??Manifold Mixup和PatchUp是對(duì)mixup數(shù)據(jù)增強(qiáng)算法的兩種改進(jìn)方法，作者都來(lái)自Yoshua Bengio團(tuán)隊(duì)。這兩種方法都是mixup方法在中間隱層的推廣，因此原文開(kāi)源代碼都需要對(duì)網(wǎng)絡(luò)各層的內(nèi)部代碼進(jìn)行修改，使用起來(lái)并不方便，不能做到即插即用。我用pytorch中的鉤子方法(hook)對(duì)這兩個(gè)方法進(jìn)行重新實(shí)現(xiàn)，這樣就可以實(shí)現(xiàn)即插即用，方便的應(yīng)用到各種網(wǎng)絡(luò)結(jié)構(gòu)中，而且我實(shí)現(xiàn)的代碼比原開(kāi)源代碼速度還能提高60%左右。
Manifold Mixup 論文：https://arxiv.org/abs/1806.05236
Manifold Mixup 官方開(kāi)源：https://github.com/vikasverma1077/manifold_mixup
PatchUp 論文：https://arxiv.org/abs/2006.07794
PatchUp 官方開(kāi)源：https://github.com/chandar-lab/PatchUp

一、Manifold Mixup簡(jiǎn)介及代碼

??manifold mixup是對(duì)mixup的擴(kuò)展，把輸入數(shù)據(jù)（raw input data）混合擴(kuò)展到對(duì)中間隱層輸出混合。至于對(duì)中間隱層混合更有效的原因，作者的解釋比較深?yuàn)W。首先給出了現(xiàn)象級(jí)的解釋，即這種混合帶來(lái)了三個(gè)優(yōu)勢(shì)：平滑決策邊界、拉大低置信空間（拉開(kāi)各類別高置信空間的間距）、展平隱層輸出的數(shù)值。至于這三點(diǎn)為什么有效，從作者說(shuō)法看這應(yīng)該是一種業(yè)界共識(shí)。然后作者又從數(shù)學(xué)上分析了第三點(diǎn)，即為什么manifold mixup可以實(shí)現(xiàn)展平中間隱層輸出。
??由于需要修改網(wǎng)絡(luò)中間層的輸出張量，如果不修改網(wǎng)絡(luò)內(nèi)部，也可以使用鉤子操作（hook）在外部進(jìn)行。核心部分代碼如下：

import torch import torch.nn as nn import torch.nn.functional as F import numpy as npdef to_one_hot(inp, num_classes):y_onehot = torch.FloatTensor(inp.size(0), num_classes).to(inp.device)y_onehot.zero_()y_onehot.scatter_(1, inp.unsqueeze(1).data, 1)return y_onehotbce_loss = nn.BCELoss() softmax = nn.Softmax(dim=1)class ManifoldMixupModel(nn.Module):def __init__(self, model, num_classes = 10, alpha = 1):super().__init__()self.model = modelself.alpha = alphaself.lam = Noneself.num_classes = num_classes##選擇需要操作的層，在ResNet中各block的層名為layer1,layer2...所以可以寫(xiě)成如下。其他網(wǎng)絡(luò)請(qǐng)自行修改self. module_list = []for n,m in self.model.named_modules():#if 'conv' in n:if n[:-1]=='layer':self.module_list.append(m)def forward(self, x, target=None):if target==None:out = self.model(x)return outelse:if self.alpha <= 0:self.lam = 1else:self.lam = np.random.beta(self.alpha, self.alpha)k = np.random.randint(-1, len(self.module_list))self.indices = torch.randperm(target.size(0)).cuda()target_onehot = to_one_hot(target, self.num_classes)target_shuffled_onehot = target_onehot[self.indices]if k == -1:x = x * self.lam + x[self.indices] * (1 - self.lam)out = self.model(x)else:modifier_hook = self.module_list[k].register_forward_hook(self.hook_modify)out = self.model(x)modifier_hook.remove()target_reweighted = target_onehot* self.lam + target_shuffled_onehot * (1 - self.lam)loss = bce_loss(softmax(out), target_reweighted)return out, lossdef hook_modify(self, module, input, output):output = self.lam * output + (1 - self.lam) * output[self.indices]return output

調(diào)用代碼如下：

net = ResNet18() net = ManifoldMixupModel(net,num_classes=10, alpha=args.alpha) def train(epoch):net.train()for batch_idx, (inputs, targets) in enumerate(trainloader):inputs, targets = inputs.cuda(), targets.cuda()outputs, loss = net(inputs, targets)optimizer.zero_grad()loss.backward()optimizer.step()def test(epoch):net.eval()with torch.no_grad():for batch_idx, (inputs, targets) in enumerate(testloader):inputs, targets = inputs.cuda(), targets.cuda()outputs = net(inputs)

