前言：

本文为记录自己在Nerf学习道路的一些笔记，包括对论文以及其代码的思考内容。公众号： AI知识物语 B站讲解：出门吃三碗饭

本篇文章主要针对其代码来学习其内容，关于代码的理解可能会有出入，欢迎批评指正！！！

1：Paper with code 获取代码

（论文是论文：https://arxiv.org/abs/2003.08934）
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis

代码地址，自取

Nerf 与unity结合

项目目录如下，这是nerf-wild的代码，不要clone错了

2：nerf.py

位置编码函数
有点类似这个函数（代码和下图公告不完全一致！）
注释里有解释其具体函数
（代码） x to (x, sin(2^k x), cos(2^k x), …)
（参考函数）

class PosEmbedding(nn.Module):
    def __init__(self, max_logscale, N_freqs, logscale=True):
        """
        Defines a function that embeds x to (x, sin(2^k x), cos(2^k x), ...)
        """
        super().__init__()
        self.funcs = [torch.sin, torch.cos]

        if logscale:
            self.freqs = 2**torch.linspace(0, max_logscale, N_freqs)
        else:
            self.freqs = torch.linspace(1, 2**max_logscale, N_freqs)

    def forward(self, x):
        """
        Inputs:
            x: (B, 3)

        Outputs:
            out: (B, 6*N_freqs+3)
        """
        out = [x]
        for freq in self.freqs:
            for func in self.funcs:
                out += [func(freq*x)]

        return torch.cat(out, -1)

Nerf网络结构 （重点）

class NeRF(nn.Module):
    def __init__(self, typ,
                 D=8, W=256, skips=[4],
                 in_channels_xyz=63, in_channels_dir=27,
                 encode_appearance=False, in_channels_a=48,
                 encode_transient=False, in_channels_t=16,
                 beta_min=0.03):
        """
        ---Parameters for the original NeRF---
        D: number of layers for density (sigma) encoder
        W: number of hidden units in each layer
        skips: add skip connection in the Dth layer
        in_channels_xyz: number of input channels for xyz (3+3*10*2=63 by default)
        in_channels_dir: number of input channels for direction (3+3*4*2=27 by default)
        in_channels_t: number of input channels for t

        ---Parameters for NeRF-W (used in fine model only as per section 4.3)---
        ---cf. Figure 3 of the paper---
        encode_appearance: whether to add appearance encoding as input (NeRF-A)
        in_channels_a: appearance embedding dimension. n^(a) in the paper
        encode_transient: whether to add transient encoding as input (NeRF-U)
        in_channels_t: transient embedding dimension. n^(tau) in the paper
        beta_min: minimum pixel color variance
        """
        super().__init__()
        self.typ = typ
        self.D = D
        self.W = W
        self.skips = skips
        self.in_channels_xyz = in_channels_xyz
        self.in_channels_dir = in_channels_dir

        self.encode_appearance = False if typ=='coarse' else encode_appearance
        self.in_channels_a = in_channels_a if encode_appearance else 0
        self.encode_transient = False if typ=='coarse' else encode_transient
        self.in_channels_t = in_channels_t
        self.beta_min = beta_min

        # xyz encoding layers
        for i in range(D):
            if i == 0:
                layer = nn.Linear(in_channels_xyz, W)
            elif i in skips:
                layer = nn.Linear(W+in_channels_xyz, W)
            else:
                layer = nn.Linear(W, W)
            layer = nn.Sequential(layer, nn.ReLU(True))
            setattr(self, f"xyz_encoding_{i+1}", layer)
        self.xyz_encoding_final = nn.Linear(W, W)

        # direction encoding layers
        self.dir_encoding = nn.Sequential(
                        nn.Linear(W+in_channels_dir+self.in_channels_a, W//2), nn.ReLU(True))

        # static output layers
        self.static_sigma = nn.Sequential(nn.Linear(W, 1), nn.Softplus())
        self.static_rgb = nn.Sequential(nn.Linear(W//2, 3), nn.Sigmoid())

        if self.encode_transient:
            # transient encoding layers
            self.transient_encoding = nn.Sequential(
                                        nn.Linear(W+in_channels_t, W//2), nn.ReLU(True),
                                        nn.Linear(W//2, W//2), nn.ReLU(True),
                                        nn.Linear(W//2, W//2), nn.ReLU(True),
                                        nn.Linear(W//2, W//2), nn.ReLU(True))
            # transient output layers
            self.transient_sigma = nn.Sequential(nn.Linear(W//2, 1), nn.Softplus())
            self.transient_rgb = nn.Sequential(nn.Linear(W//2, 3), nn.Sigmoid())
            self.transient_beta = nn.Sequential(nn.Linear(W//2, 1), nn.Softplus())

    def forward(self, x, sigma_only=False, output_transient=True):
        """
        Encodes input (xyz+dir) to rgb+sigma (not ready to render yet).
        For rendering this ray, please see rendering.py

        Inputs:
            x: the embedded vector of position (+ direction + appearance + transient)
            sigma_only: whether to infer sigma only.
            has_transient: whether to infer the transient component.

