def test_bloom(compression_ratio, expected_rmse):
interactions = movielens.get_movielens_dataset('100K')
train, test = random_train_test_split(interactions,
random_state=RANDOM_STATE)
user_embeddings = BloomEmbedding(interactions.num_users,
32,
compression_ratio=compression_ratio,
num_hash_functions=2)
item_embeddings = BloomEmbedding(interactions.num_items,
32,
compression_ratio=compression_ratio,
num_hash_functions=2)
network = BilinearNet(interactions.num_users,
interactions.num_items,
user_embedding_layer=user_embeddings,
item_embedding_layer=item_embeddings)
model = ExplicitFactorizationModel(loss='regression',
n_iter=10,
batch_size=1024,
learning_rate=1e-2,
l2=1e-5,
representation=network,
use_cuda=CUDA)
model.fit(train)
print(model)
rmse = rmse_score(model, test)
print(rmse)
assert rmse - EPSILON < expected_rmse
def train_spotlight_models(train, test, dataset_testing, embedding_dims, n_iters, batch_sizes, l2s, learning_rates, is_save = False):
"""
takes train, test, dataset_testing datasets as spotlight.interactions.
train multiple spotlight models using ExplicitFactorizationModel, with given parameters.
parameters are given in embedding_dims, n_iters, batch_sizes, l2s, learning_rates.
return predictions of train, test, dataset_testing datasets as well as rmse on train and test.
"""
# initialize train_rmses and test_rmses, these store rmse on train and test set
train_rmses = np.array([])
test_rmses = np.array([])
# initialize preds_train_trains, preds_train_tests, preds_tests; these store predictions of models on train, test and dataset_testing datasets
preds_train_trains = []
preds_train_tests = []
preds_tests = []
# traverse all parameter combinations
# embedding_din, n_iter, batch_size, l2 regularization, learning_rate
for embedding_dim in embedding_dims:
for n_iter in n_iters:
for batch_size in batch_sizes:
for l2 in l2s:
for learning_rate in learning_rates:
# initialize model with parameter, ues GPU is torch.cuda.is_available() returns True, otherwise use CPU
model = ExplicitFactorizationModel(loss='regression',
embedding_dim=embedding_dim, # latent dimensionality
n_iter=n_iter, # number of epochs of training
batch_size=batch_size, # minibatch size
l2=l2, # strength of L2 regularization
learning_rate=learning_rate,
use_cuda=torch.cuda.is_available())
# print which model is being trained
print("embedding_dim={}, n_iter={}, batch_size={}, l2={}, learning_rate={}".format(embedding_dim, n_iter, batch_size, l2, learning_rate))
# fit model
model.fit(train, verbose=True)
# find rmse on train
train_rmse = rmse_score(model, train)
# find rmse on test
test_rmse = rmse_score(model, test)
# store rmse on train and test sets
train_rmses = np.append(train_rmses, train_rmse)
test_rmses = np.append(test_rmses, test_rmse)
# print train and test rmses
print('Train RMSE {:.3f}, test RMSE {:.3f}'.format(train_rmse, test_rmse))
# if is_save given, save the models to disk
if is_save:
torch.save(model, "models/embedding_dim={}, n_iter={}, batch_size={}, l2={}, learning_rate={}".format(embedding_dim, n_iter, batch_size, l2, learning_rate))
# find predictions of train, test and dataset_testing datasets
preds_train_train = model.predict(train.user_ids,train.item_ids)
preds_train_test = model.predict(test.user_ids,test.item_ids)
preds_test = model.predict(dataset_testing.user_ids,dataset_testing.item_ids)
#store those predictions
preds_train_trains.append(preds_train_train)
preds_train_tests.append(preds_train_test)
preds_tests.append(preds_test)
# return stored predictions on train, test, dataset_testing; return rmses on train and test
return preds_train_trains, preds_train_tests, preds_tests, train_rmses, test_rmses
def best_params_spotlight(losses,
n_iters,
batch_sizes,
l2s,
learning_rates,
embedding_dims,
train_data,
t=Timer()):
rmses = dict()
params = dict()
