def test_put_operation_gradient_is_correct():
a = Tensor([[1.0, 2.0], [3.0, 4.0]])
b = a.put([0], 1, 2.0, inplace=False)
assert np.equal(b.value, np.array([[2.0, 2.0], [2.0, 4.0]])).all()
b.backward()
assert np.equal(a.gradient, np.array([[0.0, 0.0], [0.0, 0.0]])).all()
def test_take_operation_gradient_is_correct():
a = Tensor([[1.0, 2.0], [3.0, 4.0]])
b = a.take([0], 1)
assert np.equal(b.value, np.array([[1.0], [3.0]])).all()
b.backward()
assert np.equal(a.gradient, np.array([[1.0, 0.0], [1.0, 0.0]])).all()
def test_sub_operation_gradient_is_correct():
a = Tensor([1.0, 2.0])
b = Tensor([1.0])
c = a - b
assert np.equal(c.value, np.array([0.0, 1.0])).all()
c.backward()
assert np.equal(a.gradient, np.array([1.0, 1.0])).all()
assert np.equal(b.gradient, np.array([-2.0])).all()
def test_mul_operation_gradient_is_correct():
a = Tensor([1.0, 2.0])
b = Tensor([2.0])
c = a * b
assert np.equal(c.value, np.array([2.0, 4.0])).all()
c.backward()
assert np.equal(a.gradient, np.array([2.0, 2.0])).all()
assert np.equal(b.gradient, np.array([3.0])).all()
def test_div_operation_gradient_is_correct():
a = Tensor([1.0, 2.0])
b = Tensor([2.0])
c = a / b
assert np.equal(c.value, np.array([0.5, 1.0])).all()
c.backward()
assert np.equal(a.gradient, np.array([0.5, 0.5])).all()
assert np.equal(b.gradient, np.array([-0.75])).all()
def test_dot_operation_gradient_is_correct():
a = Tensor([[1.0, 2.0]])
b = Tensor([[1.0], [2.0]])
c = a @ b
assert np.equal(c.value, np.array([5.0])).all()
c.backward()
assert np.equal(a.gradient, np.array([1.0, 2.0])).all()
assert np.equal(b.gradient, np.array([[1.0], [2.0]])).all()
def test_gradient_is_summed_correctly():
a = Tensor(np.ones((1, 2, 1, 1, 2)))
b = Tensor(np.ones((1, 2, 2, 1)))
c = a + b
assert np.equal(c.value,
np.ones((1, 2, 1, 1, 2)) + np.ones((1, 2, 2, 1))).all()
c.backward()
assert a.gradient.shape == a.shape
assert b.gradient.shape == b.shape
def test_stack_operation_gradient_is_correct():
a = Tensor([1.0])
b = Tensor([2.0])
c = Tensor([3.0])
d = Stack.forward([a, b, c], axis=1)
assert np.equal(d.value, np.array([[1.0, 2.0, 3.0]])).all()
d.backward()
assert np.equal(a.gradient, np.array([1.0])).all()
assert np.equal(b.gradient, np.array([1.0])).all()
assert np.equal(c.gradient, np.array([1.0])).all()
def generate_fib(n, c=1.0, prev=0, cc=0):
if cc == 0:
inputs.append(Tensor([1.0], requires_grad=False))
f = 1.0
c = 0.0
else:
f = c + prev
inputs.append(Tensor([f], requires_grad=False))
cc += 1
if cc <= n:
generate_fib(n, f, c, cc)
def test_getitem_operation_gradient_is_correct():
a = Tensor([[1.0, 2.0]])
b = a[0]
assert np.equal(b.value, np.array([1.0, 2.0])).all()
b.backward()
assert np.equal(a.gradient, np.array([[1.0, 1.0]])).all()
a = Tensor([[1.0, 2.0]])
b = a[0][1]
assert np.equal(b.value, np.array([2.0])).all()
b.backward()
assert np.equal(a.gradient, np.array([[0.0, 1.0]])).all()
def test_leakyrelu_operation_gradient_is_correct():
a = Tensor([2.0, -2.0])
b = leaky_relu.forward(a, a=0.1)
assert np.equal(b.value, np.array([2.0, -0.2])).all()
b.backward()
assert np.equal(a.gradient, np.array([1.0, 0.1])).all()
lr = LeakyRelu(0.1)
a = Tensor([2.0, -2.0])
b = lr(a)
assert np.equal(b.value, np.array([2.0, -0.2])).all()
b.backward()
assert np.equal(a.gradient, np.array([1.0, 0.1])).all()
def test_relu_operation_gradient_is_correct():
a = Tensor([2.0, -2.0])
b = Relu.forward(a)
assert np.equal(b.value, np.array([2.0, 0.0])).all()
b.backward()
assert np.equal(a.gradient, np.array([1.0, 0.0])).all()
def test_fliplr_operation_gradient_is_correct():
a = Tensor([[1.0, 2.0]])
b = FlipLR.forward(a)
assert np.equal(b.value, np.array([[2.0, 1.0]])).all()
b = b.mul([[1.0, 2.0]])
b.backward()
assert np.equal(a.gradient, np.array([[2.0, 1.0]])).all()
def test_sigmoid_operation_gradient_is_correct():
a = Tensor([2.0])
b = Sigmoid.forward(a)
sig = 1 / (1 + math.exp(-2))
