def testMulGrad(self):
d1 = np.random.uniform(0, 10, size=(5, 5))
d2 = np.random.uniform(0, 10, size=(5, 5))
t1 = sn.Tensor(d1, requires_grad=True)
t2 = sn.Tensor(d2, requires_grad=True)
t3 = t1 * t2 # 5x5
initialGrad = sn.Tensor(np.ones_like(t1.data))
t3.backward(initialGrad)
assert_array_equal(t1.grad.data, t2.data)
assert_array_equal(t2.grad.data, t1.data)
def testSumGrad(self):
d1 = np.random.uniform(0, 10, size=(5, 5))
t1 = sn.Tensor(d1, requires_grad=True)
t2 = t1.sum()
t2.backward()
assert_array_equal(t2.grad.data, np.ones_like(t1.data))
assert t2.requires_grad == True
def testNoGrad(self):
data1 = np.random.uniform(-10, 10, size=(5, 5))
data2 = np.random.uniform(-10, 10, size=(5, 5))
with sn.no_grad():
a = sn.Tensor(data1, requires_grad=True)
b = sn.Tensor(data2, requires_grad=True)
c = a.sum()
d = a + b
e = sn.matmul(a, b)
f = a * b
assert a.requires_grad == False
assert b.requires_grad == False
assert c.requires_grad == False
assert d.requires_grad == False
assert e.requires_grad == False
assert f.requires_grad == False
import synapse as sn
from synapse import Tensor
from synapse.autograd.tensor import Node
from synapse.nn.loss import MSE
from synapse.nn.activations import ReLU
from synapse.testing.graph import showParents
data = np.arange(1, 10, dtype=np.float)
data = np.expand_dims(data, 0)
x = Tensor(data)
print(x.shape)
W = Tensor(np.zeros_like(data), requiresGrad=True)
y = Tensor(2 * data)
maxEpoch = 1000
lr = 0.1
initialGrad = sn.Tensor([[1.0]])
for epoch in range(maxEpoch):
predicted = W * x
loss = MSE(predicted, y)
loss.backward()
W.data = W.data - lr * W.grad.data
W.zeroGrad()
print(loss.data)
print(f"Prediction: {W.data}")