t.start()
for loss in losses:
params['loss'] = loss
for n_iter in n_iters:
params['n_iter'] = n_iter
for batch_size in batch_sizes:
params['batch_size'] = batch_size
for l2 in l2s:
params['l2'] = l2
for learning_rate in learning_rates:
params['learning_rate'] = learning_rate
for embedding_dim in embedding_dims:
params['embedding_dim'] = embedding_dim
model = ExplicitFactorizationModel(
loss='regression',
embedding_dim=
embedding_dim, # latent dimensionality
n_iter=n_iter, # number of epochs of training
batch_size=batch_size, # minibatch size
l2=l2, # strength of L2 regularization
learning_rate=learning_rate,
use_cuda=torch.cuda.is_available())
params['model'] = model
train_tr_data, test_tr_data = random_train_test_split(
train_data,
random_state=np.random.RandomState(42))
model.fit(train_tr_data, verbose=True)
rmse = rmse_score(model, test_tr_data)
rmses[rmse] = params
print(
"-----------Time: {}, Loss: {}, n_iter: {}, l2: {}, batch_size: {}, learning_rate: {}, embedding_dim: {}, rmse: {}-------------\n\n"
.format(t.stop(), loss, n_iter, l2, batch_size,
learning_rate, embedding_dim, rmse))
# restart timer
t.start()
def train(user_ids, item_ids, ratings, num_dimensions, verbose):
dataset = Interactions(np.array(user_ids, dtype=np.int32),
np.array(item_ids, dtype=np.int32),
ratings=np.array(ratings, dtype=np.float32))
is_cuda_available = False if device.type == 'cpu' else True
m = ExplicitFactorizationModel(loss='logistic',
use_cuda=is_cuda_available,
embedding_dim=num_dimensions)
m.fit(dataset, verbose=verbose)
user_embeddings = m._net.user_embeddings.weight.detach().cpu().numpy()
return user_embeddings
def test_explicit_serialization(data):
train, test = data
model = ExplicitFactorizationModel(loss='regression',
n_iter=3,
batch_size=1024,
learning_rate=1e-3,
l2=1e-5,
use_cuda=CUDA)
model.fit(train)
rmse_original = rmse_score(model, test)
rmse_recovered = rmse_score(_reload(model), test)
assert rmse_original == rmse_recovered
def build_model(data, loss, embedding_dim, n_iter, batch_size, l2,
learning_rate, **kwargs):
model = ExplicitFactorizationModel(
loss=loss,
embedding_dim=embedding_dim, # latent dimensionality
n_iter=n_iter, # number of epochs of training
batch_size=batch_size, # minibatch size
l2=l2, # strength of L2 regularization
learning_rate=learning_rate,
use_cuda=torch.cuda.is_available())
train, test = random_train_test_split(
data, random_state=np.random.RandomState(42))
model.fit(train, verbose=True)
test_rmse = rmse_score(model, test)
return test_rmse
def test_poisson():
interactions = movielens.get_movielens_dataset('100K')
train, test = random_train_test_split(interactions,
random_state=RANDOM_STATE)
model = ExplicitFactorizationModel(loss='poisson',
n_iter=10,
batch_size=1024,
learning_rate=1e-3,
l2=1e-6)
model.fit(train)
rmse = rmse_score(model, test)
assert rmse < 1.0
def test_poisson():
interactions = movielens.get_movielens_dataset('100K')
train, test = random_train_test_split(interactions,
random_state=RANDOM_STATE)
model = ExplicitFactorizationModel(loss='poisson',
n_iter=10,
batch_size=1024,
learning_rate=1e-3,
l2=1e-6)
model.fit(train)
rmse = rmse_score(model, test)
assert rmse < 1.0
def test_check_input():
# Train for single iter.
interactions = movielens.get_movielens_dataset('100K')
train, test = random_train_test_split(interactions,
random_state=RANDOM_STATE)
model = ExplicitFactorizationModel(loss='regression',
n_iter=1,
batch_size=1024,
learning_rate=1e-3,
l2=1e-6)
model.fit(train)
# Modify data to make imcompatible with original model.
train.user_ids[0] = train.user_ids.max() + 1
with pytest.raises(ValueError):
model.fit(train)
def train_spotlight_models_using_all_data(train, dataset_testing, embedding_dims, n_iters, batch_sizes, l2s, learning_rates, verbose=True):
"""
takes train dataset as spotlight.interactions.
