|
@@ -7,9 +7,11 @@ |
|
|
# software distributed under the License is distributed on an |
|
|
# software distributed under the License is distributed on an |
|
|
# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
|
|
# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
|
|
import numpy as np |
|
|
import numpy as np |
|
|
|
|
|
import pytest |
|
|
|
|
|
|
|
|
import megengine as mge |
|
|
import megengine as mge |
|
|
from megengine.module import LeakyReLU |
|
|
|
|
|
|
|
|
from megengine.jit.tracing import set_symbolic_shape |
|
|
|
|
|
from megengine.module import LeakyReLU, PReLU |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def test_leaky_relu(): |
|
|
def test_leaky_relu(): |
|
@@ -21,3 +23,19 @@ def test_leaky_relu(): |
|
|
|
|
|
|
|
|
np_output = np.maximum(0, data) + negative_slope * np.minimum(0, data) |
|
|
np_output = np.maximum(0, data) + negative_slope * np.minimum(0, data) |
|
|
np.testing.assert_equal(output.numpy(), np_output) |
|
|
np.testing.assert_equal(output.numpy(), np_output) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
@pytest.mark.parametrize("shape", [(1, 64, 15, 15), (64,)]) |
|
|
|
|
|
@pytest.mark.parametrize("use_symbolic", [False, True]) |
|
|
|
|
|
def test_prelu(shape, use_symbolic): |
|
|
|
|
|
old_flag = set_symbolic_shape(use_symbolic) |
|
|
|
|
|
data = np.random.random(size=shape) |
|
|
|
|
|
|
|
|
|
|
|
num_channel = 1 if len(shape) == 1 else shape[1] |
|
|
|
|
|
prelu = PReLU(num_parameters=num_channel, init=0.25) |
|
|
|
|
|
output = prelu(mge.Tensor(data)) |
|
|
|
|
|
|
|
|
|
|
|
np_output = np.maximum(data, 0) + prelu.weight.numpy() * np.minimum(data, 0) |
|
|
|
|
|
set_symbolic_shape(old_flag) |
|
|
|
|
|
|
|
|
|
|
|
np.testing.assert_allclose(output.numpy(), np_output, atol=1e-5) |