def __init__(self, user_vectors, user_biases, item_vectors, item_biases):
self._user_vectors = align(user_vectors)
self._user_biases = align(user_biases)
self._item_vectors = align(item_vectors)
self._item_biases = align(item_biases)
self._lib = get_lib()
def _predict_bench(self, user_id, out):
return self._lib.predict_xnor_256(align(self._user_vectors[user_id]),
self._item_vectors,
self._user_biases[user_id],
self._item_biases,
self._user_norms[user_id],
self._item_norms, out)
def predict(self, user_id, item_ids=None):
if item_ids is None:
item_ids = slice(0, None, None)
return self._lib.predict_float_256(align(self._user_vectors[user_id]),
self._item_vectors[item_ids],
self._user_biases[user_id],
self._item_biases[item_ids])
def test_predict_float_256():
lib = get_lib()
num_items = 1024
for latent_dim in (1, 5, 16, 32, 128, 256):
user_vector = align(np.random.random(latent_dim).astype(np.float32))
item_vectors = align(np.random.random((num_items, latent_dim)).astype(np.float32))
user_bias = 1.0
item_biases = align(np.random.random(num_items).astype(np.float32))
expected = _predict_float_256(user_vector,
item_vectors,
user_bias,
item_biases)
ffi_result = lib.predict_float_256(user_vector,
item_vectors,
user_bias,
item_biases)
assert np.allclose(expected, ffi_result)
def __init__(self, user_vectors, user_biases, item_vectors, item_biases):
assert item_vectors.shape[1] >= 32
self._user_norms = align(np.abs(user_vectors).mean(axis=1))
self._item_norms = align(np.abs(item_vectors).mean(axis=1))
self._user_vectors = align(binarize_array(user_vectors))
self._user_biases = align(user_biases)
self._item_vectors = align(binarize_array(item_vectors))
self._item_biases = align(item_biases)
self._lib = get_lib()