assert np.equal(b.value, np.array([sig])).all()
b.backward()
assert np.equal(a.gradient, np.array([sig * (1 - sig)])).all()
def test_tanh_operation_gradient_is_correct():
a = Tensor([2.0])
b = Tanh.forward(a)
tanh = np.tanh(2.0)
assert np.equal(b.value, np.array([tanh])).all()
b.backward()
assert np.equal(a.gradient, np.array([1 - tanh**2])).all()
def test_softmax_operation_gradient_is_correct():
a = Tensor([1.0, 2.0, 3.0])
b = softmax.forward(a)
assert np.isclose(b.value, np.array([0.09003057, 0.24472847,
0.66524096])).all()
b.backward()
assert np.isclose(a.gradient,
np.array([0.08192507, 0.18483645, 0.22269543])).all()
a = Tensor([1.0, 2.0, 3.0])
sm = Softmax(0)
b = sm(a)
assert np.isclose(b.value, np.array([0.09003057, 0.24472847,
0.66524096])).all()
b.backward()
assert np.isclose(a.gradient,
np.array([0.08192507, 0.18483645, 0.22269543])).all()
def forward(cls, *args, **kargs):
from np.ad.nn import Tensor, Context, Variable
if isinstance(args[0], (tuple, list)):
parents = list(args[0])
args = [parents]
else:
parents = list(args)
args = parents
rg = False
for i in range(len(parents)):
p = parents[i]
if isinstance(p, Variable):
parents[i] = p = p.tensor
elif not isinstance(p, Tensor):
parents[i] = p = Tensor(p, requires_grad=False, retain_grad=False)
rg = rg or p.requires_grad
context = Context()
result = cls.forward_(context, *args, **kargs)
return Tensor(result, operation=cls, parents=parents, context=context, requires_grad=rg)
def test_pow_operation_gradient_is_correct():
a = Tensor([1.0, 2.0])
b = Tensor([2.0])
c = a**b
assert np.equal(c.value, np.array([1.0, 4.0])).all()
c.backward()
bg = 4.0 * np.log(np.array([2.0]))
assert np.equal(a.gradient, np.array([1.0, 2.0])).all()
assert np.equal(b.gradient, bg).all()
a = Tensor([2.0, 3.0])
b = Tensor([1.5, 2.0])
c = a**b
res = np.array([2.0**1.5, 9.0])
assert np.equal(c.value, res).all()
c.backward()
bg = res * np.log(np.array([2.0, 3.0]))
assert np.equal(a.gradient, np.array([2.0**0.5, 3.0])).all()
assert np.equal(b.gradient, bg).all()
def test_transpose_operation_gradient_is_correct():
a = Tensor([[1.0, 2.0]])
b = a.transpose()
assert np.equal(b.value, np.array([[1.0], [2.0]])).all()
b.backward()
assert np.equal(a.gradient, np.array([[1.0, 1.0]])).all()
a = Tensor([[1.0], [2.0]])
b = a.transpose(axes=(1, 0))
assert np.equal(b.value, np.array([[1.0, 2.0]])).all()
b.backward()
assert np.equal(a.gradient, np.array([[1.0], [1.0]])).all()
def test_reshape_operation_gradient_is_correct():
a = Tensor([[1.0, 2.0]])
b = a.reshape((2, 1))
assert np.equal(b.value, np.array([[1.0], [2.0]])).all()
b.backward()
assert np.equal(a.gradient, np.array([[1.0, 1.0]])).all()
a = Tensor([[1.0], [2.0]])
b = a.reshape((1, 2))
assert np.equal(b.value, np.array([[1.0, 2.0]])).all()
b.backward()
assert np.equal(a.gradient, np.array([[1.0], [1.0]])).all()
def test_sum_operation_gradient_is_correct():
a = Tensor([1.0, 2.0])
b = a.sum()
assert np.equal(b.value, np.array([3.0])).all()
b.backward()
assert np.equal(a.gradient, np.array([1.0, 1.0])).all()
a = Tensor([[1.0, 2.0]])
b = a.sum(axis=0, keepdims=True)
assert np.equal(b.value, np.array([1.0, 2.0])).all()
b.backward()
assert np.equal(a.gradient, np.array([1.0, 1.0])).all()
def test_conv2dtranspose_operation_gradient_is_correct():
####################################################################################################################
cx = 2
cy = 1
cz = 0
batch = 0
channel = 1
kx, ky, kz = 2, 1, 1
sx, sy, sz = 1, 1, 1
x = Tensor(np.arange(1, cy * cx + 1).reshape(cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d_transpose.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(y.value, np.array([[[2.0, 5.0, 2.0]]])).all()
nx, ny, nz = conv_t_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(np.arange(1, ny * nx + 1).reshape(1, ny, nx))
y.backward()
assert np.equal(x.gradient, np.array([[4.0, 7.0]])).all()
