def test_sanity_check():
x = Value(-4.0, _op="x")
z = 2 * x + 2 + x
q = z.relu() + z * x
h = (z * z).relu()
y = h + q + q * x
y.backward()
xmg, ymg = x, y
x = torch.Tensor([-4.0]).double()
x.requires_grad = True
z = 2 * x + 2 + x
q = z.relu() + z * x
h = (z * z).relu()
y = h + q + q * x
y.backward()
xpt, ypt = x, y
# forward pass went well
assert ymg.data == ypt.data.item()
print(f"{ymg.data = }\n{ypt.data.item() = }\n")
# backward pass went well
print(f"{xmg.grad = }\n{xpt.grad.item() = }\n")
assert xmg.grad == xpt.grad.item()
def __init__(self, nin, nonlin=True):
self.w = [
Value(random.uniform(-1, 1), requires_grad=True)
for _ in range(nin)
]
self.b = Value(0, requires_grad=True)
self.nonlin = nonlin
def test_Tensor_sanity_check_scalar():
x = Tensor([[Value(-4.0)]])
two_1 = Tensor([[Value(2)]])
two_2 = Tensor([[Value(2)]])
z = two_1 * x + two_2 + x
q = z.relu() + z * x
h = (z * z).relu()
y = h + q + q * x
y.data[0][0].backward()
xmg, ymg = x.data[0][0], y.data[0][0]
x = torch.Tensor([-4.0]).double()
x.requires_grad = True
z = 2 * x + 2 + x
q = z.relu() + z * x
h = (z * z).relu()
y = h + q + q * x
y.backward()
xpt, ypt = x, y
# forward pass went well
assert ymg.data == ypt.data.item()
# backward pass went well
assert xmg.grad == xpt.grad.item()
def sigmoid(self):
out = Value(math.e**self / (math.e**self + 1), (self, ), f'sigmoid')
def _backward():
self.grad += math.e**self / ((math.e**self + 1) * (math.e**self + 1))
out._backward = _backward
return out
def log(self, **kwargs):
out = Value(math.log(self.data), (self, ), f'log')
def _backward():
self.grad += 1 / self.data * out.grad
out._backward = _backward
return out
def __pow__(self, other):
other = other if isinstance(other, Value) else Value(other)
out = Value(self.data**other.data, (self, other), f'pow')
def _backward():
self.grad += (other.data * self.data**(other.data - 1)) * out.grad
if other.data != 0:
other.grad += (self.data**other.data) * math.log(abs(
other.data)) * out.grad
out._backward = _backward
return out
def compare(func, *inputs):
"""Compare torch and micrograd implementations for func, applied to inputs.
`func` takes in any number of arguments, each of which is a Value or size-1 Tesnor and returns
any number of outputs.
:param func: a mathematical function using Value or Tensor primitive functions.
:param inputs: a list of scalars, which are arguments to func.
:returns:
:rtype:
"""
assert len(inputs) > 0
vs = [Value(x) for x in inputs]
ts = [torch.Tensor([x]) for x in inputs]
for t in ts:
t.requires_grad = True
# micrograd, pytorch output
ov = func(*vs)
ot = func(*ts)
assert ov.data == ot.data.item(), f'values: {ov} != {ot}'
ov.backward()
ot.backward()
for v, t in zip(vs, ts):
assert v.grad == t.grad.item(), f'gradients: {v} != {t}'
def test_sanity_check():
x = Value(-4.0)
z = 2 * x + 2 + x
z.backward()
assert z.data == -10.0
assert x.grad == 3.0
def __init__(self, in_features, out_features, bias: bool = True):
"""Initializing model"""
stddev = 1 / np.sqrt(in_features)
row = [
Value(item)
for item in np.random.uniform(-stddev, stddev, size=out_features)
]
matrix = Tensor([list(row) for _ in range(in_features)])
self.W = matrix
self.b = row if bias is True else 0
def test_higher_order():
x = Value(3)
y = Value(2)
f = 2 * x * x * y + y * x
x.grad = y.grad = 0
f.backward()
dfx = x.grad
dfy = y.grad
assert dfx.data == 4 * x.data * y.data + y.data
assert dfy.data == 2 * x.data * x.data + x.data
x.grad = y.grad = 0
dfx.backward()
dfxx = x.grad
dfxy = y.grad
x.grad = y.grad = 0
dfy.backward()
dfyx = x.grad
dfyy = y.grad
print(dfxx, dfyx, dfyy, dfxy)
assert dfyx.data == dfxy.data == (4 * x.data + 1)
assert dfyy == 0
assert dfxx.data == 4 * y.data
def test_sanity_check():
x = Value(-4.0)
z = 2 * x + 2 + x
q = z.relu() + z * x