        Outputs (concatenated):
            if sigma_ony:
                static_sigma
            elif output_transient:
                static_rgb, static_sigma, transient_rgb, transient_sigma, transient_beta
            else:
                static_rgb, static_sigma
        """
        if sigma_only:
            input_xyz = x
        elif output_transient:
            input_xyz, input_dir_a, input_t = \
                torch.split(x, [self.in_channels_xyz,
                                self.in_channels_dir+self.in_channels_a,
                                self.in_channels_t], dim=-1)
        else:
            input_xyz, input_dir_a = \
                torch.split(x, [self.in_channels_xyz,
                                self.in_channels_dir+self.in_channels_a], dim=-1)
            

        xyz_ = input_xyz
        for i in range(self.D):
            if i in self.skips:
                xyz_ = torch.cat([input_xyz, xyz_], 1)
            xyz_ = getattr(self, f"xyz_encoding_{i+1}")(xyz_)

        static_sigma = self.static_sigma(xyz_) # (B, 1)
        if sigma_only:
            return static_sigma

        xyz_encoding_final = self.xyz_encoding_final(xyz_)
        dir_encoding_input = torch.cat([xyz_encoding_final, input_dir_a], 1)
        dir_encoding = self.dir_encoding(dir_encoding_input)
        static_rgb = self.static_rgb(dir_encoding) # (B, 3)
        static = torch.cat([static_rgb, static_sigma], 1) # (B, 4)

        if not output_transient:
            return static

        transient_encoding_input = torch.cat([xyz_encoding_final, input_t], 1)
        transient_encoding = self.transient_encoding(transient_encoding_input)
        transient_sigma = self.transient_sigma(transient_encoding) # (B, 1)
        transient_rgb = self.transient_rgb(transient_encoding) # (B, 3)
        transient_beta = self.transient_beta(transient_encoding) # (B, 1)

        transient = torch.cat([transient_rgb, transient_sigma,
                               transient_beta], 1) # (B, 5)

        return torch.cat([static, transient], 1) # (B, 9)

这段代码定义了一个名为NeRF的模块类，用于将输入x编码为RGB颜色和密度(sigma)的输出。具体的功能如下：

初始化函数__init__接受以下参数：

typ：模型类型，可以是coarse或fine
D：密度(sigma)编码器的层数 （默认8层）
W：每层中的隐藏单元数 （默认256个单元）
skips：在第D层中添加跳跃连接的层数列表 （默认第4层跳转）
in_channels_xyz：坐标(xyz)的输入通道数，默认为63（63为经过PosEmbedding后的一个通道数）
in_channels_dir：方向的输入通道数，默认为27 （同理）
encode_appearance：是否添加外观编码作为输入，默认为False
in_channels_a：外观嵌入的维度，默认为48 （ appearance embedding）
encode_transient：是否添加瞬变编码作为输入，默认为False
in_channels_t：瞬变嵌入的维度，默认为16 （ transient embedding）
beta_min：像素颜色方差的最小值，默认为0.03
在初始化函数中，根据模型类型typ和是否编码外观encode_appearance、是否编码瞬变encode_transient进行设置。

初始化xyz编码器层。使用nn.Linear和nn.ReLU构建了多层全连接网络，并使用setattr动态地给模块设置属性，属性名为xyz_encoding_{i+1}，其中i为层的索引。

        for i in range(D):
            if i == 0:
                layer = nn.Linear(in_channels_xyz, W)
            elif i in skips:
                layer = nn.Linear(W+in_channels_xyz, W)
            else:
                layer = nn.Linear(W, W)
            layer = nn.Sequential(layer, nn.ReLU(True))
            setattr(self, f"xyz_encoding_{i+1}", layer)
        self.xyz_encoding_final = nn.Linear(W, W)

初始化方向编码器层。使用nn.Linear和nn.ReLU构建了一个全连接网络。

        self.dir_encoding = nn.Sequential(
                        nn.Linear(W+in_channels_dir+self.in_channels_a, W//2), nn.ReLU(True))