assert np.equal(k.gradient, np.array([[8.0, 5.0]])).all()
####################################################################################################################
cx = 2
cy = 2
cz = 0
batch = 0
channel = 1
kx, ky, kz = 2, 2, 1
sx, sy, sz = 1, 1, 1
x = Tensor(np.arange(1, cy * cx + 1).reshape(cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d_transpose.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(
y.value,
np.array([[[4.0, 11.0, 6.0], [14.0, 30.0, 14.0], [6.0, 11.0,
4.0]]])).all()
nx, ny, nz = conv_t_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(np.arange(1, ny * nx + 1).reshape(1, ny, nx))
y.backward()
assert np.equal(x.gradient, np.array([[23.0, 33.0], [53.0, 63.0]])).all()
assert np.equal(k.gradient, np.array([[77.0, 67.0, 47.0, 37.0]])).all()
####################################################################################################################
cx = 2
cy = 2
cz = 0
batch = 0
channel = 1
kx, ky, kz = 2, 2, 1
sx, sy, sz = 2, 2, 1
x = Tensor(np.arange(1, cy * cx + 1).reshape(cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d_transpose.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(
y.value,
np.array([[[4.0, 3.0, 8.0, 6.0], [2.0, 1.0, 4.0, 2.0],
[12.0, 9.0, 16.0, 12.0], [6.0, 3.0, 8.0, 4.0]]])).all()
nx, ny, nz = conv_t_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(np.arange(1, ny * nx + 1).reshape(1, ny, nx))
y.backward()
assert np.equal(x.gradient, np.array([[26.0, 46.0], [106.0, 126.0]])).all()
assert np.equal(k.gradient, np.array([128.0, 118.0, 88.0, 78.0])).all()
####################################################################################################################
cx = 2
cy = 1
cz = 0
batch = 0
channel = 2
kx, ky, kz = 2, 1, 1
sx, sy, sz = 1, 1, 1
x = Tensor(np.arange(1, cy * cx + 1).reshape(cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d_transpose.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(y.value, np.array([[[2.0, 5.0, 2.0]], [[4.0, 11.0,
6.0]]])).all()
nx, ny, nz = conv_t_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(np.arange(1, channel * ny * nx + 1).reshape(channel, ny, nx))
y.backward()
assert np.equal(x.gradient, np.array([[58.0, 78.0]])).all()
assert np.equal(k.gradient, np.array([[8.0, 5.0], [17.0, 14.0]])).all()
####################################################################################################################
cx = 2
cy = 2
cz = 1
batch = 0
channel = 2
kx, ky, kz = 2, 2, 1
sx, sy, sz = 2, 2, 1
x = Tensor(np.arange(1, cz * cy * cx + 1).reshape(cz, cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d_transpose.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(
y.value,
np.array([[[4.0, 3.0, 8.0, 6.0], [2.0, 1.0, 4.0, 2.0],
[12.0, 9.0, 16.0, 12.0], [6.0, 3.0, 8.0, 4.0]],
[[8.0, 7.0, 16.0, 14.0], [6.0, 5.0, 12.0, 10.0],
[24.0, 21.0, 32.0, 28.0], [18.0, 15.0, 24.0,
20.0]]])).all()
nx, ny, nz = conv_t_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(
np.arange(1, channel * nz * ny * nx + 1).reshape(channel * nz, ny, nx))
y.backward()
assert np.equal(x.gradient, np.array([[792.0, 936.0], [1368.0,
1512.0]])).all()
assert np.equal(
k.gradient,
np.array([[128.0, 118.0, 88.0, 78.0], [288.0, 278.0, 248.0,
238.0]])).all()
####################################################################################################################
cx = 2
cy = 2
cz = 2
batch = 0
channel = 2
kx, ky, kz = 2, 2, 2
sx, sy, sz = 2, 2, 2
x = Tensor(np.arange(1, cz * cy * cx + 1).reshape(cz, cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d_transpose.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(
y.value,
np.array([[[8.0, 7.0, 16.0, 14.0], [6.0, 5.0, 12.0, 10.0],
[24.0, 21.0, 32.0, 28.0], [18.0, 15.0, 24.0, 20.0]],
[[4.0, 3.0, 8.0, 6.0], [2.0, 1.0, 4.0, 2.0],
[12.0, 9.0, 16.0, 12.0], [6.0, 3.0, 8.0, 4.0]],
[[40.0, 35.0, 48.0, 42.0], [30.0, 25.0, 36.0, 30.0],
