search

lartpang / Machine-Deep-Learning

:wave: ML/DL学习笔记（基础+论文）
262 stars 58 forks source link

技巧-Pytorch #39

Open ghost opened 5 years ago

ghost commented 5 years ago

损失函数整理

https://blog.csdn.net/zhangxb35/article/details/72464152?utm_source=itdadao&utm_medium=referral

ghost commented 5 years ago

保存/载入模型

载入

    model_path = os.path.join(
        shufflenetv2_path, 'shufflenetv2_x1_69.402_88.374.pth.tar')
    model = ShuffleNetV2()

    if pretrained:
        print(f"=> loading model '{model_path}'")
        pretrained_dict = torch.load(model_path)

        model_dict = model.state_dict()

        # 1. filter out unnecessary keys
        pretrained_dict = {k: v for k, v in pretrained_dict.items()
                           if k in model_dict}
        # 2. overwrite entries in the existing state dict
        model_dict.update(pretrained_dict)
        # 3. load the new state dict
        model.load_state_dict(model_dict)

保存后读入

# 存储一个epoch后的模型(权重和偏置项), 以便后期使用
filename = ('%s/feature-current.pth' % check_root_feature)
torch.save(model.state_dict(), filename)

# 存储优化器状态
filename_opti = ('%s/opti-current.pth' % check_root_opti)
torch.save(optimizer_feature.state_dict(), filename_opti)

# 载入上一次的训练结果(权重和偏置项), 进一步的训练
model.load_state_dict(
    torch.load(check_root_feature + '/feature-current.pth')
)
# 载入优化器状态
optimizer_feature.load_state_dict(
    torch.load(check_root_opti + '/opti-current.pth')
)
fileroot = ('%s/feature-current.pth' % check_root_feature)
# 基于torch.save(model.state_dict(), filename)存储方法的对应的恢复方法
model.load_state_dict(torch.load(fileroot))
ghost commented 5 years ago

展示数据

# Have a look at data
inputs, classes = next(iter(dataloaders['train']))
out = torchvision.utils.make_grid(inputs)
imshow(out, title=[class_names[x] for x in classes])
ghost commented 5 years ago

反卷积计算

image

ghost commented 5 years ago

显存优化

ghost commented 5 years ago

pytorch加速技巧

ghost commented 5 years ago

TensorBoardX

# demo.py

import torch
import torchvision.utils as vutils
import numpy as np
import torchvision.models as models
from torchvision import datasets
from tensorboardX import SummaryWriter

resnet18 = models.resnet18(False)
writer = SummaryWriter()
sample_rate = 44100
freqs = [262, 294, 330, 349, 392, 440, 440, 440, 440, 440, 440]

for n_iter in range(100):

    dummy_s1 = torch.rand(1)
    dummy_s2 = torch.rand(1)
    # data grouping by `slash`
    writer.add_scalar('data/scalar1', dummy_s1[0], n_iter)
    writer.add_scalar('data/scalar2', dummy_s2[0], n_iter)

    writer.add_scalars('data/scalar_group', {'xsinx': n_iter * np.sin(n_iter),
                                             'xcosx': n_iter * np.cos(n_iter),
                                             'arctanx': np.arctan(n_iter)}, n_iter)

    dummy_img = torch.rand(32, 3, 64, 64)  # output from network
    if n_iter % 10 == 0:
        x = vutils.make_grid(dummy_img, normalize=True, scale_each=True)
        writer.add_image('Image', x, n_iter)

        dummy_audio = torch.zeros(sample_rate * 2)
        for i in range(x.size(0)):
            # amplitude of sound should in [-1, 1]
            dummy_audio[i] = np.cos(freqs[n_iter // 10] * np.pi * float(i) / float(sample_rate))
        writer.add_audio('myAudio', dummy_audio, n_iter, sample_rate=sample_rate)

        writer.add_text('Text', 'text logged at step:' + str(n_iter), n_iter)

        for name, param in resnet18.named_parameters():
            writer.add_histogram(name, param.clone().cpu().data.numpy(), n_iter)

        # needs tensorboard 0.4RC or later
        writer.add_pr_curve('xoxo', np.random.randint(2, size=100), np.random.rand(100), n_iter)

dataset = datasets.MNIST('mnist', train=False, download=True)
images = dataset.test_data[:100].float()
label = dataset.test_labels[:100]

features = images.view(100, 784)
writer.add_embedding(features, metadata=label, label_img=images.unsqueeze(1))