train multiple spotlight models using ExplicitFactorizationModel, with given parameters.
parameters are given in embedding_dims, n_iters, batch_sizes, l2s, learning_rates.
saves trained models into disk
"""
# store predictions on test set
preds_tests = []
# traverse all parameter combinations
# embedding_din, n_iter, batch_size, l2 regularization, learning_rate
for embedding_dim in embedding_dims:
for n_iter in n_iters:
for batch_size in batch_sizes:
for l2 in l2s:
for learning_rate in learning_rates:
# initialize model
model = ExplicitFactorizationModel(loss='regression',
embedding_dim=embedding_dim, # latent dimensionality
n_iter=n_iter, # number of epochs of training
batch_size=batch_size, # minibatch size
l2=l2, # strength of L2 regularization
learning_rate=learning_rate,
use_cuda=torch.cuda.is_available())
# print if given True
if verbose:
print("embedding_dim={}, n_iter={}, batch_size={}, l2={}, learning_rate={}".format(embedding_dim, n_iter, batch_size, l2, learning_rate))
# fit model using train dataset
model.fit(train, verbose=verbose)
preds_test = model.predict(dataset_testing.user_ids,dataset_testing.item_ids)
preds_tests.append(preds_test)
# save model to disk
torch.save(model, "models_all_data/embedding_dim={}, n_iter={}, batch_size={}, l2={}, learning_rate={}".format(embedding_dim, n_iter, batch_size, l2, learning_rate))
# return stored predictions on dataset_testing
return preds_tests
def test_logistic():
interactions = movielens.get_movielens_dataset('100K')
# Convert to binary
interactions.ratings = (interactions.ratings > 3).astype(np.float32)
# Convert from (0, 1) to (-1, 1)
interactions.ratings = interactions.ratings * 2 - 1
train, test = random_train_test_split(interactions,
random_state=RANDOM_STATE)
model = ExplicitFactorizationModel(loss='logistic',
n_iter=10,
batch_size=1024,
learning_rate=1e-3,
l2=1e-6,
use_cuda=CUDA)
model.fit(train)
rmse = rmse_score(model, test)
assert rmse - EPSILON < 1.05
def obtener_modelos(self):
"""
Método obtener_modelos. Obtiene, entrena y guarda el modelo escogido.
Este método solo se utiliza en la interfaz de texto.
"""
global train, modelo
# Se obtiene el modelo, se entrena con parámetros por defecto y se guarda
if self.opcion_modelo == 1:
modelo = ExplicitFactorizationModel(loss='logistic', use_cuda=torch.cuda.is_available())
modelo.fit(train, verbose=True)
guardar_modelos_dl(modelo, 'el modelo de factorización explícito')
elif self.opcion_modelo == 2:
modelo = ImplicitFactorizationModel(loss='bpr', use_cuda=torch.cuda.is_available())
modelo.fit(train, verbose=True)
guardar_modelos_dl(modelo, 'el modelo de factorización implícito')
else:
modelo = ImplicitSequenceModel(loss='bpr', representation='pooling', use_cuda=torch.cuda.is_available())
modelo.fit(train, verbose=True)
guardar_modelos_dl(modelo, 'el modelo de secuencia explícito')
from spotlight.interactions import Interactions
user_ids = np.array(lite['user']).astype(np.int32)
item_ids = np.array(lite['item']).astype(np.int32)
ratings = np.array(lite['rating']).astype(np.float32)
times = np.array(lite['time']).astype(np.int32)
dataset = Interactions(user_ids, item_ids, ratings, times)
# Prepare train test
train, test = user_based_train_test_split(dataset)
# train, test = random_train_test_split(dataset)
# Test baseline
model = ExplicitFactorizationModel(n_iter=20)
model.fit(train, verbose=True)
print('RMSE', rmse_score(model, test))
from scipy.sparse import coo_matrix
ratings = coo_matrix((dataset.ratings, (dataset.user_ids, dataset.item_ids)),