初始化静态输出层。使用nn.Linear和nn.Softplus构建了两个全连接网络，分别用于预测密度(sigma)和RGB颜色。

self.static_sigma = nn.Sequential(nn.Linear(W, 1), nn.Softplus())
        self.static_rgb = nn.Sequential(nn.Linear(W//2, 3), nn.Sigmoid())

如果模型需要编码瞬变，则初始化瞬变编码器层和瞬变输出层。使用nn.Linear和nn.ReLU构建了多层全连接网络，并使用nn.Softplus和nn.Sigmoid构建了多个全连接网络，分别用于预测瞬变的密度(sigma)、RGB颜色和像素颜色方差。

        if self.encode_transient:
            # transient encoding layers
            self.transient_encoding = nn.Sequential(
                                        nn.Linear(W+in_channels_t, W//2), nn.ReLU(True),
                                        nn.Linear(W//2, W//2), nn.ReLU(True),
                                        nn.Linear(W//2, W//2), nn.ReLU(True),
                                        nn.Linear(W//2, W//2), nn.ReLU(True))
            # transient output layers
            self.transient_sigma = nn.Sequential(nn.Linear(W//2, 1), nn.Softplus())
            self.transient_rgb = nn.Sequential(nn.Linear(W//2, 3), nn.Sigmoid())
            self.transient_beta = nn.Sequential(nn.Linear(W//2, 1), nn.Softplus())

前向传播函数forward接受输入x，以及是否只输出密度(sigma)、是否输出瞬变组件的标志。

如果只输出密度(sigma)，则将输入x作为xyz的编码输入，并通过静态密度(sigma)网络得到输出。

如果需要输出瞬变组件，则将输入x分割为xyz的编码输入、方向和外观的输入以及瞬变的输入，并分别经过对应的编码器层得到输出。

将xyz的编码输出和方向/外观的编码输入连接起来，并通过方向编码器层得到输出。然后，通过静态RGB网络得到静态RGB颜色输出。

如果不需要输出瞬变组件，则将静态RGB颜色和静态密度(sigma)连接起来作为输出。

如果需要输出瞬变组件，则将xyz的编码输出和瞬变的输入连接起来，并通过瞬变编码器层得到输出。然后，通过瞬变的RGB、密度(sigma)和像素颜色方差网络得到瞬变组件的输出。

最终，将静态组件和瞬变组件连接起来作为输出。输出形状为(B, 9)，其中B为批量大小。

简略网络图参考

下面是我简单测试后输出的网络层

NeRF(
  (xyz_encoding_1): Sequential(
    (0): Linear(in_features=63, out_features=256, bias=True)
    (1): ReLU(inplace=True)
  )
  (xyz_encoding_2): Sequential(
    (0): Linear(in_features=256, out_features=256, bias=True)
    (1): ReLU(inplace=True)
  )
  (xyz_encoding_3): Sequential(
    (0): Linear(in_features=256, out_features=256, bias=True)
    (1): ReLU(inplace=True)
  )
  (xyz_encoding_4): Sequential(
    (0): Linear(in_features=256, out_features=256, bias=True)
    (1): ReLU(inplace=True)
  )
  (xyz_encoding_5): Sequential(
    (0): Linear(in_features=319, out_features=256, bias=True)
    (1): ReLU(inplace=True)
  )
  (xyz_encoding_6): Sequential(
    (0): Linear(in_features=256, out_features=256, bias=True)
    (1): ReLU(inplace=True)
  )
  (xyz_encoding_7): Sequential(
    (0): Linear(in_features=256, out_features=256, bias=True)
    (1): ReLU(inplace=True)
  )
  (xyz_encoding_8): Sequential(
    (0): Linear(in_features=256, out_features=256, bias=True)
    (1): ReLU(inplace=True)
  )
  (xyz_encoding_final): Linear(in_features=256, out_features=256, bias=True)
  (dir_encoding): Sequential(
    (0): Linear(in_features=283, out_features=128, bias=True)
    (1): ReLU(inplace=True)
  )
  (static_sigma): Sequential(
    (0): Linear(in_features=256, out_features=1, bias=True)
    (1): Softplus(beta=1, threshold=20)
  )
  (static_rgb): Sequential(
    (0): Linear(in_features=128, out_features=3, bias=True)
    (1): Sigmoid()
  )
)

下面这个图是论文NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis 的网络图，其前半部分大致是相同的，全连接层以及对应的relu， skip操作等等
，

