- 对于：
torch.arange(0,10).reshape(2,-1)==torch.arange(0,10).reshape(-1,5)
没有区别。 - 读字典：
dict.values()
dict.keys()
dict['key1']
list(dict) # keys
- list的拆分:
a,b = [['a','b','c'],[1,2,3]]
print(a,'\n',b)
"""
output:
['a', 'b', 'c']
[1, 2, 3]
"""
- random.sample()
# eg1
a = ['a','b','c','d','e','f','g']
random.sample(a,4)
"""
output:
['a', 'd', 'f', 'c']
"""
#eg2
b = {'a':1,'b':2,'c':3,'d':4,'e':5}
random.sample(b,2)
"""
output:
TypeError: Population must be a sequence or set. For dicts, use list(d).
故改为：
"""
{each_key:b[each_key] for each_key in random.sample(list(b),2)}
"""
{'b': 2, 'e': 5}
"""
下采样代码：
# 创建数据字典
c = {'a':[i for i in range(10)],'b':[i for i in range(100,110,1)]}
# 展平字典的values
value_flatten = [i for each in list(c.values()) for i in each]
# sum(c.values(),[])
# 采样
sampled_value = random.sample(value_flatten,10)
#寻位还原字典形式
zeros = {}
for each_value in sampled_value:
for each_keys in c:
if each_value in c[each_keys]:
if each_keys not in list(zeros):
zeros[each_keys] = []
zeros[each_keys].append(each_value)
"""
各个对象采样加起来一共10个。
采样前：
{'a': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
'b': [100, 101, 102, 103, 104, 105, 106, 107, 108, 109]}
采样后：
{'a': [6, 5, 3, 0, 4], 'b': [108, 100, 104, 105, 109]}
"""
- class_weight
# 创建数据
zeros = {str(key):random.sample(range(10),random.randint(5,8)) for key in range(300,310)}
ones = {str(key):random.sample(range(10),random.randint(5,8)) for key in range(400,415)}
"""
zeros:
{'300': [4, 1, 8, 5, 0],
'301': [9, 6, 2, 1, 3, 5],
...
'309': [6, 3, 7, 4, 8, 0, 5]},
ones:
{'400': [6, 2, 1, 5, 9, 7],
'401': [0, 3, 6, 7, 8, 9, 4],
'402': [1, 4, 5, 8, 3],
...
'414': [9, 0, 6, 2, 3, 4, 5, 7]}
"""
(1) class weight
instance_of_zero = sum(len(i) for i in zeros.values()) # 64
instance_of_one = sum(len(i) for i in ones.values()) # 88
min_class = min(len(zeros), len(ones)) # 10
max_class = max(len(zeros), len(ones)) # 15
class_weight = min_class / max_class # 0.667
index_to_alter = [len(zeros), len(ones)].index(max_class) # 1。读出哪一类的subject更多
weights = [1, 1]
weights[index_to_alter] = class_weight # [1,0.667]
# 以字典存储每个sample的weight
weights_dict = {}
for i in zeros:
weights_dict[i] = weights[0]
for i in ones:
weights_dict[i] = weights[1]
"""
weights_dict:
{'300': 1,
'301': 1,
'302': 1,
....
'400': 0.6666666666666666,
'401': 0.6666666666666666,
'402': 0.6666666666666666,
...}
"""
(2) micro weight
WEIGHT_TYPE = 'micro'
min_zeros = min(len(i) for i in zeros.values()) # 5
min_ones = min(len(i) for i in ones.values()) # 5
micro_weights = {}
for i in zeros:
tmp = min_zeros / len(zeros[i])
micro_weights[i] = tmp
for i in ones:
tmp = min_ones / len(ones[i])
micro_weights[i] = tmp
weights_dict = micro_weights
(3) combine(1)&(2)
# 以上两种结合
WEIGHT_TYPE = 'both'
min_zeros = min(len(i) for i in zeros.values()) # 5
min_ones = min(len(i) for i in ones.values()) # 5
micro_weights = {}
for i in zeros:
tmp = min_zeros / len(zeros[i])
if WEIGHT_TYPE == 'both':
tmp = tmp * weights[0]
micro_weights[i] = tmp
for i in ones:
tmp = min_ones / len(ones[i])
if WEIGHT_TYPE == 'both':
tmp = tmp * weights[1]
micro_weights[i] = tmp
weights_dict = micro_weights
(4) class weight (segment_based)
WEIGHT_TYPE = 'instance'
if WEIGHT_TYPE == 'instance':
min_tmp = min(instance_of_zero, instance_of_one)
max_tmp = max(instance_of_zero, instance_of_one)
tmp_weight = min_tmp / max_tmp
index_to_alter = [instance_of_zero, instance_of_one].index(max_tmp)
weights = [1, 1]
weights[index_to_alter] = tmp_weight
instance_weights = {}
for i in zeros:
instance_weights[i] = weights[0]
for i in ones:
instance_weights[i] = weights[1]
weights_dict = instance_weights
- indx
import numpy as np
a = np.array([i for i in range(10)])
idx = [5,6,7]
a[idx]
# output: array([5, 6, 7])
- 训练的数据格式
- float->torch.Tensor(x)
- int->torch.LongTensor(x)
def create_tensor_data(x, cuda):
"""
Converts the data from numpy arrays to torch tensors
Inputs
x: The input data
cuda: Bool - Set to true if using the GPU
Output
x: Data converted to a tensor
"""
if 'float' in str(x.dtype):
x = torch.Tensor(x)
elif 'int' in str(x.dtype):
x = torch.LongTensor(x)
else:
raise Exception("Error!")
if cuda:
x = x.cuda()
return x