# export scalar data to JSON for external processing
writer.export_scalars_to_json("./all_scalars.json")
writer.close()
ghost commented 5 years ago

梯度剪裁

optimizer.zero_grad()        
loss, hidden = model(data, hidden, targets)
loss.backward()

# https://pytorch.org/docs/master/nn.html#clip-grad-norm
# torch.nn.utils.clip_grad_norm_(parameters, max_norm, norm_type=2)
# torch.nn.utils.clip_grad_value_(parameters, clip_value)
torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip)
optimizer.step()

既然在BP过程中会产生梯度消失/爆炸（就是偏导无限接近0，导致长时记忆无法更新），那么最简单粗暴的方法，设定阈值，当梯度小于/大于阈值时，对梯度进行限制.

ghost commented 5 years ago

参数初始化

关于初始化函数的选择

官方文档里还有其他的: https://blog.csdn.net/dss_dssssd/article/details/83959474

  1. 均匀分布 torch.nn.init.uniform_(tensor, a=0, b=1), 服从U(a,b)
  2. 正太分布 torch.nn.init.normal_(tensor, mean=0, std=1) 服从N(mean,std)
  3. 初始化为常数 torch.nn.init.constant_(tensor, val) 初始化整个矩阵为常数val
  4. Xavier 基本思想是通过网络层时，输入和输出的方差相同，包括前向传播和后向传播. 对于Xavier初始化方式，pytorch提供了uniform和normal两种:
    • torch.nn.init.xavier_uniform_(tensor, gain=1) 均匀分布U(−a,a)
    • torch.nn.init.xavier_normal_(tensor, gain=1) 正态分布N(0,std)

      Xavier在tanh中表现的很好，但在Relu激活函数中表现的很差，所何凯明提出了针对于Relu的初始化方法。

  5. 在ReLU网络中，假定每一层有一半的神经元被激活，另一半为0，所以，要保持方差不变，只需要在Xavier的基础上再除以2. 也就是说在方差推到过程中，式子左侧除以2. pytorch也提供了两个版本：
    • torch.nn.init.kaiming_uniform_(tensor, a=0, mode=‘fan_in’, nonlinearity=‘leaky_relu’)，均匀分布U(−bound,bound)
    • torch.nn.init.kaiming_normal_(tensor, a=0, mode=‘fan_in’, nonlinearity=‘leaky_relu’), 正态分布N(0,std)

      两函数的参数： a：该层后面一层的激活函数中负的斜率(默认为ReLU，此时a=0) mode：‘fan_in’ (default) 或者 ‘fan_out’. 使用fan_in保持weights的方差在前向传播中不变；使用fan_out保持weights的方差在反向传播中不变

针对于Relu的激活函数，基本使用He initialization，pytorch也是使用kaiming 初始化卷积层参数的

单层网络

在创建model后直接调用torch.nn.init里的初始化函数

layer1 = torch.nn.Linear(10,20)
torch.nn.init.xavier_uniform_(layer1.weight)
torch.nn.init.constant_(layer1.bias, 0)

重写reset_parameters()方法，并不推荐

多层网络

使用nn.Squential或自定义多层网络

使用torch.nn.Module.apply(fn)

将函数fn递归的运用在每个子模块上，这些子模块由self.children()返回.

注意：此种初始化方式采用的递归，而在python中，对递归层数是有限制的，所以当网络结构很深时，可能会递归层数过深的错误

import torch
from torch import nn

# hyper parameters
in_dim=1
n_hidden_1=1
n_hidden_2=1
out_dim=1

class Net(nn.Module):
    def __init__(self, in_dim, n_hidden_1, n_hidden_2, out_dim):
        super().__init__()
        self.layer = nn.Sequential(
            nn.Linear(in_dim, n_hidden_1), 
            nn.ReLU(True),
            nn.Linear(n_hidden_1, n_hidden_2),
            nn.ReLU(True),
            nn.Linear(n_hidden_2, out_dim)
        )

    def forward(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        return x

# 注意, 在类的外边
# 1. 根据网络层的不同定义不同的初始化方式     
def weight_init(m):
    if isinstance(m, nn.Linear):
        nn.init.xavier_normal_(m.weight)
        nn.init.constant_(m.bias, 0)
    # 也可以判断是否为conv2d，使用相应的初始化方式 
    elif isinstance(m, nn.Conv2d):
        nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
     # 是否为批归一化层
    elif isinstance(m, nn.BatchNorm2d):
        nn.init.constant_(m.weight, 1)
        nn.init.constant_(m.bias, 0)
# 2. 初始化网络结构        
model = Net(in_dim, n_hidden_1, n_hidden_2, out_dim)
# 3. 将weight_init应用在子模块上
model.apply(weight_init)