3：损失函数

这段代码定义了两个损失函数类：

3.1ColorLoss

ColorLoss类用于计算颜色损失，初始化函数__init__接受一个参数coef，用于调节损失的权重。在初始化函数中，定义了一个均方误差损失函数nn.MSELoss，并将其赋值给属性self.loss。

在前向传播中，计算了粗采样的颜色coarse-color输出与目标值之间的均方误差损失c_l。如果inputs中还包含了细采样fine-color颜色输出rgb_fine，则计算细化颜色输出与目标值之间的均方误差损失f_l。最后，将粗采样颜色损失和细采样颜色损失加权求和，并乘以权重coef作为最终的损失输出。

3.2NerfWLoss

NerfWLoss类用于计算NeRF-W论文中的损失函数，初始化函数__init__接受两个参数coef和lambda_u（默认0.01），分别表示损失的权重和论文中的lambda_u超参数。

在前向传播中，计算了粗采样颜色输出与目标值之间的均方误差损失c_l。如果inputs中还包含了细化颜色输出rgb_fine，则根据是否包含了瞬变组件计算不同的细采样颜色损失。（同理）

如果没有瞬变组件（transient head），使用均方误差损失函数计算细化颜色输出与目标值之间的损失f_l。
如果包含了瞬变组件（transient head），使用公式13中的第一项计算细采样颜色损失f_l，其中分子部分为细采样颜色输出与目标值之间的平方差，分母部分为输入inputs中的beta值的平方乘以2，再取均值。

同时，计算了输入inputs中的beta值的对数的平均值，并加上3以保证其为正数，作为公式13中的第二项beta损失b_l。

ret['b_l'] = 3 + torch.log(inputs['beta']).mean() # +3 to make it positive

最后，计算了输入inputs中的瞬变组件（transient head）密度(sigma)的平均值乘以lambda_u超参数，作为公式13中的第三项sigma损失s_l。

（补充：这里我看其他大佬有写解释的一种说法是，为了避免beta值过大，导致第1项为0，设了第2项，可以让其稍微偏离0）

最后，将粗糙颜色损失、细化颜色损失、beta损失和sigma损失分别乘以权重coef作为最终的损失输出。

其他内容具体可以参考代码

class ColorLoss(nn.Module):
    def __init__(self, coef=1):
        super().__init__()
        self.coef = coef
        self.loss = nn.MSELoss(reduction='mean')

    def forward(self, inputs, targets):
        loss = self.loss(inputs['rgb_coarse'], targets)
        if 'rgb_fine' in inputs:
            loss += self.loss(inputs['rgb_fine'], targets)

        return self.coef * loss


class NerfWLoss(nn.Module):
    """
    Equation 13 in the NeRF-W paper.
    Name abbreviations:
        c_l: coarse color loss
        f_l: fine color loss (1st term in equation 13)
        b_l: beta loss (2nd term in equation 13)
        s_l: sigma loss (3rd term in equation 13)
    """
    def __init__(self, coef=1, lambda_u=0.01):
        """
        lambda_u: in equation 13
        """
        super().__init__()
        self.coef = coef
        self.lambda_u = lambda_u

    def forward(self, inputs, targets):
        ret = {}
        ret['c_l'] = 0.5 * ((inputs['rgb_coarse']-targets)**2).mean()
        if 'rgb_fine' in inputs:
            if 'beta' not in inputs: # no transient head, normal MSE loss
                ret['f_l'] = 0.5 * ((inputs['rgb_fine']-targets)**2).mean()
            else:
                ret['f_l'] = \
                    ((inputs['rgb_fine']-targets)**2/(2*inputs['beta'].unsqueeze(1)**2)).mean()
                ret['b_l'] = 3 + torch.log(inputs['beta']).mean() # +3 to make it positive
                ret['s_l'] = self.lambda_u * inputs['transient_sigmas'].mean()

        for k, v in ret.items():
            ret[k] = self.coef * v

        return ret

loss_dict = {'color': ColorLoss,
             'nerfw': NerfWLoss}

train.py

这部分比较简单，主要是对上述工作的一些调用
定义了一个名为NeRFSystem的类，它是一个继承自LightningModule的模型。NeRFSystem用于实现基于NeRF（Neural Radiance Fields）的渲染系统。

