def testPadForward(self):
x = type('', (), {})()
x.value = np.ndarray([3, 5, 6, 2])
x.value.fill(0)
x.value[0, :, :, 0] = [[1, 2, 3, 4, 5, 6], [7, 8, 9, 10, 11, 12],
[13, 14, 15, 16, 17, 18],
[19, 20, 21, 22, 23, 24],
[25, 26, 27, 28, 29, 30]]
x.value[0, :, :, 1] = [[1, 2, 3, 3, 2, 1], [1, 1, 2, 2, 1, 1],
[1, 1, 1, 1, 1, 1], [1, 2, 3, 3, 2, 1],
[1, 1, 2, 2, 1, 1]]
x.grad = x.value
f = type('', (), {})()
f.value = np.ndarray([2, 3, 2, 1])
f.value[:, :, 0, 0] = [[1, 1, 1], [1, 1, 1]]
f.value[:, :, 1, 0] = [[1, 0, -1], [1, 0, -1]]
f.grad = f.value
conv = Conv(f, x, 3, 1)
conv.forward()
self.assertEqual((3, 2, 2, 1), conv.value.shape)
self.assertEqual(edf.DT(1), conv.value[0, 0, 0, 0])
self.assertEqual(edf.DT(13), conv.value[0, 0, 1, 0])
self.assertEqual(edf.DT(63), conv.value[0, 1, 0, 0])
self.assertEqual(edf.DT(115), conv.value[0, 1, 1, 0])
def testForward(self):
x = type('', (), {})()
x.value = np.ndarray([2, 4, 6, 2], np.dtype(np.float64))
x.value.fill(0)
x.value[1, :, :, 0] = np.transpose([[1, 6, 2, 5], [0, 3, 2, 2],
[7, 9, 5, 0], [0, 2, 0, 3],
[0, 0, 6, 0], [3, 1, 4, 9]])
x.grad = x.value
ave = AvePool(x, 3, 2)
ave.forward()
self.assertEqual((2, 1, 2, 2), ave.value.shape)
self.assertEqual(edf.DT(35) / 9, ave.value[1, 0, 0, 0])
self.assertEqual(edf.DT(29) / 9, ave.value[1, 0, 1, 0])
ave2 = AvePool(x, 2, 2)
ave2.forward()
self.assertEqual((2, 2, 3, 2), ave2.value.shape)
self.assertEqual(edf.DT(10) / 4, ave2.value[1, 0, 0, 0])
self.assertEqual(edf.DT(18) / 4, ave2.value[1, 0, 1, 0])
self.assertEqual(edf.DT(4) / 4, ave2.value[1, 0, 2, 0])
self.assertEqual(edf.DT(11) / 4, ave2.value[1, 1, 0, 0])
self.assertEqual(edf.DT(8) / 4, ave2.value[1, 1, 1, 0])
self.assertEqual(edf.DT(19) / 4, ave2.value[1, 1, 2, 0])
def __init__(self, x, ksz=2, stride=None):
edf.components.append(self)
self.x = x
self.ksz = ksz
if stride is None:
self.stride = ksz
else:
self.stride = stride
self.grad = None if x.grad is None else edf.DT(0)
self.square = np.ndarray([self.ksz, self.ksz])
self.square.fill(1)
def __init__(self, f, k, stride=1, pad=0):
edf.components.append(self)
self.f = f
self.k = k
pad = np.array(pad)
if pad.shape == ():
self.xpad = self.ypad = pad
else:
self.ypad = pad[0]
self.xpad = pad[1]
self.stride = stride
self.grad = None if f.grad is None and k.grad is None else edf.DT(0)
def testBackward(self):
x = type('', (), {})()
x.value = np.ndarray([2, 4, 6, 2], np.dtype(np.float64))
x.value.fill(0)
x.grad = np.ndarray([2, 4, 6, 2], np.dtype(np.float64))
x.grad.fill(0)
ave = AvePool(x, 3, 1)
ave.value = np.ndarray([2, 2, 4, 2], np.dtype(np.float64))
ave.value.fill(0)
ave.grad = np.ndarray([2, 2, 4, 2], np.dtype(np.float64))
ave.grad.fill(0)
ave.grad[0, :, :, 0] = [[1, 3, 5, 7], [2, 4, 6, 8]]
ave.backward()
self.assertEqual(edf.DT(1) / 9, x.grad[0, 0, 0, 0])
self.assertEqual(edf.DT(4) / 9, x.grad[0, 0, 1, 0])