[56.0, 49.0, 64.0, 56.0], [42.0, 35.0, 48.0, 40.0]],
[[20.0, 15.0, 24.0, 18.0], [10.0, 5.0, 12.0, 6.0],
[28.0, 21.0, 32.0, 24.0], [14.0, 7.0, 16.0, 8.0]],
[[16.0, 15.0, 32.0, 30.0], [14.0, 13.0, 28.0, 26.0],
[48.0, 45.0, 64.0, 60.0], [42.0, 39.0, 56.0, 52.0]],
[[12.0, 11.0, 24.0, 22.0], [10.0, 9.0, 20.0, 18.0],
[36.0, 33.0, 48.0, 44.0], [30.0, 27.0, 40.0, 36.0]],
[[80.0, 75.0, 96.0, 90.0], [70.0, 65.0, 84.0, 78.0],
[112.0, 105.0, 128.0, 120.0], [98.0, 91.0, 112.0, 104.0]],
[[60.0, 55.0, 72.0, 66.0], [50.0, 45.0, 60.0, 54.0],
[84.0, 77.0, 96.0, 88.0], [70.0, 63.0, 80.0,
72.0]]])).all()
nx, ny, nz = conv_t_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(
np.arange(1, channel * nz * ny * nx + 1).reshape(nz * channel, ny, nx))
y.backward()
"""assert np.equal(x.gradient, np.array([
[[8.0, 7.0],
[6.0, 5.0]],
[[4.0, 3.0],
[2.0, 1.0]]
])).all()"""
assert np.equal(
k.gradient,
np.array([[1840., 1804., 1696., 1660., 1264., 1228., 1120., 1084.],
[4144., 4108., 4000., 3964., 3568., 3532., 3424.,
3388.]])).all()
####################################################################################################################
cx = 1
cy = 1
cz = 2
batch = 2
channel = 2
kx, ky, kz = 1, 1, 1
sx, sy, sz = 1, 1, 1
x = Tensor(
np.arange(1, batch * cz * cy * cx + 1).reshape(batch, cz, cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d_transpose.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(
y.value,
np.array([[[[1.0]], [[2.0]], [[2.0]], [[4.0]]],
[[[3.0]], [[4.0]], [[6.0]], [[8.0]]]])).all()
nx, ny, nz = conv_t_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(
np.arange(1, batch * channel * nz * ny * nx + 1).reshape(
batch, nz * channel, ny, nx))
y.backward()
assert np.equal(x.gradient,
np.array([[[[12.0]], [[18.0]]], [[[36.0]],
[[42.0]]]])).all()
assert np.equal(k.gradient, np.array([[44.0], [64.0]])).all()
def test_powi_operation_gradient_is_correct():
a = Tensor([1.0, 2.0])
c = a.powi(2)
assert np.equal(c.value, np.array([1.0, 4.0])).all()
c.backward()
assert np.equal(a.gradient, np.array([2.0, 4.0])).all()
from np.ad.nn import Tensor, MSE_Loss
from np.ad.operation import conv1d
import numpy as np
inp = []
inp.append([[1.0, 2.0, 3.0],
[4.0, 5.0, 6.0],
[7.0, 8.0, 9.0]])
inp.append([[2.0, 4.0, 6.0],
[8.0, 10.0, 12.0],
[14.0, 16.0, 18.0]])
input = Tensor(np.array(inp), requires_grad=True)
k = Tensor(np.array([[1.0, 2.0, 1.0, 5.0],
[1.0, 3.0, 4.0, 1.0]]))
tar = []
tar.append([[12.0, 16.0],
[24.0, 28.0],
[24.0, 32.0],
[48.0, 56.0]])
tar.append([[24.0, 32.0],
[48.0, 56.0],
[48.0, 64.0],
[56.0, 112.0]])
target = Tensor(np.array(tar))
mse = MSE_Loss()
for e in range(30000):
y = conv1d.forward(input, k, kx=2, ky=2, channel=2, stride=(1, 1))
loss = mse(y, target, (0, 1, 2))
def test():
l = 0
correct = 0
for i, (x, label) in enumerate(testloader):
x = Tensor(x.view(-1, INPUT_SIZE).numpy(), requires_grad=False)
label = Tensor(label.numpy().reshape(label.size(0), 1),
requires_grad=False)
x_ = model(x)
x_label = Tensor((x_.shape[0], OUTPUT_SIZE), requires_grad=False)
x_label.put_(label, 1, 1.0)
preds = x_.argmax(1).reshape((label.shape[0], 1))
correct += (preds == label).sum().item()
loss = criterion(x_, x_label, axis=(0, 1))
l += loss
return l / len(testset), correct / len(testset)
def test_conv2d_operation_gradient_is_correct():
####################################################################################################################
cx = 4
cy = 1
cz = 0
batch = 0
channel = 1
kx, ky, kz = 2, 1, 1
sx, sy, sz = 1, 1, 1
x = Tensor(np.arange(1, cy * cx + 1).reshape(cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(y.value, np.array([[[5.0, 8.0, 11.0]]])).all()
nx, ny, nz = conv_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(np.arange(1, ny * nx + 1).reshape(1, ny, nx))
y.backward()
assert np.equal(x.gradient, np.array([[1.0, 4.0, 7.0, 6.0]])).all()
assert np.equal(k.gradient, np.array([[14.0, 20.0]])).all()
