def test_broadcast_sub2(self):
t1 = Tensor([[1, 2, 3], [4, 5, 6]], requires_grad=True)
t2 = Tensor([[7, 8, 9]], requires_grad=True)
t3 = t1 - t2 # shape(2,3)
assert t3.data.tolist() == [[-6, -6, -6], [-3, -3, -3]]
t3.backward(Tensor([[1, 1, 1], [1, 1, 1]]))
assert t1.grad.data.tolist() == [[1, 1, 1], [1, 1, 1]]
assert t2.grad.data.tolist() == [[-2, -2, -2]]
def test_simple_mul(self):
t1 = Tensor([[1, 2], [3, 4], [5, 6]], requires_grad=True)
t2 = Tensor([[10], [20]], requires_grad=True)
t3 = t1 @ t2
assert t3.data.tolist() == [[50], [110], [170]]
grad = Tensor([[-1], [-2], [-3]])
t3.backward(grad)
assert np.all(t1.grad.data == grad.data @ t2.data.T)
assert np.all(t2.grad.data == t1.data.T @ grad.data)
def test_broadcast_add2(self):
t1 = Tensor([[1, 2, 3], [4, 5, 6]], requires_grad=True)
t2 = Tensor([[7, 8, 9]], requires_grad=True)
t3 = t1 + t2 # shape(2,3)
assert t3.data.tolist() == [[8, 10, 12], [11, 13, 15]]
t3.backward(Tensor([[1, 1, 1], [1, 1, 1]]))
assert t1.grad.data.tolist() == [[1, 1, 1], [1, 1, 1]]
assert t2.grad.data.tolist() == [[2, 2, 2]]
def test_simple_sub(self):
t1 = Tensor([1, 2, 3], requires_grad=True)
t2 = Tensor([4, 5, 6], requires_grad=True)
t3 = t1 - t2
assert t3.data.tolist() == [-3, -3, -3]
t3.backward(Tensor([-1, -2, -3]))
assert t1.grad.data.tolist() == [-1, -2, -3]
assert t2.grad.data.tolist() == [1, 2, 3]
t1 -= 0.1
assert t1.grad is None
assert t1.data.tolist() == [0.9, 1.9, 2.9]
def test_simple_add(self):
t1 = Tensor([1, 2, 3], requires_grad=True)
t2 = Tensor([4, 5, 6], requires_grad=True)
t3 = t1 + t2
assert t3.data.tolist() == [5, 7, 9]
t3.backward(Tensor([-1, -2, -3]))
assert t1.grad.data.tolist() == [-1, -2, -3]
assert t2.grad.data.tolist() == [-1, -2, -3]
t1 += 0.1
assert t1.grad is None
assert t1.data.tolist() == [1.1, 2.1, 3.1]
def __init__(self,
num_channels: int,
eps: float = 1e-05,
momentum=0.1,
affine: bool = True,
track_running_stats: bool = True) -> None:
# Call super constructor
super(BatchNorm1d, self).__init__()
# Save parameter
self.eps = eps
self.momentum = momentum
self.track_running_stats = track_running_stats
# Init learnable parameter if utilized
self.gamma = Parameter(data=np.ones(num_channels)) if affine else None
self.beta = Parameter(data=np.zeros(num_channels)) if affine else None
# Init running mean and std if needed
self.running_mean = Tensor(0.0) if self.track_running_stats else None
self.running_std = Tensor(1.0) if self.track_running_stats else None
def test_broadcast_sub(self):
"""
eg: t1.shape==(10,5) t2.shape=(5,) => t1 + t2 ,viewed as(1,5)
t2=[1,2,3,4,5] => viewed v2 as[[1,2,3,4,5]]
eg:
t1 as (10,5)
t2 as (1,5) ias [[1,2,3,4,,5]]
:return:
"""
t1 = Tensor([[1, 2, 3], [4, 5, 6]], requires_grad=True)
t2 = Tensor([7, 8, 9], requires_grad=True)
t3 = t1 - t2 # shape(2,3)
assert t3.data.tolist() == [[-6, -6, -6], [-3, -3, -3]]
t3.backward(Tensor([[1, 1, 1], [1, 1, 1]]))
assert t1.grad.data.tolist() == [[1, 1, 1], [1, 1, 1]]
assert t2.grad.data.tolist() == [-2, -2, -2]
"""
The idea here is that we'd like to use our library
to minimize a function
say x**2
"""
from autograd import Tensor
x = Tensor([10, -10, 10, -5, 6, 3, 1], requires_grad=True)
# we want to minimize the sum of squares
for i in range(100):
x.zero_grad()
sum_of_squares = (x * x).sum() # is a 0-tensor
sum_of_squares.backward()
delta_x = 0.1 * x.grad
x -= delta_x
print(i, sum_of_squares)
from autograd import Tensor x = Tensor([[1., 2., 3.]]) w = Tensor([[2.], [3.], [4.]], requires_grad=True) b = Tensor([.0], requires_grad=True) y_ = x.matmul(w).add(b) y = Tensor([60.]) loss = y_.sub(y).pow(Tensor(2.)).div(Tensor(2.)) loss.backward() print(loss.narray) print(w.grad) print(b.grad)
import numpy as np
from autograd import Tensor
t1 = Tensor(1.)