self.assertEqual(edf.DT(9) / 9, x.grad[0, 0, 2, 0])
self.assertEqual(edf.DT(15) / 9, x.grad[0, 0, 3, 0])
self.assertEqual(edf.DT(12) / 9, x.grad[0, 0, 4, 0])
self.assertEqual(edf.DT(7) / 9, x.grad[0, 0, 5, 0])
self.assertEqual(edf.DT(3) / 9, x.grad[0, 1, 0, 0])
self.assertEqual(edf.DT(10) / 9, x.grad[0, 1, 1, 0])
self.assertEqual(edf.DT(21) / 9, x.grad[0, 1, 2, 0])
self.assertEqual(edf.DT(33) / 9, x.grad[0, 1, 3, 0])
self.assertEqual(edf.DT(26) / 9, x.grad[0, 1, 4, 0])
self.assertEqual(edf.DT(15) / 9, x.grad[0, 1, 5, 0])
self.assertEqual(edf.DT(3) / 9, x.grad[0, 2, 0, 0])
self.assertEqual(edf.DT(10) / 9, x.grad[0, 2, 1, 0])
self.assertEqual(edf.DT(21) / 9, x.grad[0, 2, 2, 0])
self.assertEqual(edf.DT(33) / 9, x.grad[0, 2, 3, 0])
self.assertEqual(edf.DT(26) / 9, x.grad[0, 2, 4, 0])
self.assertEqual(edf.DT(15) / 9, x.grad[0, 2, 5, 0])
self.assertEqual(edf.DT(2) / 9, x.grad[0, 3, 0, 0])
self.assertEqual(edf.DT(6) / 9, x.grad[0, 3, 1, 0])
self.assertEqual(edf.DT(12) / 9, x.grad[0, 3, 2, 0])
self.assertEqual(edf.DT(18) / 9, x.grad[0, 3, 3, 0])
self.assertEqual(edf.DT(14) / 9, x.grad[0, 3, 4, 0])
self.assertEqual(edf.DT(8) / 9, x.grad[0, 3, 5, 0])
def testBackward2(self):
x = type('', (), {})()
x.value = np.ndarray([2, 4, 6, 2], np.dtype(np.float64))
x.value.fill(0)
x.value[0, :, :, 0] = np.transpose([[1, 4, 5, 5], [6, 3, 2, 5],
[6, 5, 4, 4], [0, 2, 0, 3],
[3, 3, 0, 6], [3, 1, 2, 9]])
x.grad = np.ndarray([2, 4, 6, 2], np.dtype(np.float64))
x.grad.fill(0)
maxp = MaxPool(x, 3, 1)
maxp.forward()
self.assertEqual((2, 2, 4, 2), maxp.value.shape, '')
maxp.grad = np.ndarray([2, 2, 4, 2], np.dtype(np.float64))
maxp.grad.fill(0)
maxp.grad[0, :, :, 0] = [[1, 2, 3, 4], [5, 6, 7, 8]]
maxp.backward()
self.assertEqual(edf.DT(0), x.grad[0, 0, 0, 0])
self.assertEqual(edf.DT(3), x.grad[0, 0, 1, 0])
self.assertEqual(edf.DT(6), x.grad[0, 0, 2, 0])
self.assertEqual(edf.DT(0), x.grad[0, 0, 3, 0])
self.assertEqual(edf.DT(4), x.grad[0, 0, 4, 0])
self.assertEqual(edf.DT(4), x.grad[0, 0, 5, 0])
self.assertEqual(edf.DT(0), x.grad[0, 1, 0, 0])
self.assertEqual(edf.DT(0), x.grad[0, 1, 1, 0])
self.assertEqual(edf.DT(11), x.grad[0, 1, 2, 0])
self.assertEqual(edf.DT(0), x.grad[0, 1, 3, 0])
self.assertEqual(edf.DT(4), x.grad[0, 1, 4, 0])
self.assertEqual(edf.DT(0), x.grad[0, 1, 5, 0])
self.assertEqual(edf.DT(5), x.grad[0, 2, 0, 0])
self.assertEqual(edf.DT(0), x.grad[0, 2, 1, 0])
self.assertEqual(edf.DT(0), x.grad[0, 2, 2, 0])
self.assertEqual(edf.DT(0), x.grad[0, 2, 3, 0])
self.assertEqual(edf.DT(0), x.grad[0, 2, 4, 0])
self.assertEqual(edf.DT(0), x.grad[0, 2, 5, 0])
self.assertEqual(edf.DT(5), x.grad[0, 3, 0, 0])
self.assertEqual(edf.DT(11), x.grad[0, 3, 1, 0])
self.assertEqual(edf.DT(0), x.grad[0, 3, 2, 0])
self.assertEqual(edf.DT(0), x.grad[0, 3, 3, 0])
self.assertEqual(edf.DT(7), x.grad[0, 3, 4, 0])