在__init__方法中，模型初始化了一些参数和组件。创建了一些位置编码器（PosEmbedding）用于对位置信息进行嵌入。模型还定义了一个粗糙（coarse）NeRF模型（self.nerf_coarse）和一个细化（fine）NeRF模型（self.nerf_fine），并将它们存储在self.models字典中。

forward方法实现了批量推断。它根据输入的光线和时间戳，调用render_rays函数对光线进行渲染。渲染使用了定义的NeRF模型和嵌入器，以及其他一些参数，例如采样数量、是否使用视差、扰动等。最后，将渲染结果返回。

总体来说，NeRFSystem的作用是实现了一个NeRF渲染系统，可以对输入的光线进行渲染，并返回渲染结果。

class NeRFSystem(LightningModule):
    def __init__(self, hparams):
        super().__init__()
        self.hparams = hparams

        self.loss = loss_dict['nerfw'](coef=1)

        self.models_to_train = []
        self.embedding_xyz = PosEmbedding(hparams.N_emb_xyz-1, hparams.N_emb_xyz)
        self.embedding_dir = PosEmbedding(hparams.N_emb_dir-1, hparams.N_emb_dir)
        self.embeddings = {'xyz': self.embedding_xyz,
                           'dir': self.embedding_dir}

        if hparams.encode_a:
            self.embedding_a = torch.nn.Embedding(hparams.N_vocab, hparams.N_a)
            self.embeddings['a'] = self.embedding_a
            self.models_to_train += [self.embedding_a]
        if hparams.encode_t:
            self.embedding_t = torch.nn.Embedding(hparams.N_vocab, hparams.N_tau)
            self.embeddings['t'] = self.embedding_t
            self.models_to_train += [self.embedding_t]

        self.nerf_coarse = NeRF('coarse',
                                in_channels_xyz=6*hparams.N_emb_xyz+3,
                                in_channels_dir=6*hparams.N_emb_dir+3)
        self.models = {'coarse': self.nerf_coarse}
        if hparams.N_importance > 0:
            self.nerf_fine = NeRF('fine',
                                  in_channels_xyz=6*hparams.N_emb_xyz+3,
                                  in_channels_dir=6*hparams.N_emb_dir+3,
                                  encode_appearance=hparams.encode_a,
                                  in_channels_a=hparams.N_a,
                                  encode_transient=hparams.encode_t,
                                  in_channels_t=hparams.N_tau,
                                  beta_min=hparams.beta_min)
            self.models['fine'] = self.nerf_fine
        self.models_to_train += [self.models]

    def get_progress_bar_dict(self):
        items = super().get_progress_bar_dict()
        items.pop("v_num", None)
        return items

    def forward(self, rays, ts):
        """Do batched inference on rays using chunk."""
        B = rays.shape[0]
        results = defaultdict(list)
        for i in range(0, B, self.hparams.chunk):
            rendered_ray_chunks = \
                render_rays(self.models,
                            self.embeddings,
                            rays[i:i+self.hparams.chunk],
                            ts[i:i+self.hparams.chunk],
                            self.hparams.N_samples,
                            self.hparams.use_disp,
                            self.hparams.perturb,
                            self.hparams.noise_std,
                            self.hparams.N_importance,
                            self.hparams.chunk, # chunk size is effective in val mode
                            self.train_dataset.white_back)

            for k, v in rendered_ray_chunks.items():
                results[k] += [v]

        for k, v in results.items():
            results[k] = torch.cat(v, 0)
        return results

    def setup(self, stage):
        dataset = dataset_dict[self.hparams.dataset_name]
        kwargs = {'root_dir': self.hparams.root_dir}
        if self.hparams.dataset_name == 'phototourism':
            kwargs['img_downscale'] = self.hparams.img_downscale
            kwargs['val_num'] = self.hparams.num_gpus
            kwargs['use_cache'] = self.hparams.use_cache
        elif self.hparams.dataset_name == 'blender':
            kwargs['img_wh'] = tuple(self.hparams.img_wh)
            kwargs['perturbation'] = self.hparams.data_perturb
        self.train_dataset = dataset(split='train', **kwargs)
        self.val_dataset = dataset(split='val', **kwargs)

    def configure_optimizers(self):
        self.optimizer = get_optimizer(self.hparams, self.models_to_train)
        scheduler = get_scheduler(self.hparams, self.optimizer)
        return [self.optimizer], [scheduler]

    def train_dataloader(self):
        return DataLoader(self.train_dataset,
                          shuffle=True,
                          num_workers=4,
                          batch_size=self.hparams.batch_size,
                          pin_memory=True)

    def val_dataloader(self):
        return DataLoader(self.val_dataset,
                          shuffle=False,
                          num_workers=4,
                          batch_size=1, # validate one image (H*W rays) at a time
                          pin_memory=True)
    
