def test_large_seq2col_gpu_against_cpu(nW):
cupy_ops = CupyOps()
numpy_ops = NumpyOps()
# Use array with a large enough batch to require multiple
# CUDA grids.
batch_size = 128 * 128 * 2 # threads per block * blocks * 2
X = numpy_ops.xp.random.randn(batch_size * 2).astype("float32").reshape(
-1, 2)
X_gpu = cupy_ops.asarray2f(X)
# Use somewhat interesting sequence lengths.
lengths = numpy_ops.asarray1i([1, 4, 2, 1] * (batch_size // 8))
lengths_gpu = cupy_ops.asarray1i(lengths)
cols = numpy_ops.seq2col(X, nW=nW, lengths=lengths)
cols_gpu = cupy_ops.seq2col(X_gpu, nW=nW, lengths=lengths_gpu)
assert_allclose(cols, cols_gpu.get())
def get_dropout_mask(self, shape: Shape,
drop: Optional[float]) -> FloatsXd:
if drop is None or drop <= 0:
return self.xp.ones(shape, dtype="f")
else:
raise ValueError(
"During prediction, dropout should not be applied")
array1d = OPS.xp.asarray([1, 2, 3], dtype="f")
array1dint = OPS.xp.asarray([1, 2, 3], dtype="i")
array2d = OPS.xp.asarray([[4, 2, 3, 4], [1, 5, 3, 1], [9, 8, 5, 7]], dtype="f")
array2dint = OPS.xp.asarray([[1, 2, 3], [4, 5, 6]], dtype="i")
array3d = OPS.xp.zeros((3, 3, 3), dtype="f")
ragged = Ragged(array2d, OPS.xp.asarray([2, 1], dtype="i"))
padded = Padded(array3d, array1d, OPS.asarray1i([1, 2, 3, 4]),
OPS.asarray1i([1, 2, 3, 4]))
width = array2d.shape[1]
vectors = numpy.zeros((array2dint.max(), 1), dtype="f")
def assert_data_match(Y, out_data):
assert type(Y) == type(out_data)
if isinstance(out_data, OPS.xp.ndarray):
assert isinstance(Y, OPS.xp.ndarray)
assert out_data.ndim == Y.ndim
elif isinstance(out_data, Ragged):
assert isinstance(Y, Ragged)
assert out_data.data.ndim == Y.data.ndim
assert out_data.lengths.ndim == Y.lengths.ndim
elif isinstance(out_data, Padded):