####################################################################################################################
cx = 4
cy = 4
cz = 0
batch = 0
channel = 1
kx, ky, kz = 2, 2, 1
sx, sy, sz = 1, 1, 1
x = Tensor(np.arange(1, cy * cx + 1).reshape(cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(
y.value,
np.array([[[44.0, 54.0, 64.0], [84.0, 94.0, 104.0],
[124.0, 134.0, 144.0]]])).all()
nx, ny, nz = conv_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(np.arange(1, ny * nx + 1).reshape(1, ny, nx))
y.backward()
assert np.equal(
x.gradient,
np.array([[1.0, 4.0, 7.0, 6.0], [7.0, 23.0, 33.0, 24.0],
[19.0, 53.0, 63.0, 42.0], [21.0, 52.0, 59.0, 36.0]])).all()
assert np.equal(k.gradient, np.array([[348.0, 393.0, 528.0, 573.0]])).all()
####################################################################################################################
cx = 4
cy = 4
cz = 0
batch = 0
channel = 1
kx, ky, kz = 2, 2, 1
sx, sy, sz = 2, 2, 1
x = Tensor(np.arange(1, cy * cx + 1).reshape(cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(y.value, np.array([[[44.0, 64.0], [124.0, 144.0]]])).all()
nx, ny, nz = conv_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(np.arange(1, ny * nx + 1).reshape(1, ny, nx))
y.backward()
assert np.equal(
x.gradient,
np.array([[1.0, 2.0, 2.0, 4.0], [3.0, 4.0, 6.0, 8.0],
[3.0, 6.0, 4.0, 8.0], [9.0, 12.0, 12.0, 16.0]])).all()
assert np.equal(k.gradient, np.array([[78.0, 88.0, 118.0, 128.0]])).all()
####################################################################################################################
cx = 4
cy = 2
cz = 0
batch = 0
channel = 2
kx, ky, kz = 2, 2, 1
sx, sy, sz = 2, 2, 1
x = Tensor(np.arange(1, cy * cx + 1).reshape(cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(y.value, np.array([[[44.0, 64.0]], [[100.0,
152.0]]])).all()
nx, ny, nz = conv_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(np.arange(1, ny * nx + 1).reshape(1, ny, nx))
y.backward()
assert np.equal(
x.gradient, np.array([[6.0, 8.0, 12.0, 16.0], [10.0, 12.0, 20.0,
24.0]])).all()
assert np.equal(
k.gradient, np.array([[7.0, 10.0, 19.0, 22.0], [7.0, 10.0, 19.0,
22.0]])).all()
####################################################################################################################
cx = 4
cy = 2
cz = 1
batch = 0
channel = 2
kx, ky, kz = 2, 2, 1
sx, sy, sz = 2, 2, 1
x = Tensor(np.arange(1, cz * cy * cx + 1).reshape(cz, cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(y.value, np.array([[[44.0, 64.0]], [[100.0,
152.0]]])).all()
nx, ny, nz = conv_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(
np.arange(1, channel * nz * ny * nx + 1).reshape(channel * nz, ny, nx))
y.backward()
assert np.equal(
x.gradient,
np.array([[16.0, 20.0, 22.0, 28.0], [24.0, 28.0, 34.0, 40.0]])).all()
assert np.equal(
k.gradient,
np.array([[7.0, 10.0, 19.0, 22.0], [15.0, 22.0, 43.0, 50.0]])).all()
####################################################################################################################
cx = 4
cy = 4
cz = 4
batch = 0
channel = 2
kx, ky, kz = 2, 2, 2
sx, sy, sz = 2, 2, 2
x = Tensor(np.arange(1, cz * cy * cx + 1).reshape(cz, cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(
y.value,
np.array([[[560.0, 632.0], [848.0, 920.0]],
[[1712.0, 1784.0], [2000.0, 2072.0]],
[[1296.0, 1496.0], [2096.0, 2296.0]],
[[4496.0, 4696.0], [5296.0, 5496.0]]])).all()
nx, ny, nz = conv_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(
np.arange(1, channel * nz * ny * nx + 1).reshape(nz * channel, ny, nx))
y.backward()
# assert np.equal(x.gradient, np.array().all()
assert np.equal(
k.gradient,
np.array([[1084., 1120., 1228., 1264., 1660., 1696., 1804., 1840.],
[2492., 2592., 2892., 2992., 4092., 4192., 4492.,
4592.]])).all()
####################################################################################################################
cx = 2
cy = 2
