import numpy as np
class Tensor(object):
def __init__(self, data,
autograd=False,
creators=None,
creation_op=None,
id=None):
self.data = np.array(data)
self.autograd = autograd
self.grad = None
if (id is None):
self.id = np.random.randint(0, 100000)
else:
self.id = id
self.creators = creators
self.creation_op = creation_op
self.children = {}
if (creators is not None):
for c in creators:
if (self.id not in c.children):
c.children[self.id] = 1
else:
c.children[self.id] += 1
def all_children_grads_accounted_for(self):
for id, cnt in self.children.items():
if (cnt != 0):
return False
return True
def backward(self, grad=None, grad_origin=None):
if (self.autograd):
if (grad is None):
grad = Tensor(np.ones_like(self.data))
if (grad_origin is not None):
if (self.children[grad_origin.id] == 0):
raise Exception("cannot backprop more than once")
else:
self.children[grad_origin.id] -= 1
if (self.grad is None):
self.grad = grad
else:
self.grad += grad
# grads must not have grads of their own
assert grad.autograd == False
# only continue backpropping if there's something to
# backprop into and if all gradients (from children)
# are accounted for override waiting for children if
# "backprop" was called on this variable directly
if (self.creators is not None and
(self.all_children_grads_accounted_for() or
grad_origin is None)):
if (self.creation_op == "add"):
self.creators[0].backward(self.grad, self)
self.creators[1].backward(self.grad, self)
if (self.creation_op == "sub"):
self.creators[0].backward(Tensor(self.grad.data), self)
self.creators[1].backward(Tensor(self.grad.__neg__().data), self)
if (self.creation_op == "mul"):
new = self.grad * self.creators[1]
self.creators[0].backward(new, self)
new = self.grad * self.creators[0]
self.creators[1].backward(new, self)
if (self.creation_op == "mm"):
c0 = self.creators[0]
c1 = self.creators[1]
new = self.grad.mm(c1.transpose())
c0.backward(new)
new = self.grad.transpose().mm(c0).transpose()
c1.backward(new)
if (self.creation_op == "transpose"):
self.creators[0].backward(self.grad.transpose())
if ("sum" in self.creation_op):
dim = int(self.creation_op.split("_")[1])
self.creators[0].backward(self.grad.expand(dim,
self.creators[0].data.shape[dim]))
if ("expand" in self.creation_op):
dim = int(self.creation_op.split("_")[1])
self.creators[0].backward(self.grad.sum(dim))
if (self.creation_op == "neg"):
self.creators[0].backward(self.grad.__neg__())
if (self.creation_op == "sigmoid"):
ones = Tensor(np.ones_like(self.grad.data))
self.creators[0].backward(self.grad * (self * (ones - self)))
if (self.creation_op == "tanh"):
ones = Tensor(np.ones_like(self.grad.data))
self.creators[0].backward(self.grad * (ones - (self * self)))
if (self.creation_op == "index_select"):
new_grad = np.zeros_like(self.creators[0].data)
indices_ = self.index_select_indices.data.flatten()
grad_ = grad.data.reshape(len(indices_), -1)
for i in range(len(indices_)):
new_grad[indices_[i]] += grad_[i]
self.creators[0].backward(Tensor(new_grad))
def __add__(self, other):
if (self.autograd and other.autograd):
return Tensor(self.data + other.data,
autograd=True,
creators=[self, other],
creation_op="add")
return Tensor(self.data + other.data)
def __neg__(self):
if (self.autograd):
return Tensor(self.data * -1,
autograd=True,
creators=[self],
creation_op="neg")
return Tensor(self.data * -1)
def __sub__(self, other):
if (self.autograd and other.autograd):
return Tensor(self.data - other.data,
autograd=True,
creators=[self, other],
creation_op="sub")
return Tensor(self.data - other.data)
def __mul__(self, other):
if (self.autograd and other.autograd):
return Tensor(self.data * other.data,
autograd=True,
creators=[self, other],
creation_op="mul")
return Tensor(self.data * other.data)
def sum(self, dim):
if (self.autograd):
return Tensor(self.data.sum(dim),
autograd=True,
creators=[self],
creation_op="sum_" + str(dim))
return Tensor(self.data.sum(dim))
def expand(self, dim, copies):
trans_cmd = list(range(0, len(self.data.shape)))
trans_cmd.insert(dim, len(self.data.shape))
new_data = self.data.repeat(copies).reshape(list(self.data.shape) + [copies]).transpose(trans_cmd)
if (self.autograd):
return Tensor(new_data,
autograd=True,
creators=[self],
creation_op="expand_" + str(dim))
return Tensor(new_data)
def transpose(self):
if (self.autograd):
return Tensor(self.data.transpose(),
autograd=True,
creators=[self],
creation_op="transpose")
return Tensor(self.data.transpose())
def mm(self, x):
if (self.autograd):
return Tensor(self.data.dot(x.data),
autograd=True,
creators=[self, x],
creation_op="mm")
return Tensor(self.data.dot(x.data))
def sigmoid(self):
if (self.autograd):
return Tensor(1 / (1 + np.exp(-self.data)),
autograd=True,
creators=[self],
creation_op="sigmoid")
return Tensor(1 / (1 + np.exp(-self.data)))
def tanh(self):
if (self.autograd):
return Tensor(np.tanh(self.data),
autograd=True,
creators=[self],
creation_op="tanh")
return Tensor(np.tanh(self.data))
def index_select(self, indices):
if (self.autograd):
new = Tensor(self.data[indices.data],
autograd=True,
creators=[self],
creation_op="index_select")
new.index_select_indices = indices
return new
return Tensor(self.data[indices.data])
def __repr__(self):
return str(self.data.__repr__())
def __str__(self):
return str(self.data.__str__())
class Layer(object):
def __init__(self):
self.parameters = list()
def get_parameters(self):
return self.parameters
class Tanh(Layer):
def __init__(self):
super().__init__()
def forward(self, input):
return input.tanh()
class Embedding(Layer):
def __init__(self, vocab_size, dim):
super().__init__()
self.vocab_size = vocab_size
self.dim = dim
# this random initialiation style is just a convention from word2vec
self.weight = (np.random.rand(vocab_size, dim) - 0.5) / dim
class Sigmoid(Layer):
def __init__(self):
super().__init__()
def forward(self, input):
return input.sigmoid()
x = Tensor(np.eye(5), autograd=True)
x.index_select(Tensor([[1,2,3],[2,3,4]])).backward()
print(x.grad)
''' 2 he 出现了2次，所以对应的行时2 2 2 2 2
[[0. 0. 0. 0. 0.]
[1. 1. 1. 1. 1.]
[2. 2. 2. 2. 2.]
[2. 2. 2. 2. 2.]
[1. 1. 1. 1. 1.]]
'''
