-
Notifications
You must be signed in to change notification settings - Fork 0
/
tensor.py
91 lines (72 loc) · 2.8 KB
/
tensor.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
class Tensor:
def __init__(self,
value,
operand=(),
operation=None,
leaf=True,
name=None):
self.value = value
self.grad = 0
self.operand = operand
self.operation = operation
self.grad_fxn = lambda: None
self.gradients_calculated = False
self.leaf = leaf
self.name = name
def __repr__(self):
return f"Tensor(value={self.value})"
def __add__(self, tensor):
out = Tensor(self.value + tensor.value, operand=(self, tensor), operation='+', leaf=False)
def add_grad_fxn():
self.grad += 1 * out.grad
tensor.grad += 1 * out.grad
out.grad_fxn = add_grad_fxn
return out
def __sub__(self, tensor):
out = Tensor(self.value - tensor.value, operand=(self, tensor), operation='-', leaf=False)
def sub_grad_fxn():
self.grad += 1 * out.grad
tensor.grad += -1 * out.grad
out.grad_fxn = sub_grad_fxn
return out
def __mul__(self, tensor):
out = Tensor(self.value * tensor.value, operand=(self, tensor), operation='*', leaf=False)
def mul_grad_fxn():
self.grad += tensor.value * out.grad
tensor.grad += self.value * out.grad
out.grad_fxn = mul_grad_fxn
return out
def __truediv__(self, tensor):
out = Tensor(self.value / tensor.value, operand=(self, tensor), operation='/', leaf=False)
def div_grad_fxn():
self.grad += (1/tensor.value) * out.grad
tensor.grad += -self.value /(tensor.value ** 2) * out.grad
out.grad_fxn = div_grad_fxn
return out
def __pow__(self, power):
out = Tensor(self.value**power, operand=(self, ), operation=f'**{power}', leaf=False)
def pow_grad_fxn():
self.grad += power * (self.value ** (power - 1))
out.grad_fxn = pow_grad_fxn
return out
def backward(self):
sorted, visited = [], set()
def topological_sort(node):
if node not in visited:
visited.add(node)
for v in node.operand:
topological_sort(v)
sorted.append(node)
topological_sort(self)
for node in reversed(sorted):
node.gradients_calculated = True
node.grad_fxn()
def reset_grads(self, ):
visited = set()
def dfs(node):
if node not in visited:
visited.add(node)
node.grad = 0.0
for v in node.operand:
dfs(v)
dfs(self)