scores))
max_value = max(auc_scores)
max_index = auc_scores.index(max_value)
# plt.plot(x, auc_scores, '-D', markevery=[max_index])
plt.plot(x, auc_scores)
plt.ylabel('Accuracy')
plt.xlabel('Epochs')
plt.title('Accuracy per model with genres and clusters metadata')
plt.legend(['vgg19: ' + str(np.round(max(final_test_auc[0]) * 100, decimals=2)) + '%',
'inception_v3: ' + str(np.round(max(final_test_auc[1]) * 100, decimals=2)) + '%',
'resnet50: ' + str(np.round(max(final_test_auc[2]) * 100, decimals=2)) + '%',
])
plt.savefig('auc-comparison.jpg')
plt.show()
plt.figure()
for precision_scores in final_test_precision:
x = np.arange(len(precision_scores))
max_value = max(precision_scores)
max_index = precision_scores.index(max_value)
# plt.plot(x, precision_scores, '-D', markevery=[max_index])
plt.plot(x, precision_scores)
plt.ylabel('Precision')
plt.xlabel('Epochs')
plt.title('Precision per model with genres and clusters metadata')
plt.legend(['vgg19: ' + str(np.round(max(final_test_precision[0]) * 100, decimals=2)) + '%',
'inception_v3: ' + str(np.round(max(final_test_precision[1]) * 100, decimals=2)) + '%',
'resnet50: ' + str(np.round(max(final_test_precision[2]) * 100, decimals=2)) + '%',
])
plt.savefig('precision-comparison.jpg')
plt.show()