def test_previous_gradient_wrt_x_accounted(self):
op = Conv2D(self.x, self.y, padding='VALID', stride=1)
op.register(op)
op.new_context()
op.forward()
gradient = np.arange(16).reshape((2, 2, 2, 2))
op.accumulate(op, gradient)
expected = [[[[1.0, 3.0, 5.0, 7.0, 9.0],
[14.0, 26.0, 38.0, 50.0, 62.0],
[74.0, 86.0, 98.0, 110.0, 122.0],
[103.0, 113.0, 123.0, 133.0, 143.0]],
[[36.0, 56.0, 76.0, 96.0, 116.0],
[296.0, 352.0, 408.0, 464.0, 520.0],
[576.0, 632.0, 688.0, 744.0, 800.0],
[520.0, 556.0, 592.0, 628.0, 664.0]],
[[275.0, 293.0, 311.0, 329.0, 347.0],
[762.0, 806.0, 850.0, 894.0, 938.0],
[982.0, 1026.0, 1070.0, 1114.0, 1158.0],
[657.0, 683.0, 709.0, 735.0, 761.0]]],
[[[9.0, 43.0, 77.0, 111.0, 145.0],
[190.0, 266.0, 342.0, 418.0, 494.0],
[570.0, 646.0, 722.0, 798.0, 874.0],
[431.0, 473.0, 515.0, 557.0, 599.0]],
[[532.0, 616.0, 700.0, 784.0, 868.0],
[1608.0, 1792.0, 1976.0, 2160.0, 2344.0],
[2528.0, 2712.0, 2896.0, 3080.0, 3264.0],
[1656.0, 1756.0, 1856.0, 1956.0, 2056.0]],
[[763.0, 813.0, 863.0, 913.0, 963.0],
[1898.0, 2006.0, 2114.0, 2222.0, 2330.0],
[2438.0, 2546.0, 2654.0, 2762.0, 2870.0],
[1465.0, 1523.0, 1581.0, 1639.0, 1697.0]]]]
actual = self.x.gradient
self.assertTrue((expected == actual).all())
def test_gradient_wrt_x_with_no_padding_and_stride_equals_1(self):
da = [[[[1.0, 5.0, 9.0, 13.0, 17.0], [22.0, 30.0, 38.0, 46.0, 54.0],
[62.0, 70.0, 78.0, 86.0, 94.0], [41.0, 45.0, 49.0, 53.0,
57.0]],
[[62.0, 70.0, 78.0, 86.0, 94.0],
[164.0, 180.0, 196.0, 212.0, 228.0],
[244.0, 260.0, 276.0, 292.0, 308.0],
[142.0, 150.0, 158.0, 166.0, 174.0]],
[[61.0, 65.0, 69.0, 73.0, 77.0],
[142.0, 150.0, 158.0, 166.0, 174.0],
[182.0, 190.0, 198.0, 206.0, 214.0],
[101.0, 105.0, 109.0, 113.0, 117.0]]],
[[[1.0, 5.0, 9.0, 13.0, 17.0], [22.0, 30.0, 38.0, 46.0, 54.0],
[62.0, 70.0, 78.0, 86.0, 94.0], [41.0, 45.0, 49.0, 53.0,
57.0]],
[[62.0, 70.0, 78.0, 86.0, 94.0],
[164.0, 180.0, 196.0, 212.0, 228.0],
[244.0, 260.0, 276.0, 292.0, 308.0],
[142.0, 150.0, 158.0, 166.0, 174.0]],
[[61.0, 65.0, 69.0, 73.0, 77.0],
[142.0, 150.0, 158.0, 166.0, 174.0],
[182.0, 190.0, 198.0, 206.0, 214.0],
[101.0, 105.0, 109.0, 113.0, 117.0]]]]
op = Conv2D(self.x, self.y, padding='VALID', stride=1)
op.register(op)
op.new_context()
result = op.forward()
gradient = np.ones_like(result)
op.accumulate(op, gradient)
actual = self.x.gradient
self.assertTrue((da == actual).all())
def test_forward_output_with_no_padding_and_stride_equals_1(self):
op = Conv2D(self.x, self.y, padding='VALID', stride=1)
actual = op.forward()
self.assertEqual(actual.shape, (2, 2, 2, 2))
expected = [[[[20410, 20920], [24760, 25420]],
[[37810, 38920], [42160, 43420]]],
[[[72610, 74920], [76960, 79420]],
[[90010, 92920], [94360, 97420]]]]
self.assertTrue((expected == actual).all())
def test_gradient_wrt_y_with_no_padding_and_stride_equals_1(self):
db = [[[[340.0, 340.0], [348.0, 348.0], [356.0, 356.0], [364.0, 364.0],
[372.0, 372.0]],
[[380.0, 380.0], [388.0, 388.0], [396.0, 396.0], [404.0, 404.0],
[412.0, 412.0]],
[[420.0, 420.0], [428.0, 428.0], [436.0, 436.0], [444.0, 444.0],
[452.0, 452.0]]],
[[[500.0, 500.0], [508.0, 508.0], [516.0, 516.0], [524.0, 524.0],
[532.0, 532.0]],
[[540.0, 540.0], [548.0, 548.0], [556.0, 556.0], [564.0, 564.0],