在__init__中迭代循环self.modules()来初始化网络参数

import torch
from torch import nn

# hyper parameters
in_dim=1
n_hidden_1=1
n_hidden_2=1
out_dim=1

class Net(nn.Module):
    def __init__(self, in_dim, n_hidden_1, n_hidden_2, out_dim):
        super().__init__()
        self.layer = nn.Sequential(
            nn.Linear(in_dim, n_hidden_1), 
            nn.ReLU(True),
            nn.Linear(n_hidden_1, n_hidden_2),
            nn.ReLU(True),
            nn.Linear(n_hidden_2, out_dim)
        )
        # 迭代循环初始化参数
        for m in self.modules():
            if isinstance(m, nn.Linear):
                # 这里只是示例, 不见得要这样定常数
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, -100)
            # 也可以判断是否为conv2d，使用相应的初始化方式 
            elif isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(
                    m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight.item(), 1)
                nn.init.constant_(m.bias.item(), 0)  

    def forward(self, x):
        x = self.layer(x)
        return x

model = Net(in_dim, n_hidden_1, n_hidden_2, out_dim)

# 打印参数信息
def print_weight(m):
    if isinstance(m, nn.Linear):
        print("weight", m.weight.item())
        print("bias:", m.bias.item())
        print("next...")

model.apply(print_weight)

参考链接

ghost commented 5 years ago

Unfold & Fold

a = torch.randn([2, 3, 4, 4])
a
Out[6]: 
tensor([[[[-0.5390, -0.4974,  0.4392,  0.0885],
          [ 0.3316,  0.2863,  0.5387,  0.9645],
          [-0.2879, -0.0852,  0.1790,  2.1958],
          [ 0.2817,  0.4995,  0.6190, -0.5218]],
         [[-0.1495, -0.9248,  1.9004,  0.2535],
          [-0.9124,  0.7679, -0.2503,  0.0491],
          [-1.0860, -0.0838, -0.8773, -1.4696],
          [ 1.5713, -0.9741, -0.1584,  1.1361]],
         [[-0.1027,  2.1711,  0.0953, -0.9208],
          [-1.2121, -1.2770,  1.1427, -0.3149],
          [-0.0458, -1.5204, -0.1037,  0.6764],
          [ 0.3862,  0.6306, -1.0143, -0.1202]]],
        [[[ 1.2074, -1.0920,  0.9833,  0.7729],
          [-0.3728, -0.4250, -0.3600, -0.7940],
          [ 0.6346, -0.4655,  0.9664,  0.4688],
          [-1.0701,  0.0883,  0.2658,  0.0234]],
         [[ 1.5879,  0.5017,  0.4271,  0.6949],
          [ 0.4801, -0.0612, -1.4131,  0.0766],
          [-1.0388, -0.7434, -0.1933, -0.7082],
          [ 0.2480,  0.3196, -2.1165, -0.3998]],
         [[-0.8971, -0.5938, -1.5611,  0.3487],
          [ 1.0478, -0.3852,  0.1441, -2.2990],
          [-0.7650,  0.4652, -1.0962,  1.5915],
          [ 0.7840,  0.2409,  0.3218,  0.4186]]]])

import torch
unfold = nn.Unfold(2, 1, 0, 2)

b = unfold(a)
b
Out[19]: 
tensor([[[-0.5390,  0.4392, -0.2879,  0.1790],
         [-0.4974,  0.0885, -0.0852,  2.1958],
         [ 0.3316,  0.5387,  0.2817,  0.6190],
         [ 0.2863,  0.9645,  0.4995, -0.5218],
         [-0.1495,  1.9004, -1.0860, -0.8773],
         [-0.9248,  0.2535, -0.0838, -1.4696],
         [-0.9124, -0.2503,  1.5713, -0.1584],
         [ 0.7679,  0.0491, -0.9741,  1.1361],
         [-0.1027,  0.0953, -0.0458, -0.1037],
         [ 2.1711, -0.9208, -1.5204,  0.6764],
         [-1.2121,  1.1427,  0.3862, -1.0143],
         [-1.2770, -0.3149,  0.6306, -0.1202]],
        [[ 1.2074,  0.9833,  0.6346,  0.9664],
         [-1.0920,  0.7729, -0.4655,  0.4688],
         [-0.3728, -0.3600, -1.0701,  0.2658],
         [-0.4250, -0.7940,  0.0883,  0.0234],
         [ 1.5879,  0.4271, -1.0388, -0.1933],
         [ 0.5017,  0.6949, -0.7434, -0.7082],
         [ 0.4801, -1.4131,  0.2480, -2.1165],
         [-0.0612,  0.0766,  0.3196, -0.3998],
         [-0.8971, -1.5611, -0.7650, -1.0962],
         [-0.5938,  0.3487,  0.4652,  1.5915],
         [ 1.0478,  0.1441,  0.7840,  0.3218],
         [-0.3852, -2.2990,  0.2409,  0.4186]]])