cz = 1
batch = 2
channel = 2
kx, ky, kz = 2, 2, 1
sx, sy, sz = 1, 1, 1
x = Tensor(
np.arange(1, batch * cz * cy * cx + 1).reshape(batch, cz, cy, cx))
k = Tensor(
np.arange(1,
channel * kz * ky * kx + 1).reshape(channel, kz * ky * kx))
y = conv2d.forward(x,
k,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
assert np.equal(y.value,
np.array([[[[30.0]], [[70.0]]], [[[70.0]],
[[174.0]]]])).all()
nx, ny, nz = conv_shape(cx, cy, cz, kx, ky, kz, sx, sy, sz)
y = y.mul(
np.arange(1, batch * channel * nz * ny * nx + 1).reshape(
batch, nz * channel, ny, nx))
y.backward()
assert np.equal(
x.gradient,
np.array([[[[11.0, 14.0], [17.0, 20.0]]], [[[23.0, 30.0],
[37.0, 44.0]]]])).all()
assert np.equal(k.gradient,
np.array([[16., 20., 24., 28.], [22., 28., 34.,
40.]])).all()
from np.ad.nn import Tensor, MSE_Loss
import np.ad.functional as F
from np.ad.operation import conv2d
import numpy as np
cx = 4
cy = 4
cz = 2
batch = 2
channel = 2
kx, ky, kz = 2, 2, 1
sx, sy, sz = 1, 1, 1
input = Tensor(
np.arange(1, batch * cz * cy * cx + 1,
dtype=np.float).reshape(batch, cz, cy, cx))
k = Tensor(
np.arange(1, channel * kz * ky * kx + 1,
dtype=np.float).reshape(channel, kz * ky * kx))
tar_k = Tensor(k.value * 2.0)
v, _, _ = F.conv2d(input.value,
tar_k.value,
kx=kx,
ky=ky,
kz=kz,
channel=channel,
stride=(sx, sy, sz))
tar_y = Tensor(v)
mse = MSE_Loss()
for e in range(0):
y = conv2d.forward(input,
def test_conv1dtranspose_operation_gradient_is_correct():
x = Tensor([[1.0, 2.0, 3.0]])
k = Tensor([[1.0, 2.0]])
y = conv1d_transpose.forward(x, k, kx=2, ky=1, channel=1, stride=(1, 1))
assert np.equal(y.value, np.array([[2.0, 5.0, 8.0, 3.0]])).all()
y = y.mul([[1.0, 2.0, 3.0, 4.0]])
y.backward()
assert np.equal(x.gradient, np.array([[4.0, 7.0, 10.0]])).all()
assert np.equal(k.gradient, np.array([[20.0, 14.0]])).all()
x = Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
k = Tensor([[1.0, 2.0]])
y = conv1d_transpose.forward(x, k, kx=2, ky=1, channel=1, stride=(1, 1))
assert np.equal(y.value,
np.array([[2.0, 5.0, 8.0, 3.0], [8.0, 14.0, 17.0,
6.0]])).all()
y = y.mul([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]])
y.backward()
assert np.equal(x.gradient, np.array([[4.0, 7.0, 10.0], [16.0, 19.0,
22.0]])).all()
assert np.equal(k.gradient, np.array([[127.0, 106.0]])).all()
x = Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
k = Tensor([[1.0, 2.0, 3.0, 4.0]])
y = conv1d_transpose.forward(x, k, kx=2, ky=2, channel=1, stride=(1, 1))
assert np.equal(
y.value,
np.array([[4.0, 11.0, 18.0, 9.0], [18.0, 37.0, 47.0, 21.0],
[8.0, 14.0, 17.0, 6.0]])).all()
y = y.mul([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0],
[9.0, 10.0, 11.0, 12.0]])
y.backward()
assert np.equal(x.gradient,
np.array([[26.0, 36.0, 46.0], [66.0, 76.0, 86.0]])).all()
assert np.equal(k.gradient, np.array([[211.0, 190.0, 127.0, 106.0]])).all()
x = Tensor([[1.0, 2.0, 3.0, 4.0, 5.0]])
k = Tensor([[1.0, 2.0]])
y = conv1d_transpose.forward(x, k, kx=2, ky=1, channel=1, stride=(3, 1))
assert np.equal(
y.value,
np.array([[
2.0, 1.0, 0.0, 4.0, 2.0, 0.0, 6.0, 3.0, 0.0, 8.0, 4.0, 0.0, 10.0,
5.0
]])).all()
y = y.mul([[
1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0,
14.0
]])
y.backward()
assert np.equal(x.gradient, np.array([[4.0, 13.0, 22.0, 31.0,
40.0]])).all()
assert np.equal(k.gradient, np.array([[150.0, 135.0]])).all()
x = Tensor([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0],
[9.0, 10.0, 11.0, 12.0]])
k = Tensor([[1.0, 2.0]])
y = conv1d_transpose.forward(x, k, kx=2, ky=1, channel=1, stride=(1, 2))
assert np.equal(
y.value,
np.array([[2.0, 5.0, 8.0, 11.0, 4.0], [0.0, 0.0, 0.0, 0.0, 0.0],
[10.0, 17.0, 20.0, 23.0, 8.0], [0.0, 0.0, 0.0, 0.0, 0.0],
[18.0, 29.0, 32.0, 35.0, 12.0]])).all()
y = y.mul([[1.0, 2.0, 3.0, 4.0, 5.0], [6.0, 7.0, 8.0, 9.0, 10.0],
[11.0, 12.0, 13.0, 14.0, 15.0], [16.0, 17.0, 18.0, 19.0, 20.0],
[21.0, 22.0, 23.0, 24.0, 25.0]])
