def test_matmul_relu(self):
"""Tests that matrix multiply can be combined with a simple relu."""
lhs = tile.Value.from_ndims(2)
rhs = tile.Value.from_dimensions((3, None))
out = op.relu(op.matmul(lhs, rhs))
func = tile.compose(self._ctx,
self._dev,
inputs=[('lhs', lhs), ('rhs', rhs)],
outputs=[('out', out)])
invoker = plaidml.Invoker(self._ctx, func)
invoker.set_input('lhs', self.make_inited_tensor((3, 3)))
invoker.set_input('rhs', self.make_inited_tensor((3, 3)))
output = self.make_output_tensor(invoker.get_output_shape('out'))
invoker.set_output('out', output)
invoker.invoke()
with output.mmap_current() as view:
self.assertEqual(view[0], 1.0 + 8.0 + 21.0)
self.assertEqual(view[1], 2.0 + 10.0 + 24.0)
self.assertEqual(view[2], 3.0 + 12.0 + 27.0)
self.assertEqual(view[(1, 0)], 4.0 + 20.0 + 42.0)
self.assertEqual(view[(1, 1)], 8.0 + 25.0 + 48.0)
self.assertEqual(view[(1, 2)], 12.0 + 30.0 + 54.0)
self.assertEqual(view[6], 7.0 + 32.0 + 63.0)
self.assertEqual(view[7], 14.0 + 40.0 + 72.0)
self.assertEqual(view[8], 21.0 + 48.0 + 81.0)
def relu(data):
return (op.relu(data), )
def prelu(x, slope):
if slope.shape.ndims == 1 and x.shape.ndims > 2:
slope = op.reshape(slope, [slope.shape.dims[0]] + [1] *
(x.shape.ndims - 2))
return (op.relu(x, alpha=slope), )
def leaky_relu(x, alpha=0.01):
return (op.relu(x, alpha), )
def leaky_relu(unused_ctx, x, alpha=0.01):
return (op.relu(x, alpha), )
def prelu(unused_ctx, x, slope):
# N.B. According to the operator specification, this is the V6 behavior of PRelu.
# The ONNX backend tests, however, expect that at V6, the V1 operator behavior will be used
# (broadcasting a one-dimensional slope to the channels dimension of the input tensor);
# at V7, they expect the unidirectional broadcast behavior (implemented by this method).
return (op.relu(x, alpha=slope), )
def relu(unused_ctx, data):
return (op.relu(data), )
