pytorch 基本資料初始化

2021-09-25 21:48:05 字數 4211 閱讀 3206

import torch


#張量初始化


# print ('一般性初始化：', torch.tensor([1,2,3]))


# print('隨機初始化：', torch.arange(0,4,step = 1))


# print('隨機初始化：',torch.range(0,3,step = 1))


# the step is the skipwindow


# print('隨機初始化：',torch.linspace(0,3,steps = 4))


# print('隨機初始化：',torch.logspace(0,3,steps = 4))


# the steps is numbers


# weight = torch.full((3,4), 0.6)


# # the percentage args


# print(torch.bernoulli(weight))


# # generate 0 or 1


# # weights = torch.tensor([[1,100],[100,1],[1,1]],dtype = torch.float32)


# # print (torch.mutinormal(weights,1))


# # the weight and the select times


# print('生成隨機序列：',torch.randperm(10))


# # 0 - 9 without sort tensor


# print('生成均勻分布：',torch.randint(low = 0, high = 4, size = (3, 4)))


# print('生成均勻分布：',torch.randint_like(torch.ones(3, 4), low = 0, high = 4))


# # include low but without high


# print('生成標準均勻分布：', torch.rand(3, 4))


# print('生成標準均勻分布：', torch.rand_like(torch.ones(3, 4)))


# # 0 - 1


# print('生成標準正太分布：',torch.randn(3, 4))


# print('生成標準正太分布：',torch.randn_like(torch.ones(3, 4)))


# mean = torch.tensor([1., 2.])


# std = torch.tensor([4., 2.])


# print('生成一般正態分佈：',torch.normal(mean, std))


# # notice that the mean and the std should be float


# test_tensor = torch.tensor([1,2,3,4]).reshape(1,2,2,1)


# print(test_tensor.dim())


# print(test_tensor.size())


# print(test_tensor.numel())


# 張量選擇


# tensor_to_be_selected = torch.arange(12).reshape(3, 4)


# index_selected_by = torch.tensor([0, 2])


# print ('index select:', tensor_to_be_selected.index_select(1, index_selected_by))


# # index_select dim is remain


# mask_selected_by = torch.tensor([[0, 0, 0, 1],[1, 0, 1, 0],[0, 1, 0, 0]], dtype = torch.uint8)


# print ('mask select:', tensor_to_be_selected.masked_select(mask_selected_by))


# # masked_select dim is reduced to 1


# take_selected_by = torch.tensor([2,3,4])


# print ('take select:', tensor_to_be_selected.take(take_selected_by))


# # take_select dim is reduced to 1


# print(torch.arange(4).reshape(2, 2).repeat(4, 2))


# # the result dims is (2*4,2*2)


# tensor_one = torch.tensor([[1, 2]])


# tensor_two = torch.tensor([[3, 4]])


# print(torch.cat((tensor_one, tensor_two), 0))


# # 1x2 cat 1x2 ——>2x2


# print(torch.cat((tensor_one, tensor_two), 1))


# # 1x2 cat 1x2 ——>1x4


# tensor_to_be_selected = torch.arange(12, dtype = torch.float32).reshape(3, 4)


# print(tensor_to_be_selected.reciprocal())


# print(tensor_to_be_selected.sqrt())


# print(tensor_to_be_selected.rsqrt())


# positive_tensor = torch.tensor([0.5],dtype = torch.float32)


# nagtive_tensor = torch.tensor([-0.5],dtype = torch.float32)


# print('符號函式：',torch.sign(positive_tensor), torch.sign(nagtive_tensor))


# print('絕對值函式：', torch.abs(positive_tensor), torch.abs(nagtive_tensor))


# print('向下取整：',torch.floor(positive_tensor), torch.floor(nagtive_tensor))


# print('向上取整：',torch.ceil(positive_tensor), torch.ceil(nagtive_tensor))


# print('四捨五入：',torch.round(positive_tensor), torch.round(nagtive_tensor))


# print('截斷取整：',torch.trunc(positive_tensor),torch.trunc(nagtive_tensor))


# print('取小數：',torch.frac(positive_tensor), torch.frac(nagtive_tensor))


# tensor_one = torch.tensor([1, 2, 3])


# tensor_two = torch.tensor([2, 1, 3])


# print('一維張量點積：', torch.dot(tensor_one, tensor_two))


# tensor_one = torch.arange(4).reshape(2, 2)


# tensor_two = torch.tensor([1, 2])


# print('二維點積一維：', torch.mv(tensor_one, tensor_two))


# tensor_one = torch.arange(4).reshape(2, 2)


# tensor_two = torch.tensor([[1, 2], [3, 4]])


# tensor_one = torch.arange(1, 5, dtype = torch.float32)


# print('方差：',tensor_one.var())


# print('最大值：',tensor_one.max())


# print('中位數：',tensor_one.median())


# print('乘積：',tensor_one.prod())


# print('第三大的數字是：',tensor_one.kthvalue(3))

pytorch 關於權重（初始化 固定）

emb那種可以直接from pretrained 利用requires grad false，這樣只是不更新，但是還是會計算並占用視訊記憶體 class net nn.module def init self super net,self init self.conv1 nn.conv2d 1,6,...

資料初始化

資料初始化的面試題 a 乙個類的初始化過程 b 子父類的構造執行過程 c 分層初始化 看程式寫結果 a 成員變數的問題 int x 10 成員變數是基本型別 student s new student 成員變數是引用型別 b 乙個類的初始化過程 先父後子，先靜後非靜 成員變數的初始化 預設初始化 顯...

初始化 指定初始化

id alloc 物件的誕生過程，主要是從作業系統獲得一塊足夠大的記憶體，以存放該類的全部例項變數，並將其指定為存放記憶體物件的實力變數的位置。alloc方法同時將這塊記憶體全部設定為0。結果是 bool變數初始化為no,所有的int型別變數為0,float變數為0.0，所有的指標為nil.obje...