y.backward()
assert np.equal(
x.gradient,
np.array([[4.0, 7.0, 10.0, 13.0], [34.0, 37.0, 40.0, 43.0],
[64.0, 67.0, 70.0, 73.0]])).all()
assert np.equal(k.gradient, np.array([[1388.0, 1310.0]])).all()
x = Tensor([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0],
[9.0, 10.0, 11.0, 12.0]])
k = Tensor([[1.0, 2.0]])
y = conv1d_transpose.forward(x, k, kx=2, ky=1, channel=1, stride=(2, 2))
assert np.equal(
y.value,
np.array([[2.0, 1.0, 4.0, 2.0, 6.0, 3.0, 8.0, 4.0],
[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[10.0, 5.0, 12.0, 6.0, 14.0, 7.0, 16.0, 8.0],
[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[18.0, 9.0, 20.0, 10.0, 22.0, 11.0, 24.0, 12.0]])).all()
y = y.mul([[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0],
[9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0],
[17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0],
[25.0, 26.0, 27.0, 28.0, 29.0, 30.0, 31.0, 32.0],
[33.0, 34.0, 35.0, 36.0, 37.0, 38.0, 39.0, 40.0]])
y.backward()
assert np.equal(
x.gradient,
np.array([[4.0, 10.0, 16.0, 22.0], [52.0, 58.0, 64.0, 70.0],
[100.0, 106.0, 112.0, 118.0]])).all()
assert np.equal(k.gradient, np.array([[2180.0, 2102.0]])).all()
x = Tensor([[1.0, 2.0, 3.0]])
k = Tensor([[1.0, 2.0], [3.0, 4.0]])
y = conv1d_transpose.forward(x, k, kx=2, ky=1, channel=2, stride=(1, 1))
assert np.equal(y.value,
np.array([[2.0, 5.0, 8.0, 3.0], [4.0, 11.0, 18.0,
9.0]])).all()
y = y.mul([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]])
y.backward()
assert np.equal(x.gradient, np.array([[68.0, 88.0, 108.0]])).all()
assert np.equal(k.gradient, np.array([[20.0, 14.0], [44.0, 38.0]])).all()
def test():
l = 0
correct = 0
for i, (x, label) in enumerate(testloader):
x = Tensor(x.numpy(), requires_grad=False)
label = Tensor(label.numpy(), requires_grad=False).reshape((x.size(0), 1, 1))
x_ = model(x)
x_label = Tensor((x.size(0), 1, OUTPUT_SIZE), requires_grad=False)
x_label.put_(label, 2, 1.0)
preds = x_.argmax(2).reshape((x.size(0), 1, 1))
correct += (preds == label).sum(axis=(0, 1, 2)).item()
loss = criterion(x_, x_label, axis=(0, 1, 2))
l += loss.item()
optim.zero_grad()
return l / len(testset), correct / len(testset)
def test_conv1d_operation_gradient_is_correct():
x = Tensor([[1.0, 2.0, 3.0, 4.0, 5.0]])
k = Tensor([[1.0, 2.0, 3.0]])
y = conv1d.forward(x, k, kx=3, ky=1, channel=1, stride=(1, 1))
assert np.equal(y.value, np.array([[14.0, 20.0, 26.0]])).all()
y = y.mul([[1.0, 2.0, 3.0]])
y.backward()
assert np.equal(x.gradient, np.array([[1.0, 4.0, 10.0, 12.0, 9.0]])).all()
assert np.equal(k.gradient, np.array([[14.0, 20.0, 26.0]])).all()
x = Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
k = Tensor([[1.0, 2.0]])
y = conv1d.forward(x, k, kx=2, ky=1, channel=1, stride=(1, 1))
assert np.equal(y.value, np.array([[5.0, 8.0], [14.0, 17.0]])).all()
y = y.mul([[1.0, 2.0], [3.0, 4.0]])
y.backward()
assert np.equal(x.gradient, np.array([[1.0, 4.0, 4.0], [3.0, 10.0,
8.0]])).all()
assert np.equal(k.gradient, np.array([[37.0, 47.0]])).all()
x = Tensor([[1.0, 2.0, 3.0]])
k = Tensor([[1.0, 2.0], [3.0, 4.0]])
y = conv1d.forward(x, k, kx=2, ky=1, channel=2, stride=(1, 1))
assert np.equal(y.value, np.array([[5.0, 8.0], [11.0, 18.0]])).all()
y = y.mul([[1.0, 2.0], [3.0, 4.0]])
y.backward()
assert np.equal(x.gradient, np.array([[10.0, 28.0, 20.0]])).all()
assert np.equal(k.gradient, np.array([[5.0, 8.0], [11.0, 18.0]])).all()
x = Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
k = Tensor([[1.0, 2.0], [3.0, 4.0]])
y = conv1d.forward(x, k, kx=2, ky=1, channel=2, stride=(1, 1))
assert np.equal(
y.value, np.array([[5.0, 8.0], [14.0, 17.0], [11., 18.],
[32., 39.]])).all()
y = y.mul([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]])
y.backward()
assert np.equal(x.gradient,
np.array([[16.0, 42.0, 28.0], [24.0, 62.0, 40.0]])).all()
assert np.equal(k.gradient, np.array([[37.0, 47.0], [85.0, 111.0]])).all()
x = Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
k = Tensor([[1.0, 2.0, 3.0, 4.0]])
y = conv1d.forward(x, k, kx=2, ky=2, channel=1, stride=(1, 1))
assert np.equal(y.value, np.array([[37.0, 47.0]])).all()
y = y.mul([[1.0, 2.0]])
y.backward()
assert np.equal(x.gradient, np.array([[1.0, 4.0, 4.0], [3.0, 10.0,
8.0]])).all()
assert np.equal(k.gradient, np.array([[5.0, 8.0, 14.0, 17.0]])).all()
x = Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
k = Tensor([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]])
y = conv1d.forward(x, k, kx=2, ky=2, channel=2, stride=(1, 1))
assert np.equal(y.value, np.array([[37.0, 47.0], [85.0, 111.0]])).all()
y = y.mul([[1.0, 2.0], [3.0, 4.0]])
y.backward()
assert np.equal(x.gradient,
np.array([[16.0, 42.0, 28.0], [24.0, 62.0, 40.0]])).all()
assert np.equal(
k.gradient, np.array([[5.0, 8.0, 14.0, 17.0], [11.0, 18.0, 32.0,
39.0]])).all()
x = Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]])
k = Tensor([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]])
y = conv1d.forward(x, k, kx=2, ky=2, channel=2, stride=(1, 1))
assert np.equal(
y.value,
np.array([[37.0, 47.0], [67.0, 77.0], [85.0, 111.0], [163.0,
189.0]])).all()
y = y.mul([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]])
y.backward()
assert np.equal(
x.gradient,
np.array([[26.0, 64.0, 40.0], [76.0, 184.0, 112.0],
[58.0, 136.0, 80.0]])).all()
assert np.equal(
k.gradient,
np.array([[37.0, 47.0, 67.0, 77.0], [85.0, 111.0, 163.0,
189.0]])).all()
x = Tensor([[1.0, 2.0, 3.0, 4.0, 5.0]])
k = Tensor([[1.0, 2.0, 3.0]])
y = conv1d.forward(x, k, kx=3, ky=1, channel=1, stride=(2, 1))
assert np.equal(y.value, np.array([[14.0, 26.0]])).all()
y = y.mul([[1.0, 2.0]])
y.backward()
assert np.equal(x.gradient, np.array([[1.0, 2.0, 5.0, 4.0, 6.0]])).all()
assert np.equal(k.gradient, np.array([[7.0, 10.0, 13.0]])).all()
x = Tensor([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0],
[9.0, 10.0, 11.0, 12.0]])
k = Tensor([[1.0, 2.0]])
y = conv1d.forward(x, k, kx=2, ky=1, channel=1, stride=(2, 2))
assert np.equal(y.value, np.array([[5.0, 11.0], [29.0, 35.0]])).all()
y = y.mul([[1.0, 2.0], [3.0, 4.0]])
y.backward()
assert np.equal(
x.gradient,
np.array([[1.0, 2.0, 2.0, 4.0], [0.0, 0.0, 0.0, 0.0],
[3.0, 6.0, 4.0, 8.0]])).all()
assert np.equal(k.gradient, np.array([[78.0, 88.0]])).all()
x = Tensor([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0],
[9.0, 10.0, 11.0, 12.0]])
k = Tensor([[1.0, 2.0], [3.0, 4.0]])
y = conv1d.forward(x, k, kx=2, ky=1, channel=2, stride=(2, 2))
assert np.equal(
y.value,
np.array([[5.0, 11.0], [29.0, 35.0], [11.0, 25.0], [67.0,
81.0]])).all()
y = y.mul([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]])
y.backward()
assert np.equal(
x.gradient,
np.array([[16.0, 22.0, 20.0, 28.0], [0.0, 0.0, 0.0, 0.0],
[24.0, 34.0, 28.0, 40.0]])).all()
assert np.equal(k.gradient, np.array([[78.0, 88.0], [174.0, 200.0]])).all()
x = Tensor([[1.0, 2.0, 3.0, 4.0, 5.0], [6.0, 7.0, 8.0, 9.0, 10.0],
[11.0, 12.0, 13.0, 14.0, 15.0], [16.0, 17.0, 18.0, 19.0, 20.0],
[21.0, 22.0, 23.0, 24.0, 25.0]])
k = Tensor([[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0],
[10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0]])
y = conv1d.forward(x, k, kx=3, ky=3, channel=2, stride=(2, 2))
assert np.equal(
y.value,
np.array([[411.0, 501.0], [861.0, 951.0], [978.0, 1230.0],
[2238.0, 2490.0]])).all()
y = y.mul([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]])
y.backward()
assert np.equal(
x.gradient,
np.array([[51.0, 57.0, 125.0, 70.0, 78.0],
[69.0, 75.0, 167.0, 94.0, 102.0],
[160.0, 176.0, 386.0, 214.0, 234.0],
[103.0, 113.0, 243.0, 132.0, 144.0],
[133.0, 143.0, 309.0, 168.0, 180.0]])).all()
assert np.equal(
k.gradient,
np.array(
[[92.0, 102.0, 112.0, 142.0, 152.0, 162.0, 192.0, 202.0, 212.0],
[204.0, 230.0, 256.0, 334.0, 360.0, 386.0, 464.0, 490.0,
516.0]])).all()