    def training_step(self, batch, batch_nb):
        rays, rgbs, ts = batch['rays'], batch['rgbs'], batch['ts']
        results = self(rays, ts)
        loss_d = self.loss(results, rgbs)
        loss = sum(l for l in loss_d.values())

        with torch.no_grad():
            typ = 'fine' if 'rgb_fine' in results else 'coarse'
            psnr_ = psnr(results[f'rgb_{typ}'], rgbs)

        self.log('lr', get_learning_rate(self.optimizer))
        self.log('train/loss', loss)
        for k, v in loss_d.items():
            self.log(f'train/{k}', v, prog_bar=True)
        self.log('train/psnr', psnr_, prog_bar=True)

        return loss

    def validation_step(self, batch, batch_nb):
        rays, rgbs, ts = batch['rays'], batch['rgbs'], batch['ts']
        rays = rays.squeeze() # (H*W, 3)
        rgbs = rgbs.squeeze() # (H*W, 3)
        ts = ts.squeeze() # (H*W)
        results = self(rays, ts)
        loss_d = self.loss(results, rgbs)
        loss = sum(l for l in loss_d.values())
        log = {'val_loss': loss}
        typ = 'fine' if 'rgb_fine' in results else 'coarse'
    
        if batch_nb == 0:
            if self.hparams.dataset_name == 'phototourism':
                WH = batch['img_wh']
                W, H = WH[0, 0].item(), WH[0, 1].item()
            else:
                W, H = self.hparams.img_wh
            img = results[f'rgb_{typ}'].view(H, W, 3).permute(2, 0, 1).cpu() # (3, H, W)
            img_gt = rgbs.view(H, W, 3).permute(2, 0, 1).cpu() # (3, H, W)
            depth = visualize_depth(results[f'depth_{typ}'].view(H, W)) # (3, H, W)
            stack = torch.stack([img_gt, img, depth]) # (3, 3, H, W)
            self.logger.experiment.add_images('val/GT_pred_depth',
                                               stack, self.global_step)

        psnr_ = psnr(results[f'rgb_{typ}'], rgbs)
        log['val_psnr'] = psnr_

        return log

    def validation_epoch_end(self, outputs):
        mean_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
        mean_psnr = torch.stack([x['val_psnr'] for x in outputs]).mean()

        self.log('val/loss', mean_loss)
        self.log('val/psnr', mean_psnr, prog_bar=True)

__init __.py

这个文件里面主要优化optimizer方面的代码
（可以对比下2份代码的优化器，差别不大，不过多介绍就提一下）
代码中给出了四种可能的优化器选择：SGD、Adam、RAdam和Ranger。根据所选的优化器，函数使用给定的超参数（如学习率、动量、权重衰减等）来初始化优化器对象。

最后，函数返回初始化的优化器对象。

因此，这段代码的作用是根据超参数和模型选择合适的优化器，并返回初始化后的优化器对象。

Nerf----------Wild

import torch
# optimizer
from torch.optim import SGD, Adam
import torch_optimizer as optim
# scheduler
from torch.optim.lr_scheduler import CosineAnnealingLR, MultiStepLR
from .warmup_scheduler import GradualWarmupScheduler

from .visualization import *

def get_parameters(models):
    """Get all model parameters recursively."""
    parameters = []
    if isinstance(models, list):
        for model in models:
            parameters += get_parameters(model)
    elif isinstance(models, dict):
        for model in models.values():
            parameters += get_parameters(model)
    else: # models is actually a single pytorch model
        parameters += list(models.parameters())
    return parameters
# 在这里是选择优化器
def get_optimizer(hparams, models):
    eps = 1e-8
    parameters = get_parameters(models)
    if hparams.optimizer == 'sgd':
        optimizer = SGD(parameters, lr=hparams.lr, 
                        momentum=hparams.momentum, weight_decay=hparams.weight_decay)
    elif hparams.optimizer == 'adam':
        optimizer = Adam(parameters, lr=hparams.lr, eps=eps, 
                         weight_decay=hparams.weight_decay)
    elif hparams.optimizer == 'radam':
        optimizer = optim.RAdam(parameters, lr=hparams.lr, eps=eps, 
                                weight_decay=hparams.weight_decay)
    elif hparams.optimizer == 'ranger':
        optimizer = optim.Ranger(parameters, lr=hparams.lr, eps=eps, 
                                 weight_decay=hparams.weight_decay)
    else:
        raise ValueError('optimizer not recognized!')

