def extract(self, x):
c1_1 = F.relu(self.conv1_1(x))
c1_2 = F.relu(self.conv1_2(c1_1))
p1 = F.average_pooling(c1_2, 2, 2)
c2_1 = F.relu(self.conv2_1(p1))
c2_2 = F.relu(self.conv2_2(c2_1))
p2 = F.average_pooling(c2_2, 2, 2)
c3_1 = F.relu(self.conv3_1(p2))
c3_2 = F.relu(self.conv3_2(c3_1))
c3_3 = F.relu(self.conv3_3(c3_2))
p3 = F.average_pooling(c3_3, 2, 2)
c4_1 = F.relu(self.conv4_1(p3))
c4_2 = F.relu(self.conv4_2(c4_1))
c4_3 = F.relu(self.conv4_3(c4_2))
p4 = F.average_pooling(c4_3, 2, 2)
c5_1 = F.relu(self.conv5_1(p4))
c5_2 = F.relu(self.conv5_2(c5_1))
c5_3 = F.relu(self.conv5_3(c5_2))
return {
'conv1_1': c1_1,
'conv1_2': c1_2,
'conv2_1': c2_1,
'conv2_2': c2_2,
'conv3_1': c3_1,
'conv3_2': c3_2,
'conv3_3': c3_3,
'conv4_1': c4_1,
'conv5_1': c5_1,
'conv5_2': c5_2,
'conv5_3': c5_3
}
def test_forward2(self):
n, c, h, w = 1, 5, 15, 15
ksize, stride, pad = 2, 2, 0
x = np.random.randn(n, c, h, w).astype('f')
y = F.average_pooling(x, ksize, stride, pad)
expected = CF.average_pooling_2d(x, ksize, stride, pad)
self.assertTrue(array_allclose(expected.data, y.data))
def _global_average_pooling_2d(x):
N, C, H, W = x.shape
h = F.average_pooling(x, (H, W), stride=1)
h = F.reshape(h, (N, C))
return h
def test_backward1(self):
n, c, h, w = 1, 5, 16, 16
ksize, stride, pad = 2, 2, 0
x = np.random.randn(n, c, h, w).astype('f') * 1000
f = lambda x: F.average_pooling(x, ksize, stride, pad)
self.assertTrue(gradient_check(f, x))
