train_blocks: List[RelationBlock] = []
test_blocks: List[RelationBlock] = []
for source, target in [(df_train, train_blocks), (df_test, test_blocks)]:
unique_users, user_map = np.unique(source.user_id, return_inverse=True)
target.append(RelationBlock(user_map, augment_user_id(unique_users)))
unique_movies, movie_map = np.unique(source.movie_id,
return_inverse=True)
target.append(RelationBlock(movie_map,
augment_movie_id(unique_movies)))
trace_path = "rmse_{0}_fold_{1}.csv".format(ALGORITHM, FOLD_INDEX)
callback: LibFMLikeCallbackBase
fm: Union[MyFMGibbsRegressor, MyFMOrderedProbit]
if ALGORITHM == "regression":
fm = myfm.MyFMRegressor(rank=DIMENSION)
callback = RegressionCallback(
n_iter=ITERATION,
X_test=X_date_test,
y_test=df_test.rating.values,
X_rel_test=test_blocks,
clip_min=df_train.rating.min(),
clip_max=df_train.rating.max(),
trace_path=trace_path,
)
else:
fm = myfm.MyFMOrderedProbit(rank=DIMENSION)
callback = OrderedProbitCallback(
n_iter=ITERATION,
X_test=X_date_test,
y_test=df_test.rating.values,
X_movie[i, i] = 1
for mid, watched in movie_vs_watched.items():
if not watched:
continue
m_iid = movie_to_internal[mid]
u_iids = [
len(movie_to_internal) + user_to_internal[uid] for uid in watched
]
X_movie[m_iid, u_iids] = 1 / max(1, len(watched))**0.5
X_movie = X_movie.tocsr()
rb_1 = RelationBlock(df_train.user_id.map(user_to_internal).values, X_user)
rb_2 = RelationBlock(df_train.movie_id.map(movie_to_internal).values, X_movie)
fmr = myfm.MyFMRegressor(rank=12)
fmr.fit(None,
df_train.rating.values,
X_rel=[rb_1, rb_2],
n_iter=200,
n_kept_samples=100)
print(fmr.fms_)
main = sps.csr_matrix((80000, 0), dtype=np.float64)
print(((fmr.predict(main, [rb_1, rb_2]) -
df_train.rating.values)**2).mean()**0.5)
with open('test.pkl', 'wb') as ofs:
import pickle
pickle.dump(fmr, ofs)