print(t1.narray.dtype)
xs = []
ys = []
for i in range(1, 10):
xs.append(Tensor([[i]]))
ys.append(Tensor([[i * 2]]))
w = Tensor([[.1]], requires_grad=True)
b = Tensor([.0], requires_grad=True)
for i in range(9):
for j in range(9):
x = xs[j]
y = ys[j]
print('x:')
print(x.narray)
print('y:')
print(y.narray)
y_ = x.matmul(w).add(b)
print('y_:')
print(y_.narray)
loss = y_.sub(y).pow(Tensor(2.))
print('loss:')
print(loss.narray)
loss.backward()
def test_simple_sum(self):
t1 = Tensor([1, 2, 3], requires_grad=True)
t2 = t1.sum()
t2.backward()
assert t1.grad.data.tolist() == [1, 1, 1]
def test_sum_with_grad(self):
t1 = Tensor([1, 2, 3], requires_grad=True)
t2 = t1.sum()
t2.backward(Tensor(3))
assert t1.grad.data.tolist() == [3, 3, 3]
self.b1 = Parameter(num_hidden)
self.w2 = Parameter(num_hidden, 4)
self.b2 = Parameter(4)
def predict(self, inputs: Tensor) -> Tensor:
# input is (batch_size, 10)
x1 = inputs @ self.w1 + self.b1 # (batch_size, num_hidden)
x2 = tanh(x1) # (batch_size, num_hidden)
x3 = x2 @ self.w2 + self.b2 # (batch_size, 4)
return x3
if __name__ == '__main__':
x_train = Tensor([binary_encode(x)
for x in range(101, 1024)]) # (922, 10) tensor
y_train = Tensor([fizz_buzz_encode(x)
for x in range(101, 1024)]) # (922, 4) tensor
optimizer = SGD(lr=1e-3)
batch_size = 32
model = FizzBuzzModel()
# Getting a different order of batches at each epoch
starts = np.arange(0, x_train.shape[0], batch_size)
for epoch in range(5000):
epoch_loss = 0.0
for start in starts:
end = start + batch_size
import numpy as np
from autograd import Tensor, Parameter, Module
from autograd.optim import SGD
x_data = Tensor(np.random.randn(100, 3))
coef = Tensor(np.array([-1, +3, -2], dtype=np.float)) # (3,)
# @矩阵乘法
# （100，3）*（3，1）
y_data = x_data @ coef + 5
class Module(Module):
def __init__(self) -> None:
self.w = Parameter(3)
self.b = Parameter()
def predict(self, in_puts: Tensor):
return in_puts @ self.w + self.b
# w = Tensor(np.random.randn(3), requires_grad=True)
# b = Tensor(np.random.randn(), requires_grad=True)
# w = Parameter(3) # tensor(3,),requires_grad =True,random values
# b = Parameter()
optimizer = SGD(lr=0.001)
# learning_rate = 0.001
batch_size = 32
module = Module()
for epoch in range(100):
epoch_loss = 0.0
def sigmoid2(x):
return Tensor(1).add(x.ng().exp())
from autograd import Tensor
import numpy as np
def sigmoid(x):
return x.exp()
def sigmoid2(x):
return Tensor(1).add(x.ng().exp())
x = Tensor([[1.1]], requires_grad=True)
# y = x.pow(Tensor(2))
# print(y.narray)
# y.backward()
# print(x.grad.narray)
# loss = y.sub(Tensor(10))
# print(loss.narray)
# loss.backward(loss)
# print(x.grad.narray)
# y = Tensor(1).div(Tensor(1).add(x.ng().exp()))
# print(y.narray)
def fizz_buzz_encode(x: int) -> List[int]:
if x % 15 == 0:
return [0, 0, 0, 1]
elif x % 5 == 0:
return [0, 0, 1, 0]
elif x % 3 == 0:
return [0, 1, 0, 0]
else:
return [1, 0, 0, 0]
# @矩阵乘法
# （100，3）*（3，1）
x_train = Tensor([binary_encode(x) for x in range(101, 1024)])
y_train = Tensor([fizz_buzz_encode(x) for x in range(101, 1024)])
class FizzBuzzModule(Module):
def __init__(self, num_hidden: int = 50) -> None:
self.w1 = Parameter(10, num_hidden)
self.b1 = Parameter(num_hidden)
self.w2 = Parameter(num_hidden, 4)
self.b2 = Parameter(4)
def predict(self, in_puts: Tensor):
# inputs (batch_size,10)
x1 = inputs @ self.w1 + self.b1 # (batch_size,num_hidden)
x2 = tanh(x1)
def binary_encode(x: int) -> List[int]:
return [x >> i & 1 for i in range(10)]
def fizz_buzz_encode(x: int) -> List[int]:
if x % 15 == 0:
return [0, 0, 0, 1]
elif x % 5 == 0:
return [0, 0, 1, 0]
elif x % 3 == 0:
return [0, 1, 0, 0]
else:
return [1, 0, 0, 0]
x_train = Tensor([binary_encode(x) for x in range(101, 1024)])
y_train = Tensor([fizz_buzz_encode(x) for x in range(101, 1024)])
class FizzBuzzModel(Module):
def __init__(self, num_hidden: int = 50) -> None:
self.fc1 = Linear(10, num_hidden)
self.fc2 = Linear(num_hidden, 4)
self.dropout = Dropout(0.1)
def forward(self, inputs: Tensor) -> Tensor:
return self.predict(inputs)
def predict(self, inputs: Tensor) -> Tensor:
# inputs will be (batch_size, 10)
x1 = self.fc1(inputs) # (batch_size, num_hidden)
from autograd import Tensor, Parameter, Module
from autograd.optim import SGD
class LinearModel(Module):
def __init__(self) -> None:
self.w = Parameter(3)
self.b = Parameter()
def predict(self, inputs: Tensor) -> Tensor:
"""Learned function: y = Ax + b"""
return inputs @ self.w + self.b
if __name__ == '__main__':
x_data = Tensor(np.random.randn(100, 3))
coef = Tensor(np.array([-1, 3, -2]))
# The function to be learned y = Ax + b + eps
# x_data is a tensor, so is y_data
# y_data = x_data @ coef + 5 + np.random.randint(-2, 2, size=(100,))
# With a perfect linear regression, we can get 0 error
y_data = x_data @ coef + 5
optimizer = SGD(lr=1e-3)
batch_size = 32
model = LinearModel()
for epoch in range(100):
epoch_loss = 0.0
from autograd import Tensor x = Tensor([[1., 2., 3.]]) w = Tensor([[2.], [3.], [4.]], requires_grad=True) b = Tensor([.0], requires_grad=True) y_ = x.matmul(w).add(b) y = Tensor([60.]) loss = y_.sub(y).pow(Tensor(2.)).div(Tensor(2.)) print(loss.narray) loss.backward() print(w.grad) print(b.grad) w.narray = w.narray + (0.001 * w.grad.T) b.narray = b.narray + (0.001 * b.grad.T) print(w.narray) print(b.narray)