    return optimizer

def get_scheduler(hparams, optimizer):
    eps = 1e-8
    if hparams.lr_scheduler == 'steplr':
        scheduler = MultiStepLR(optimizer, milestones=hparams.decay_step, 
                                gamma=hparams.decay_gamma)
    elif hparams.lr_scheduler == 'cosine':
        scheduler = CosineAnnealingLR(optimizer, T_max=hparams.num_epochs, eta_min=eps)
    elif hparams.lr_scheduler == 'poly':
        scheduler = LambdaLR(optimizer, 
                             lambda epoch: (1-epoch/hparams.num_epochs)**hparams.poly_exp)
    else:
        raise ValueError('scheduler not recognized!')

    if hparams.warmup_epochs > 0 and hparams.optimizer not in ['radam', 'ranger']:
        scheduler = GradualWarmupScheduler(optimizer, multiplier=hparams.warmup_multiplier, 
                                           total_epoch=hparams.warmup_epochs, after_scheduler=scheduler)

    return scheduler
# 选择超参数 学习率
def get_learning_rate(optimizer):
    for param_group in optimizer.param_groups:
        return param_group['lr']

def extract_model_state_dict(ckpt_path, model_name='model', prefixes_to_ignore=[]):
    checkpoint = torch.load(ckpt_path, map_location=torch.device('cpu'))
    checkpoint_ = {}
    if 'state_dict' in checkpoint: # if it's a pytorch-lightning checkpoint
        checkpoint = checkpoint['state_dict']
    for k, v in checkpoint.items():
        if not k.startswith(model_name):
            continue
        k = k[len(model_name)+1:]
        for prefix in prefixes_to_ignore:
            if k.startswith(prefix):
                print('ignore', k)
                break
        else:
            checkpoint_[k] = v
    return checkpoint_

def load_ckpt(model, ckpt_path, model_name='model', prefixes_to_ignore=[]):
    model_dict = model.state_dict()
    checkpoint_ = extract_model_state_dict(ckpt_path, model_name, prefixes_to_ignore)
    model_dict.update(checkpoint_)
    model.load_state_dict(model_dict)

Nerf ---------optimizer




import math
import torch
from torch.optim.optimizer import Optimizer, required
import itertools as it

class RAdam(Optimizer):

    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0, degenerated_to_sgd=True):
        if not 0.0 <= lr:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {}".format(eps))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
        
        self.degenerated_to_sgd = degenerated_to_sgd
        if isinstance(params, (list, tuple)) and len(params) > 0 and isinstance(params[0], dict):
            for param in params:
                if 'betas' in param and (param['betas'][0] != betas[0] or param['betas'][1] != betas[1]):
                    param['buffer'] = [[None, None, None] for _ in range(10)]
        defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, buffer=[[None, None, None] for _ in range(10)])
        super(RAdam, self).__init__(params, defaults)

    def __setstate__(self, state):
        super(RAdam, self).__setstate__(state)

    def step(self, closure=None):

        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:

            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad.data.float()
                if grad.is_sparse:
                    raise RuntimeError('RAdam does not support sparse gradients')

                p_data_fp32 = p.data.float()

                state = self.state[p]

                if len(state) == 0:
                    state['step'] = 0
                    state['exp_avg'] = torch.zeros_like(p_data_fp32)
                    state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
                else:
                    state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
                    state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)

                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
                beta1, beta2 = group['betas']

                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
                exp_avg.mul_(beta1).add_(1 - beta1, grad)

                state['step'] += 1
                buffered = group['buffer'][int(state['step'] % 10)]
                if state['step'] == buffered[0]:
                    N_sma, step_size = buffered[1], buffered[2]
                else:
                    buffered[0] = state['step']
                    beta2_t = beta2 ** state['step']
                    N_sma_max = 2 / (1 - beta2) - 1
                    N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
                    buffered[1] = N_sma

                    # more conservative since it's an approximated value
                    if N_sma >= 5:
                        step_size = math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
                    elif self.degenerated_to_sgd:
                        step_size = 1.0 / (1 - beta1 ** state['step'])
                    else:
                        step_size = -1
                    buffered[2] = step_size

                # more conservative since it's an approximated value
                if N_sma >= 5:
                    if group['weight_decay'] != 0:
                        p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
                    denom = exp_avg_sq.sqrt().add_(group['eps'])
                    p_data_fp32.addcdiv_(-step_size * group['lr'], exp_avg, denom)
                    p.data.copy_(p_data_fp32)
                elif step_size > 0:
                    if group['weight_decay'] != 0:
                        p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
                    p_data_fp32.add_(-step_size * group['lr'], exp_avg)
                    p.data.copy_(p_data_fp32)

        return loss

class PlainRAdam(Optimizer):