self.assertEqual(edf.DT(8), x.grad[0, 3, 5, 0])
def testBackward1(self):
x = type('', (), {})()
x.value = np.ndarray([2, 4, 6, 2], np.dtype(np.float64))
x.value.fill(0)
x.value[1, :, :, 0] = np.transpose([[1, 6, 2, 5], [0, 3, 2, 5],
[7, 9, 4, 4], [9, 2, 0, 3],
[3, 3, 6, 0], [3, 1, 4, 9]])
x.grad = np.ndarray([2, 4, 6, 2], np.dtype(np.float64))
x.grad.fill(0)
maxp = MaxPool(x, 2, 2)
maxp.forward()
maxp.grad = np.ndarray([2, 2, 3, 2], np.dtype(np.float64))
maxp.grad.fill(0)
maxp.grad[1, :, :, 0] = [[1, 3, 5], [2, 4, 6]]
maxp.backward()
self.assertEqual(edf.DT(0), x.grad[1, 0, 0, 0])
self.assertEqual(edf.DT(0), x.grad[1, 0, 1, 0])
self.assertEqual(edf.DT(0), x.grad[1, 0, 2, 0])
self.assertEqual(edf.DT(3), x.grad[1, 0, 3, 0])
self.assertEqual(edf.DT(5), x.grad[1, 0, 4, 0])
self.assertEqual(edf.DT(5), x.grad[1, 0, 5, 0])
self.assertEqual(edf.DT(1), x.grad[1, 1, 0, 0])
self.assertEqual(edf.DT(0), x.grad[1, 1, 1, 0])
self.assertEqual(edf.DT(3), x.grad[1, 1, 2, 0])
self.assertEqual(edf.DT(0), x.grad[1, 1, 3, 0])
self.assertEqual(edf.DT(5), x.grad[1, 1, 4, 0])
self.assertEqual(edf.DT(0), x.grad[1, 1, 5, 0])
self.assertEqual(edf.DT(0), x.grad[1, 2, 0, 0])
self.assertEqual(edf.DT(0), x.grad[1, 2, 1, 0])
self.assertEqual(edf.DT(4), x.grad[1, 2, 2, 0])
self.assertEqual(edf.DT(0), x.grad[1, 2, 3, 0])
self.assertEqual(edf.DT(0), x.grad[1, 2, 4, 0])
self.assertEqual(edf.DT(0), x.grad[1, 2, 5, 0])
self.assertEqual(edf.DT(2), x.grad[1, 3, 0, 0])
self.assertEqual(edf.DT(2), x.grad[1, 3, 1, 0])
self.assertEqual(edf.DT(4), x.grad[1, 3, 2, 0])
self.assertEqual(edf.DT(0), x.grad[1, 3, 3, 0])
self.assertEqual(edf.DT(0), x.grad[1, 3, 4, 0])
self.assertEqual(edf.DT(6), x.grad[1, 3, 5, 0])
def testForward(self):
x = type('', (), {})()
x.value = np.ndarray([2, 4, 6, 2], np.dtype(np.float64))
x.value.fill(0)
x.value[1, :, :, 0] = np.transpose([[1, 6, 2, 5], [0, 3, 2, 2],
[7, 9, 5, 0], [0, 2, 0, 3],
[0, 0, 6, 0], [3, 1, 4, 9]])
x.grad = x.value
maxp = MaxPool(x, 2, 2)
maxp.forward()
self.assertEqual((2, 2, 3, 2), maxp.value.shape)
self.assertEqual(edf.DT(6), maxp.value[1, 0, 0, 0])
self.assertEqual(edf.DT(9), maxp.value[1, 0, 1, 0])
self.assertEqual(edf.DT(3), maxp.value[1, 0, 2, 0])
self.assertEqual(edf.DT(5), maxp.value[1, 1, 0, 0])
self.assertEqual(edf.DT(5), maxp.value[1, 1, 1, 0])
self.assertEqual(edf.DT(9), maxp.value[1, 1, 2, 0])
maxp2 = MaxPool(x, 3, 1)
maxp2.forward()
self.assertEqual((2, 2, 4, 2), maxp2.value.shape)
self.assertEqual(edf.DT(9), maxp2.value[1, 0, 0, 0])
self.assertEqual(edf.DT(9), maxp2.value[1, 0, 1, 0])
self.assertEqual(edf.DT(9), maxp2.value[1, 0, 2, 0])
self.assertEqual(edf.DT(6), maxp2.value[1, 0, 3, 0])
self.assertEqual(edf.DT(9), maxp2.value[1, 1, 0, 0])
self.assertEqual(edf.DT(9), maxp2.value[1, 1, 1, 0])
self.assertEqual(edf.DT(9), maxp2.value[1, 1, 2, 0])
self.assertEqual(edf.DT(9), maxp2.value[1, 1, 3, 0])