[572.0, 572.0]],
[[580.0, 580.0], [588.0, 588.0], [596.0, 596.0], [604.0, 604.0],
[612.0, 612.0]]]]
op = Conv2D(self.x, self.y, padding='VALID', stride=1)
op.register(op)
op.new_context()
result = op.forward()
gradient = np.ones_like(result)
op.accumulate(op, gradient)
actual = self.y.gradient
self.assertTrue((db == actual).all())
def test_previous_gradient_wrt_y_accounted(self):
op = Conv2D(self.x, self.y, padding='VALID', stride=1)
op.register(op)
op.new_context()
op.forward()
gradient = np.arange(16).reshape((2, 2, 2, 2))
op.accumulate(op, gradient)
expected = [[[[3520.0, 3860.0], [3576.0, 3924.0], [3632.0, 3988.0],
[3688.0, 4052.0], [3744.0, 4116.0]],
[[3800.0, 4180.0], [3856.0, 4244.0], [3912.0, 4308.0],
[3968.0, 4372.0], [4024.0, 4436.0]],
[[4080.0, 4500.0], [4136.0, 4564.0], [4192.0, 4628.0],
[4248.0, 4692.0], [4304.0, 4756.0]]],
[[[4640.0, 5140.0], [4696.0, 5204.0], [4752.0, 5268.0],
[4808.0, 5332.0], [4864.0, 5396.0]],
[[4920.0, 5460.0], [4976.0, 5524.0], [5032.0, 5588.0],
[5088.0, 5652.0], [5144.0, 5716.0]],
[[5200.0, 5780.0], [5256.0, 5844.0], [5312.0, 5908.0],
[5368.0, 5972.0], [5424.0, 6036.0]]]]
actual = self.y.gradient
self.assertTrue((expected == actual).all())
def test_gradient_wrt_y_with_valid_padding_and_stride_equals_2(self):
db = [[[[60.0, 60.0], [62.0, 62.0], [64.0, 64.0], [66.0, 66.0],
[68.0, 68.0]],
[[70.0, 70.0], [72.0, 72.0], [74.0, 74.0], [76.0, 76.0],
[78.0, 78.0]],
[[80.0, 80.0], [82.0, 82.0], [84.0, 84.0], [86.0, 86.0],
[88.0, 88.0]]],
[[[100.0, 100.0], [102.0, 102.0], [104.0, 104.0], [106.0, 106.0],
[108.0, 108.0]],
[[110.0, 110.0], [112.0, 112.0], [114.0, 114.0], [116.0, 116.0],
[118.0, 118.0]],
[[120.0, 120.0], [122.0, 122.0], [124.0, 124.0], [126.0, 126.0],
[128.0, 128.0]]]]
op = Conv2D(self.x, self.y, padding='VALID', stride=2)
op.register(op)
op.new_context()
result = op.forward()
gradient = np.ones_like(result)
op.accumulate(op, gradient)
actual = self.y.gradient
self.assertTrue((db == actual).all())
def test_gradient_wrt_x_with_valid_padding_and_stride_equals_2(self):
da = [[[[1.0, 5.0, 9.0, 13.0, 17.0], [21.0, 25.0, 29.0, 33.0, 37.0],
[41.0, 45.0, 49.0, 53.0, 57.0], [0.0, 0.0, 0.0, 0.0, 0.0]],
[[61.0, 65.0, 69.0, 73.0, 77.0], [81.0, 85.0, 89.0, 93.0, 97.0],
[101.0, 105.0, 109.0, 113.0, 117.0], [0.0, 0.0, 0.0, 0.0,
0.0]],
[[0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0]]],
[[[1.0, 5.0, 9.0, 13.0, 17.0], [21.0, 25.0, 29.0, 33.0, 37.0],
[41.0, 45.0, 49.0, 53.0, 57.0], [0.0, 0.0, 0.0, 0.0, 0.0]],
[[61.0, 65.0, 69.0, 73.0, 77.0], [81.0, 85.0, 89.0, 93.0, 97.0],
[101.0, 105.0, 109.0, 113.0, 117.0], [0.0, 0.0, 0.0, 0.0,
0.0]],
[[0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0]]]]
op = Conv2D(self.x, self.y, padding='VALID', stride=2)
op.register(op)
op.new_context()
result = op.forward()
gradient = np.ones_like(result)
op.accumulate(op, gradient)
actual = self.x.gradient
self.assertTrue((da == actual).all())
def test_forward_output_with_valid_padding_and_stride_equals_2(self):
op = Conv2D(self.x, self.y, padding='VALID', stride=2)
actual = op.forward()
self.assertEqual(actual.shape, (2, 1, 1, 2))
expected = [[[[20410, 20920]]], [[[72610, 74920]]]]
self.assertTrue((expected == actual).all())