shape=(dataset.num_users, dataset.num_items)).tocsr()
train_seq = train.to_sequence(SEQ_LEN)
test_seq = test.to_sequence(SEQ_LEN)
model = ExplicitSequenceModel(n_iter=30, representation='lstm', batch_size=1)
model.fit(train_seq, ratings, verbose=True)
SEQ_ID = 0
user_batch = train_seq.user_ids[SEQ_ID]
MRR_fra_rating_5 = RR_fra_rating_5.mean()
print("MRR of fraction of 5* ratings: ", MRR_fra_rating_5)
"""#### So, the best model of Question 1 is Fraction of 5* ratings(fra_rating_5)
## Question 2
"""
from spotlight.factorization.explicit import ExplicitFactorizationModel
from spotlight.factorization.implicit import ImplicitFactorizationModel
#ExplicitFactorizationModel
emodel = ExplicitFactorizationModel(n_iter=10,
embedding_dim=32,
use_cuda=False)
emodel.fit(exp_train, verbose=True)
score_emodel = scoreAll(emodel)
print(calc_reciprank(exp_validation, score_emodel, train=exp_train).mean())
#ImplicitFactorizationModel
imodel = ImplicitFactorizationModel(n_iter=10,
loss='bpr',
embedding_dim=32,
use_cuda=False)
imodel.fit(exp_train, verbose=True)
score_imodel_32_on_exp = scoreAll(imodel)
print(calc_reciprank(exp_validation, score_imodel_32_on_exp, train=exp_train).mean())
#ImplicitFactorizationModel is more effective
#tune the number of latent factors
item_ids = np.array(df['beer'])
ratings = np.array(df['rating']).astype('float32')
#Explicit Factorization Model
explicit_interactions = Interactions(user_ids, item_ids, ratings)
explicit_interactions.tocoo().todense().shape
explicit_model = ExplicitFactorizationModel(loss='regression',
embedding_dim=32,
n_iter=10,
batch_size=250,
learning_rate=0.01)
explicit_model.fit(explicit_interactions)
user_df = pd.DataFrame(explicit_interactions.tocoo().todense())
#Spotlight Model
#This function uses the Spotlight model to make recommendations for a given user
def spotlight_predictions(user_id):
app_user = pd.DataFrame(user_df.iloc[user_id]).T
app_user_preds = pd.DataFrame({
'beer': beer_encoder.classes_,
'value': explicit_model.predict(np.array(app_user)), #needs to be passed as an array
}).sort_values('value').tail(20)
embedding_dim=5, # latent dimensionality
n_iter=10, # number of epochs of training
batch_size=256, # minibatch size
l2=1e-9, # strength of L2 regularization
learning_rate=2e-2,
use_cuda=torch.cuda.is_available())
# model = ImplicitFactorizationModel(loss='bpr',
# embedding_dim=128, # latent dimensionality
# n_iter=10, # number of epochs of training
# batch_size=256, # minibatch size
# l2=1e-9, # strength of L2 regularization
# learning_rate=1e-2,
# use_cuda=torch.cuda.is_available())
from spotlight.cross_validation import random_train_test_split
train, test = random_train_test_split(dataset,
random_state=np.random.RandomState(42))
print('Split into \n {} and \n {}.'.format(train, test))
model.fit(train, verbose=True)
torch.save(model, 'spotlight.model')
from spotlight.evaluation import rmse_score
train_rmse = rmse_score(model, train)
test_rmse = rmse_score(model, test)
print('Train RMSE {:.3f}, test RMSE {:.3f}'.format(train_rmse, test_rmse))
predictions = model.predict(test.user_ids, test.item_ids)
print(((predictions > 0.5) == (test.ratings > 0)).sum() / len(predictions))
from spotlight.cross_validation import random_train_test_split from spotlight.datasets.movielens import get_movielens_dataset from spotlight.evaluation import rmse_score from spotlight.factorization.explicit import ExplicitFactorizationModel dataset = get_movielens_dataset(variant='100K') train, test = random_train_test_split(dataset) model = ExplicitFactorizationModel(n_iter=1) model.fit(train) rmse = rmse_score(model, test) print(rmse)
