train = pd.concat((train, dr))
users = ground.user.unique()
model = BPR_MF(n_iter=100,
learning_rate=1e-1,
sparse=False,
batch_size=4096,
n_jobs=8,
use_cuda=False,
optimizer_func=Adam,
verbose=True,
stopping=True,
early_stopping=True)
interactions = train[['user', 'item', 'rating']].values
np.random.shuffle(interactions)
x = interactions[:, :-1]
y = interactions[:, -1]
model.fit(x, y, {'users': 0, 'items': 1},
n_users=n_users, n_items=n_items)
x = cartesian2D(ground.user.unique().reshape(-1, 1), item_cat.reshape(-1, 1))
pred = model.predict(x).reshape(ground.user.unique().shape[0], item_cat.shape[0])
np.save("predictions.txt", pred)
arg = np.argsort(-pred, 1)
np.save("rankings.txt", arg)
model = FM_MMR(n_iter=1,
learning_rate=1e-1,
logger=logger,
n_jobs=4,
use_cuda=False)
x = interactions[:, :-1]
y = interactions[:, -1]
model.fit(x,
y,
dic={
'users': 0,
'items': 1,
'artists': 2
},
n_users=n_users,
n_items=n_items,
lengths={"n_artists": n_artists})
logger = TLogger()
plt.plot(logger.epochs, logger.losses)
plt.show()
users = np.unique(x[:, 0]).reshape(-1, 1)[:10, :]
items = train[['tps_id', 'artist_id', 'mode']].drop_duplicates('tps_id').values
values = cartesian2D(users, items)
top = 3
table = np.zeros((users.shape[0], top + 1), dtype=np.object)
table[:, 0] = users[:, 0]
table[:, 1:] = model.predict(values, top=top, b=1)
print(table)
n_jobs=N_JOBS,
batch_size=BATCH_SIZE,
n_factors=FACTORS,
learning_rate=LEARNING_RATE,
use_cuda=USE_CUDA,
verbose=VERBOSE,
sparse=SPARSE)
train = pd.read_csv(train_path, header=0)
ground = pd.read_csv(ground_path, header=0)
dataset = pd.concat((train, ground))
users = dataset.user.unique()
item_catalogue = dataset[proj[1:-1]].drop_duplicates().values
values = cartesian2D(users.reshape(-1, 1), item_catalogue)
model0 = FM(n_iter=N_ITER,
model=MODEL_PATH,
n_jobs=N_JOBS,
batch_size=BATCH_SIZE,
n_factors=FACTORS,
learning_rate=LEARNING_RATE,
use_cuda=USE_CUDA,
verbose=VERBOSE,
sparse=SPARSE)
rank = model0.predict(values) \
.reshape(users.shape[0], item_catalogue.shape[0])
for b in tqdm([1.0], desc="sys.div.", leave=False):
# Create Ground Truth
ground = test[~test.index.isin(dr.index)].dropna()
train = pd.concat((train, dr))
users = ground.user.unique()
model = FM(n_iter=100,
learning_rate=1e-3,
sparse=False,
batch_size=4096,
n_jobs=8,
optimizer_func=Adam,
use_cuda=True,
verbose=True,
stopping=True,
early_stopping=True)
interactions = train[['user', 'item', 'rating']].values
np.random.shuffle(interactions)
x = interactions[:, :-1]
y = interactions[:, -1]
model.fit(x, y, {'users': 0, 'items': 1}, n_items=n_items, n_users=n_users)
x = cartesian2D(users.reshape(-1, 1), item_cat.reshape(-1, 1))
pred = model.predict(x).reshape(users.shape[0], item_cat.shape[0])
np.save("predictions.txt", pred)
arg = np.argsort(-pred, 1)
np.save("rankings.txt", arg)