h = (z * z).relu()
y = h + q + q * x
y.backward()
xmg, ymg = x, y
x = torch.Tensor([-4.0])
x.requires_grad = True
z = 2 * x + 2 + x
q = z.relu() + z * x
h = (z * z).relu()
y = h + q + q * x
y.backward()
xpt, ypt = x, y
# forward pass went well
assert ymg.data == ypt.data.item()
# backward pass went well
assert xmg.grad == xpt.grad.item()
def test_more_ops():
a = Value(-4.0)
b = Value(2.0)
a.requires_grad = True
b.requires_grad = True
c = a + b
d = a * b + b**3
c += c + 1
c += 1 + c + (-a)
d += d * 2 + (b + a).relu()
d += 3 * d + (b - a).relu()
e = c - d
f = e**2
g = f / 2.0
g += 10.0 / f
g.backward()
amg, bmg, gmg = a, b, g
a = torch.Tensor([-4.0]).double()
b = torch.Tensor([2.0]).double()
a.requires_grad = True
b.requires_grad = True
c = a + b
d = a * b + b**3
c = c + c + 1
c = c + 1 + c + (-a)
d = d + d * 2 + (b + a).relu()
d = d + 3 * d + (b - a).relu()
e = c - d
f = e**2
g = f / 2.0
g = g + 10.0 / f
g.backward()
apt, bpt, gpt = a, b, g
tol = 1e-6
# forward pass went well
assert abs(gmg.data - gpt.data.item()) < tol
# backward pass went well
assert abs(amg.grad - apt.grad.item()) < tol
assert abs(bmg.grad - bpt.grad.item()) < tol
def test_higher_order():
x = Value(3)
y = x**3
# This might seem redundant, but better we do this to ensure gradient is
# 0 before backward pass.
x.grad = 0
y.backward()
dy = x.grad
assert dy.data == (x.data**2) * 3
x.grad = 0
dy.backward()
d2y = x.grad
assert d2y.data == x.data * 6
x.grad = 0
d2y.backward()
d3y = x.grad
assert d3y.data == 6
x.grad = 0
d3y.backward()
d4y = x.grad
assert d4y.data == 0
def __repr__(self):
return f"Layer of [{', '.join(str(n) for n in self.neurons)}]"
class MLP(Module):
def __init__(self, nin, nouts):
sz = [nin] + nouts
self.layers = [
Layer(sz[i], sz[i + 1], nonlin=i != len(nouts) - 1)
for i in range(len(nouts))
]
def __call__(self, x):
for layer in self.layers:
x = layer(x)
return x
def parameters(self):
return [p for layer in self.layers for p in layer.parameters()]
def __repr__(self):
return f"MLP of [{', '.join(str(layer) for layer in self.layers)}]"
if __name__ == '__main__':
n = Neuron(2, nonlin=False)
x = [Value(1.0), Value(-2.0)]
y = n(x)
print(y)
def __init__(self, nin, nonlin=True):
self.w = [Value(random.uniform(-1, 1)) for _ in range(nin)]
self.b = Value(random.uniform(-1, 1))
self.nonlin = nonlin
# w = Value(3.0, name='w')
# x = Value(-4.0, name='x')
# y = Value(2.0, name='y')
# l = (w*x - y)**2
# l.name = 'MSEloss'
# print("l:", l)
# l.backward()
# draw_dot(l).render('test0')
# %%
# LOGloss
import math
from micrograd.engine import Value
from micrograd.trace_graph import draw_dot
w = Value(3.0, name='w')
x = Value(-4.0, name='x')
y = Value(1.0, name='y')
# Assume True y = 1
dot = Value(w.data * x.data, name='dotprod')
print("dot:", dot)
print("dot.sigmoid():", dot.sigmoid())
l = dot.cross_entropy(1)
# l.name = 'LOGloss'
print("loss:", l)
l.backward()
draw_dot(l).render('test0')
# %%
import numpy as np
def __rpow__(self, other):
if not isinstance(other, (int, float, Value)):
return NotImplemented
other = other if isinstance(other, Value) else Value(other)
return other**self
from micrograd.engine import Value
a = Value(-4.0)
b = Value(2.0)
c = a + b
d = a * b + b**3
c += c + 1
c += 1 + c + (-a)
d += d * 2 + (b + a).relu()
d += 3 * d + (b - a).relu()
e = c - d
f = e**2
g = f / 2.0
g += 10.0 / f
print(f'{g.data:.4f}') # prints 24.7041, the outcome of this forward pass
g.backward()
print(f'{a.grad:.4f}') # prints 138.8338, i.e. the numerical value of dg/da
print(f'{b.grad:.4f}') # prints 645.5773, i.e. the numerical value of dg/dbc
def __init__(self, nin, nonlin=True, seed=None):
if seed:
random.seed(seed)
self.w = [Value(random.uniform(-1, 1)) for _ in range(nin)]
self.b = Value(0)
self.nonlin = nonlin