## Build interactions object (building torch tensors underneath)
log.info("Building interactions object")
interactions = Interactions(
item_ids=ratings_df.movie_int.astype(np.int32).values,
user_ids=ratings_df.user_int.astype(np.int32).values,
num_items=len(ratings_df.movie_int.unique()),
num_users=len(ratings_df.user_int.unique()),
ratings=ratings_df.vote.astype(np.float32).values)
## Build Explicit Matrix Factorization Model
# We use logistic loss since the interaction rating is binary (-1, 1)
log.info(
"Training the recommendation engine model using Explicit Matrix Factorization"
)
model = ExplicitFactorizationModel(loss='logistic', n_iter=10)
model.fit(interactions)
# Prepare to get predictions out for each user
full_movies = movie_ind.movie_int.unique()
recommendations = []
# Convert datetime to string to ensure serialization success
timestamp = datetime.now().strftime('%Y-%m-%d %H:%M:%S.%f')[:-3]
batch_count = 0
for device, user_row in user_ind.iterrows():
# Get list of all movies this user voted on
log.info("Generating recommendations for user {}".format(device))
user = user_row.user_int
user_votes = ratings_df[ratings_df.user_int == user].movie_int.unique()
# Calculate difference in the two lists - rate those movies only
m = np.setdiff1d(full_movies, user_votes)
def trainModelUntilOverfit(dataset, modelSteps, modelIterations,
numberDataSplits, embedding_dim, learning_rate):
numUsers = dataset.num_users
numMovies = dataset.num_items
train, test = random_train_test_split(dataset, 0.2)
print('Split into \n {} and \n {}.'.format(train, test))
#add random seed
seed = np.random.RandomState(seed=55555)
model = ExplicitFactorizationModel(n_iter=modelIterations,
embedding_dim=embedding_dim,
learning_rate=learning_rate,
random_state=seed)
rmseResults = np.empty((modelSteps * numberDataSplits, 2))
indexPreviousClosest = ["0"]
if (numberDataSplits > 1):
arraySplits = dataSplit(train, numberDataSplits)
print("Data set split into", len(arraySplits), "*", (arraySplits[1]))
# Each model step fits the entire dataset
arrayOfSteps = []
splitCounter = 0
fullStepCounter = 0 # increases each time the entire data set has been visited
currentStep = 0 # increases at every split of the data set (does not reset)
for i in range(modelSteps * numberDataSplits):
print("\nStarting step", fullStepCounter)
print("Data split", splitCounter)
if (numberDataSplits == 1):
model.fit(train, verbose=True)
elif (numberDataSplits > 1):
print(arraySplits[splitCounter])
model.fit(arraySplits[splitCounter], verbose=True)
else:
print("Invalid number of data splits")
break
#predictions for any user are made for all items, matrix has shape (944, 1683)
modelPredict = np.empty((numUsers, numMovies))
for userIndex in range(numUsers):
modelPredict[userIndex, :] = model.predict(userIndex)
# We take the transpose for tsne formatting (should be more rows than columns)
modelPredict = modelPredict.T
#Measure the model's effectiveness (how good predictions are):
rmse = rmse_score(model, test)
rmseTrain = rmse_score(model, train)
rmseTest = rmse_score(model, test)
print("RMSE TEST:", rmseTest, "\n")
rmseResults[i, :] = [rmseTrain, rmseTest]
arrayOfSteps += [i]
if (stopTraining(rmseResults, arrayOfSteps)):
rmseResults = rmseResults[:len(arrayOfSteps)]
break
if (numberDataSplits > 1):
splitCounter += 1
if (splitCounter >= len(arraySplits)):
splitCounter = 0
fullStepCounter += 1
currentStep += 1
return (model, rmseResults)
