台州外贸网站磁力链最好用的搜索引擎
本文是对CLUT技术的代码讲解，原文解读请看CLUT文章讲解。
1、原文概要
CLUT利用矩阵在保持3DLUT映射能力的前提下显著降低了参数量。整体流程如下所示。
整体还是基于3D-LUT的框架，只不过添加了一个压缩自适应的变换矩阵。作者使用的损失函数在3DLUT的基础上额外添加了一个余弦相似度的损失。
2、代码结构
代码整体结构如下：
核心代码是models.py与LUT.py文件。
3 、核心代码模块
model.py 文件
1. CLUTNet类
这里是网络的整体实现，其定义了backbone、classifier、CLUT。
class CLUTNet(nn.Module): def __init__(self, nsw, dim=33, backbone='Backbone', *args, **kwargs):super().__init__()self.TrilinearInterpolation = TrilinearInterpolation()self.pre = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])self.backbone = eval(backbone)()last_channel = self.backbone.last_channelself.classifier = nn.Sequential(nn.Conv2d(last_channel, 128,1,1),nn.Hardswish(inplace=True),nn.Dropout(p=0.2, inplace=True),nn.Conv2d(128, int(nsw[:2]),1,1),)nsw = nsw.split("+")num, s, w = int(nsw[0]), int(nsw[1]), int(nsw[2])self.CLUTs = CLUT(num, dim, s, w)def fuse_basis_to_one(self, img, TVMN=None):mid_results = self.backbone(self.pre(img))weights = self.classifier(mid_results)[:,:,0,0] # n, numD3LUT, tvmn_loss = self.CLUTs(weights, TVMN)return D3LUT, tvmn_loss def forward(self, img, img_org, TVMN=None):D3LUT, tvmn_loss = self.fuse_basis_to_one(img, TVMN)img_res = self.TrilinearInterpolation(D3LUT, img_org)return {"fakes": img_res + img_org,"3DLUT": D3LUT,"tvmn_loss": tvmn_loss,}
前向中给出了计算过程，首先图像经过backbone计算中间结果，然后经过classifer得到CLUT的输入，最后给到CLUT变换得到实际使用的3DLUT。
2. CLUT类
定义了CLUT的计算过程，讲解中提到了有3个主要参数，num代表LUT的条数，s和w是压缩的参数。
class CLUT(nn.Module):def __init__(self, num, dim=33, s="-1", w="-1", *args, **kwargs):super(CLUT, self).__init__()self.num = numself.dim = dimself.s,self.w = s,w = eval(str(s)), eval(str(w))# +: compressed; -: uncompressedif s == -1 and w == -1: # standard 3DLUTself.mode = '--'self.LUTs = nn.Parameter(torch.zeros(num,3,dim,dim,dim))elif s != -1 and w == -1: self.mode = '+-'self.s_Layers = nn.Parameter(torch.rand(dim, s)/5-0.1)self.LUTs = nn.Parameter(torch.zeros(s, num*3*dim*dim))elif s == -1 and w != -1: self.mode = '-+'self.w_Layers = nn.Parameter(torch.rand(w, dim*dim)/5-0.1)self.LUTs = nn.Parameter(torch.zeros(num*3*dim, w))else: # full-version CLUTself.mode = '++'self.s_Layers = nn.Parameter(torch.rand(dim, s)/5-0.1)self.w_Layers = nn.Parameter(torch.rand(w, dim*dim)/5-0.1)self.LUTs = nn.Parameter(torch.zeros(s*num*3,w))print("n=%d s=%d w=%d"%(num, s, w), self.mode)def reconstruct_luts(self):dim = self.dimnum = self.numif self.mode == "--":D3LUTs = self.LUTselse:if self.mode == "+-":# d,s x s,num*3dd -> d,num*3dd -> d,num*3,dd -> num,3,d,dd -> num,-1CUBEs = self.s_Layers.mm(self.LUTs).reshape(dim,num*3,dim*dim).permute(1,0,2).reshape(num,3,self.dim,self.dim,self.dim)if self.mode == "-+":# num*3d,w x w,dd -> num*3d,dd -> num,3dddCUBEs = self.LUTs.mm(self.w_Layers).reshape(num,3,self.dim,self.dim,self.dim)if self.mode == "++":# s*num*3, w x w, dd -> s*num*3,dd -> s,num*3*dd -> d,num*3*dd -> num,-1CUBEs = self.s_Layers.mm(self.LUTs.mm(self.w_Layers).reshape(-1,num*3*dim*dim)).reshape(dim,num*3,dim**2).permute(1,0,2).reshape(num,3,self.dim,self.dim,self.dim)D3LUTs = cube_to_lut(CUBEs)return D3LUTsdef combine(self, weights, TVMN): # n,numdim = self.dimnum = self.numD3LUTs = self.reconstruct_luts()if TVMN is None:tvmn_loss = 0else:tvmn_loss = TVMN(D3LUTs)D3LUT = weights.mm(D3LUTs.reshape(num,-1)).reshape(-1,3,dim,dim,dim)return D3LUT, tvmn_lossdef forward(self, weights, TVMN=None):lut, tvmn_loss = self.combine(weights, TVMN)return lut, tvmn_loss
mode这里是调整压缩的模式，当然我们需要的是完全压缩的版本，即mode==“++”，可以看到首先会对w_layers与self.LUTs矩阵乘，后续在跟s_layers进行矩阵乘，这与讲解相对应。
utils/LUT.py 文件
1. cube_to_lut函数
此函数在CLUT类的前向完成处理最后会调用到。
def cube_to_lut(cube): # (n,)3,d,d,dif len(cube.shape) == 5:to_shape = [[0,2,3,1],[0,2,1,3],]else:to_shape = [[1,2,0],[1,0,2],]if isinstance(cube, torch.Tensor):lut = torch.empty_like(cube)lut[...,0,:,:,:] = cube[...,0,:,:,:].permute(*to_shape[0])lut[...,1,:,:,:] = cube[...,1,:,:,:].permute(*to_shape[1])lut[...,2,:,:,:] = cube[...,2,:,:,:]else:lut = np.empty_like(cube)lut[...,0,:,:,:] = cube[...,0,:,:,:].transpose(*to_shape[0])lut[...,1,:,:,:] = cube[...,1,:,:,:].transpose(*to_shape[1])lut[...,2,:,:,:] = cube[...,2,:,:,:]return lut
通过CLUT类我们可以看到送入到该函数的输入的shape是(num,3,self.dim,self.dim,self.dim)，因为shape的长度为5，to_shape是[0,2,3,1]以及[0,2,1,3]，也就是说实际的lut是调换通道顺序的cube变量。
3、总结
代码实现核心的部分讲解完毕，跟以往最不同的部分就在于这个CLUT的计算矩阵，把这部分看明白就行。
感谢阅读，欢迎留言或私信，一起探讨和交流。
如果对你有帮助的话，也希望可以给博主点一个关注，感谢。