b.view(2, 12, 2, 2)
Out[18]: 
tensor([[[[-0.5390,  0.4392],
          [-0.2879,  0.1790]],
         [[-0.4974,  0.0885],
          [-0.0852,  2.1958]],
         [[ 0.3316,  0.5387],
          [ 0.2817,  0.6190]],
         [[ 0.2863,  0.9645],
          [ 0.4995, -0.5218]],
         [[-0.1495,  1.9004],
          [-1.0860, -0.8773]],
         [[-0.9248,  0.2535],
          [-0.0838, -1.4696]],
         [[-0.9124, -0.2503],
          [ 1.5713, -0.1584]],
         [[ 0.7679,  0.0491],
          [-0.9741,  1.1361]],
         [[-0.1027,  0.0953],
          [-0.0458, -0.1037]],
         [[ 2.1711, -0.9208],
          [-1.5204,  0.6764]],
         [[-1.2121,  1.1427],
          [ 0.3862, -1.0143]],
         [[-1.2770, -0.3149],
          [ 0.6306, -0.1202]]],
        [[[ 1.2074,  0.9833],
          [ 0.6346,  0.9664]],
         [[-1.0920,  0.7729],
          [-0.4655,  0.4688]],
         [[-0.3728, -0.3600],
          [-1.0701,  0.2658]],
         [[-0.4250, -0.7940],
          [ 0.0883,  0.0234]],
         [[ 1.5879,  0.4271],
          [-1.0388, -0.1933]],
         [[ 0.5017,  0.6949],
          [-0.7434, -0.7082]],
         [[ 0.4801, -1.4131],
          [ 0.2480, -2.1165]],
         [[-0.0612,  0.0766],
          [ 0.3196, -0.3998]],
         [[-0.8971, -1.5611],
          [-0.7650, -1.0962]],
         [[-0.5938,  0.3487],
          [ 0.4652,  1.5915]],
         [[ 1.0478,  0.1441],
          [ 0.7840,  0.3218]],
         [[-0.3852, -2.2990],
          [ 0.2409,  0.4186]]]])

这里可以看出，unfold的操作的收集数据的顺序是按照滑窗区域里W->H->C->N的顺序进行的收集， 而且，这个过程本身也是分通道的过程，是逐通道收集的。

view

temp.view(1, 2*2*2, 1, 1)
Out[16]: 
tensor([[[[2.]],
         [[7.]],
         [[1.]],
         [[3.]],
         [[2.]],
         [[0.]],
         [[1.]],
         [[1.]]]])
temp.view(1, 2*2, 2, 1)
Out[17]: 
tensor([[[[2.],
          [7.]],
         [[1.],
          [3.]],
         [[2.],
          [0.]],
         [[1.],
          [1.]]]])
temp.view(1, 2*2, 1, 2)
Out[18]: 
tensor([[[[2., 7.]],
         [[1., 3.]],
         [[2., 0.]],
         [[1., 1.]]]])

temp
Out[30]: 
tensor([[[[3., 7., 0., 8.],
          [5., 2., 2., 4.],
          [1., 1., 4., 4.],
          [8., 6., 8., 7.]],
         [[3., 3., 4., 0.],
          [5., 0., 0., 7.],
          [3., 2., 8., 3.],
          [8., 5., 8., 8.]]]])
temp.view(1, 2*4, 1, 4)
Out[28]: 
tensor([[[[3., 7., 0., 8.]],
         [[5., 2., 2., 4.]],
         [[1., 1., 4., 4.]],
         [[8., 6., 8., 7.]],
         [[3., 3., 4., 0.]],
         [[5., 0., 0., 7.]],
         [[3., 2., 8., 3.]],
         [[8., 5., 8., 8.]]]])
F.unfold(temp, (2, 2), 1, 0, 2)
Out[29]: 
tensor([[[3., 0., 1., 4.],
         [7., 8., 1., 4.],
         [5., 2., 8., 8.],
         [2., 4., 6., 7.],
         [3., 4., 3., 8.],
         [3., 0., 2., 3.],
         [5., 0., 8., 8.],
         [0., 7., 5., 8.]]])

