PyTorch学习笔记之计算图

1. **args, **kwargs的区别

    def build_vocab(self, *args, **kwargs):
        counter = Counter()
        sources = []
        for arg in args:
            if isinstance(arg, Dataset):
                sources += [getattr(arg, name) for name, field in
                            arg.fields.items() if field is self]
            else:
                sources.append(arg)
        for data in sources:
            for x in data:
                if not self.sequential:
                    x = [x]
                counter.update(x)
        specials = list(OrderedDict.fromkeys(
            tok for tok in [self.pad_token, self.init_token, self.eos_token]
            if tok is not None))
        self.vocab = Vocab(counter, specials=specials, **kwargs)

2. np.sum

import numpy as np
np.random.seed(0)

N, D = 3, 4
x = np.random.randn(N, D)
y = np.random.randn(N, D)
z = np.random.randn(N, D)

a = x * y
b = a + z
print(b)
c = np.sum(b)
print(c)                # 6.7170085378

# search the function of np.sum
total = 0
for i in range(len(b)):
    for j in range(4):
        total += b[i][j]
print(total)            # 6.7170085378

3. use numpy to solve grad

import numpy as np
N, D = 3, 4
x = np.random.randn(N, D)
y = np.random.randn(N, D)
z = np.random.randn(N, D)

a = x * y
b = a + z
# print(b)
c = np.sum(b)
# print(c)                # 6.7170085378

grad_c = 1.0
grad_b = grad_c * np.ones((N, D))
grad_a = grad_b.copy()
grad_z = grad_b.copy()
grad_x = grad_a * y
grad_y = grad_a * x

print(grad_x)
print(grad_y)
print(grad_z)
'''
[[ 0.04998285  0.32809396 -0.49822878  1.36419309]
 [-0.52303972 -0.5881509  -0.37058995 -1.42112189]
 [-0.58705758 -0.26012336  1.31326911 -0.20088737]]
[[ 0.14893265 -0.45509058  0.21410015  0.27659   ]
 [ 0.29617438  0.98971103  2.07310583 -0.0195055 ]
 [-1.49222601 -0.64073344 -0.18269488  0.26193553]]
[[ 1.  1.  1.  1.]
 [ 1.  1.  1.  1.]
 [ 1.  1.  1.  1.]]
'''

PyTorch自动计算梯度

import torch
from torch.autograd import Variable

N, D = 3, 4
# define variables to start building a computational graph
x = Variable(torch.randn(N, D), requires_grad=True)
y = Variable(torch.randn(N, D), requires_grad=True)
z = Variable(torch.randn(N, D), requires_grad=True)

# forward pass looks just like numpy
a = x * y
b = a + z
c = torch.sum(b)

# calling c,backward() computes all gradients
c.backward()
print(x.grad.data)
print(y.grad.data)
print(z.grad.data)
'''
-0.9775 -0.0913  0.3710  1.5789
 0.0896 -0.6563  0.8976 -0.3508
-0.9378  0.7028  1.4533  0.9255
[torch.FloatTensor of size 3x4]


 0.6365  0.2388 -0.4755 -0.9860
-0.2403 -0.0468 -0.0470 -1.0132
-0.5019  0.5005 -1.9270  1.0030
[torch.FloatTensor of size 3x4]


 1  1  1  1
 1  1  1  1
 1  1  1  1
[torch.FloatTensor of size 3x4]
'''

x is a variable, requires_grad=True.

x.data is a tensor.

x.grad is a variable of gradients(same shape as x.data).

x.grad.data is a tensor of gradients.

4. 随机数

（1）random.seed(int)

• 给随机数对象一个种子值，用于产生随机序列。
• 对于同一个种子值的输入，之后产生的随机数序列也一样。
• 通常是把时间秒数等变化值作为种子值，达到每次运行产生的随机系列都不一样
• seed() 省略参数，意味着使用当前系统时间生成随机数

（2）random.random()

　　生成随机浮点数

import numpy as np
np.random.seed(0)

print(np.random.random())       # 0.5488135039273248 不随时间改变
print(np.random.random())       # 0.7151893663724195 不随时间改变

np.random.seed(0)
print(np.random.random())       # 0.5488135039273248 不随时间改变

np.random.seed()
print(np.random.random())       # 0.9623797942471012 随时间改变
np.random.seed()
print(np.random.random())       # 0.12734792669918393 随时间改变

（3）random.shuffle

• 对list列表随机打乱顺序，也就是洗牌
• shuffle只作用于list，对Str会报错比如‘abcdfed’,而['1','2','3','5','6','7']可以

import numpy as np

item = [1,2,3,4,5,6,7]
print(item)                     # [1, 2, 3, 4, 5, 6, 7]
np.random.shuffle(item)
print(item)                     # [7, 1, 2, 5, 4, 6, 3]

item2 = ['1','2','3']
np.random.shuffle(item2)
print(item2)                    # ['1', '3', '2']

参考博客：python随机数用法