二、PatchUp簡(jiǎn)介及代碼

??PatchUp方法在manifold mixup基礎(chǔ)上，又借鑒了cutMix在空間維度剪裁的思路，對(duì)中間隱層輸出也進(jìn)行剪裁，對(duì)兩個(gè)不同樣本的中間隱層剪裁塊(patches)進(jìn)行互換或插值，文中稱互換法為硬patchUp，插值法為軟patchUp。試驗(yàn)發(fā)現(xiàn)互換法在識(shí)別精度上更好，插值法在對(duì)抗攻擊的魯棒性上更好。這篇論文中沒(méi)有對(duì)方法理論進(jìn)行深度解釋，僅僅給出了一個(gè)現(xiàn)象級(jí)對(duì)比，就是patchUp方法的隱層激活值比較高。
??使用hook實(shí)現(xiàn)的核心代碼PatchUpModel類如下，注意在該代碼中強(qiáng)制k=-1就可以變成CutMix：

class PatchUpModel(nn.Module):def __init__(self, model, num_classes = 10, block_size=7, gamma=.9, patchup_type='hard',keep_prob=.9):super().__init__()self.patchup_type = patchup_typeself.block_size = block_sizeself.gamma = gammaself.gamma_adj = Noneself.kernel_size = (block_size, block_size)self.stride = (1, 1)self.padding = (block_size // 2, block_size // 2)self.computed_lam = Noneself.model = modelself.num_classes = num_classesself. module_list = []for n,m in self.model.named_modules():if n[:-1]=='layer':#if 'conv' in n:self.module_list.append(m)def adjust_gamma(self, x):return self.gamma * x.shape[-1] ** 2 / \(self.block_size ** 2 * (x.shape[-1] - self.block_size + 1) ** 2)def forward(self, x, target=None):if target==None:out = self.model(x)return outelse:self.lam = np.random.beta(2.0, 2.0)k = np.random.randint(-1, len(self.module_list))self.indices = torch.randperm(target.size(0)).cuda()self.target_onehot = to_one_hot(target, self.num_classes)self.target_shuffled_onehot = self.target_onehot[self.indices]if k == -1: #CutMixW,H = x.size(2),x.size(3)cut_rat = np.sqrt(1. - self.lam)cut_w = np.int(W * cut_rat)cut_h = np.int(H * cut_rat)cx = np.random.randint(W)cy = np.random.randint(H)bbx1 = np.clip(cx - cut_w // 2, 0, W)bby1 = np.clip(cy - cut_h // 2, 0, H)bbx2 = np.clip(cx + cut_w // 2, 0, W)bby2 = np.clip(cy + cut_h // 2, 0, H)x[:, :, bbx1:bbx2, bby1:bby2] = x[self.indices, :, bbx1:bbx2, bby1:bby2]lam = 1 - ((bbx2 - bbx1) * (bby2 - bby1) / (W * H))out = self.model(x)loss = bce_loss(softmax(out), self.target_onehot) * lam +\bce_loss(softmax(out), self.target_shuffled_onehot) * (1. - lam)else:modifier_hook = self.module_list[k].register_forward_hook(self.hook_modify)out = self.model(x)modifier_hook.remove()loss = 1.0 * bce_loss(softmax(out), self.target_a) * self.total_unchanged_portion + \bce_loss(softmax(out), self.target_b) * (1. - self.total_unchanged_portion) + \1.0 * bce_loss(softmax(out), self.target_reweighted)return out, lossdef hook_modify(self, module, input, output):self.gamma_adj = self.adjust_gamma(output)p = torch.ones_like(output[0]) * self.gamma_adjm_i_j = torch.bernoulli(p)mask_shape = len(m_i_j.shape)m_i_j = m_i_j.expand(output.size(0), m_i_j.size(0), m_i_j.size(1), m_i_j.size(2))holes = F.max_pool2d(m_i_j, self.kernel_size, self.stride, self.padding)mask = 1 - holesunchanged = mask * outputif mask_shape == 1:total_feats = output.size(1)else:total_feats = output.size(1) * (output.size(2) ** 2)total_changed_pixels = holes[0].sum()total_changed_portion = total_changed_pixels / total_featsself.total_unchanged_portion = (total_feats - total_changed_pixels) / total_featsif self.patchup_type == 'hard':self.target_reweighted = self.total_unchanged_portion * self.target_onehot +\total_changed_portion * self.target_shuffled_onehotpatches = holes * output[self.indices]self.target_b = self.target_onehot[self.indices]elif self.patchup_type == 'soft':self.target_reweighted = self.total_unchanged_portion * self.target_onehot +\self.lam * total_changed_portion * self.target_onehot +\(1 - self.lam) * total_changed_portion * self.target_shuffled_onehotpatches = holes * outputpatches = patches * self.lam + patches[self.indices] * (1 - self.lam)self.target_b = self.lam * self.target_onehot + (1 - self.lam) * self.target_shuffled_onehotelse:raise ValueError("patchup_type must be \'hard\' or \'soft\'.")output = unchanged + patchesself.target_a = self.target_onehotreturn output

??調(diào)用過(guò)程同上，其中模型包裝語(yǔ)句如下：

net = ResNet18() net = PatchUpModel(net,num_classes=10, block_size=7, gamma=.9, patchup_type='hard')

三、在CIFAR-10上試驗(yàn)結(jié)果

??試驗(yàn)主要目的是驗(yàn)證代碼可運(yùn)行。僅靠在一個(gè)簡(jiǎn)單數(shù)據(jù)集上一次試驗(yàn)非常不充分，不能公平對(duì)比效果，所以不作為各方法的性能對(duì)比。

總結(jié)

以上是生活随笔為你收集整理的[论文学习]Manifold Mixup和PatchUp的代码重新实现（实现即插即用且速度更快）的全部?jī)?nèi)容，希望文章能夠幫你解決所遇到的問(wèn)題。

如果覺(jué)得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。