unfold与view

import torch
import torch.nn.functional as F
a = torch.rand((1, 1, 4, 4))
a
Out[5]: 
tensor([[[[0.6956, 0.2741, 0.7546, 0.6516],
          [0.7810, 0.5884, 0.4314, 0.1446],
          [0.4217, 0.5753, 0.0358, 0.3593],
          [0.1191, 0.0768, 0.3927, 0.3685]]]])
b =F.unfold(a, 2, stride=2)
b
Out[10]: 
tensor([[[0.6956, 0.7546, 0.4217, 0.0358],
         [0.2741, 0.6516, 0.5753, 0.3593],
         [0.7810, 0.4314, 0.1191, 0.3927],
         [0.5884, 0.1446, 0.0768, 0.3685]]])
b.view(1, 4, 2, 2)
Out[9]: 
tensor([[[[0.6956, 0.7546],
          [0.4217, 0.0358]],
         [[0.2741, 0.6516],
          [0.5753, 0.3593]],
         [[0.7810, 0.4314],
          [0.1191, 0.3927]],
         [[0.5884, 0.1446],
          [0.0768, 0.3685]]]])
ghost commented 5 years ago

矩阵乘法

torch.addmm 对矩阵mat1和mat2进行矩阵乘操作(用@表示)。矩阵mat加到最终结果。

out=(beta∗M)+(alpha∗mat1@mat2)
torch.addmm(beta=1, mat, alpha=1, mat1, mat2, out=None)

torch.mm 对矩阵mat1和mat2进行相乘。

torch.mm(mat1, mat2, out=None)

torch.bmm 对存储在两个批batch1和batch2内的矩阵进行批矩阵乘操作。torch.bmm(batch1, batch2, out=None)

用法：

>>> batch1 = torch.randn(10, 3, 4)
>>> batch2 = torch.randn(10, 4, 5)
>>> res = torch.bmm(batch1, batch2)
>>> res.size()
torch.Size([10, 3, 5])
lartpang commented 5 years ago

Shuffle Channels

主要用在了shufflenet中，是一种修改通道顺序的操作，从各组中抽取特定通道进行组合。

    def shuffle_channels(x, groups):
        '''Channel shuffle: [N,C,H,W] -> [N,g,C/g,H,W] -> [N,C/g,g,H,W] -> [N,C,H,W]'''
        '''一共C个channel要分成g组混合的channel，先把C reshape成(g, C/g)的形状，然后转置成(C/g, g)最后平坦成C组channel'''
        N, C, H, W = x.size()
        return x.view(N, groups, C // groups, H, W).permute(0, 2, 1, 3, 4).contiguous().view(N, C, H, W)  # 因为x之前view过了，他的内存不连续了，需要contiguous来规整一下

作者：急流勇进 来源：CSDN 原文：https://blog.csdn.net/weixin_44538273/article/details/88856239 版权声明：本文为博主原创文章，转载请附上博文链接！

另一份实现：

https://github.com/jaxony/ShuffleNet/blob/e9bf42f0cda8dda518cafffd515654cc04584e7a/model.py#L36

def channel_shuffle(x, groups):
    batchsize, num_channels, height, width = x.data.size()

    channels_per_group = num_channels // groups

    # reshape
    x = x.view(batchsize, groups, 
        channels_per_group, height, width)

    # transpose
    # - contiguous() required if transpose() is used before view().
    #   See https://github.com/pytorch/pytorch/issues/764
    x = torch.transpose(x, 1, 2).contiguous()

    # flatten
    x = x.view(batchsize, -1, height, width)

    return x
lartpang commented 5 years ago

Subpixel Conv的像素处理操作

https://github.com/pytorch/examples/blob/1de2ff9338bacaaffa123d03ce53d7522d5dcc2e/super_resolution/model.py#L15

nn.PixelShuffle(upscale_factor)

该操作实现了下图的操作，只是重排了特征，没有其他操作：

image

具体公式：https://blog.csdn.net/oLingFengYu/article/details/87728077