    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0, degenerated_to_sgd=True):
        if not 0.0 <= lr:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {}".format(eps))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
                    
        self.degenerated_to_sgd = degenerated_to_sgd
        defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)

        super(PlainRAdam, self).__init__(params, defaults)

    def __setstate__(self, state):
        super(PlainRAdam, self).__setstate__(state)

    def step(self, closure=None):

        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:

            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad.data.float()
                if grad.is_sparse:
                    raise RuntimeError('RAdam does not support sparse gradients')

                p_data_fp32 = p.data.float()

                state = self.state[p]

                if len(state) == 0:
                    state['step'] = 0
                    state['exp_avg'] = torch.zeros_like(p_data_fp32)
                    state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
                else:
                    state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
                    state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)

                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
                beta1, beta2 = group['betas']

                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
                exp_avg.mul_(beta1).add_(1 - beta1, grad)

                state['step'] += 1
                beta2_t = beta2 ** state['step']
                N_sma_max = 2 / (1 - beta2) - 1
                N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)


                # more conservative since it's an approximated value
                if N_sma >= 5:
                    if group['weight_decay'] != 0:
                        p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
                    step_size = group['lr'] * math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
                    denom = exp_avg_sq.sqrt().add_(group['eps'])
                    p_data_fp32.addcdiv_(-step_size, exp_avg, denom)
                    p.data.copy_(p_data_fp32)
                elif self.degenerated_to_sgd:
                    if group['weight_decay'] != 0:
                        p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
                    step_size = group['lr'] / (1 - beta1 ** state['step'])
                    p_data_fp32.add_(-step_size, exp_avg)
                    p.data.copy_(p_data_fp32)

        return loss

class AdamW(Optimizer):

    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0, warmup = 0):
        if not 0.0 <= lr:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {}".format(eps))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
        
        defaults = dict(lr=lr, betas=betas, eps=eps,
                        weight_decay=weight_decay, warmup = warmup)
        super(AdamW, self).__init__(params, defaults)

    def __setstate__(self, state):
        super(AdamW, self).__setstate__(state)

    def step(self, closure=None):
        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:

            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad.data.float()
                if grad.is_sparse:
                    raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')

                p_data_fp32 = p.data.float()

                state = self.state[p]

                if len(state) == 0:
                    state['step'] = 0
                    state['exp_avg'] = torch.zeros_like(p_data_fp32)
                    state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
                else:
                    state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
                    state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)

                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
                beta1, beta2 = group['betas']

                state['step'] += 1

                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
                exp_avg.mul_(beta1).add_(1 - beta1, grad)

                denom = exp_avg_sq.sqrt().add_(group['eps'])
                bias_correction1 = 1 - beta1 ** state['step']
                bias_correction2 = 1 - beta2 ** state['step']
                
                if group['warmup'] > state['step']:
                    scheduled_lr = 1e-8 + state['step'] * group['lr'] / group['warmup']
                else:
                    scheduled_lr = group['lr']

                step_size = scheduled_lr * math.sqrt(bias_correction2) / bias_correction1
                
                if group['weight_decay'] != 0:
                    p_data_fp32.add_(-group['weight_decay'] * scheduled_lr, p_data_fp32)

                p_data_fp32.addcdiv_(-step_size, exp_avg, denom)

                p.data.copy_(p_data_fp32)

        return loss


#Ranger deep learning optimizer - RAdam + Lookahead combined.
#https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer

#Ranger has now been used to capture 12 records on the FastAI leaderboard.

#This version = 9.3.19  

#Credits:
#RAdam -->  https://github.com/LiyuanLucasLiu/RAdam
#Lookahead --> rewritten by lessw2020, but big thanks to Github @LonePatient and @RWightman for ideas from their code.
#Lookahead paper --> MZhang,G Hinton  https://arxiv.org/abs/1907.08610

#summary of changes: 
#full code integration with all updates at param level instead of group, moves slow weights into state dict (from generic weights), 
#supports group learning rates (thanks @SHolderbach), fixes sporadic load from saved model issues.
#changes 8/31/19 - fix references to *self*.N_sma_threshold; 
                #changed eps to 1e-5 as better default than 1e-8.


class Ranger(Optimizer):

    def __init__(self, params, lr=1e-3, alpha=0.5, k=6, N_sma_threshhold=5, betas=(.95, 0.999), eps=1e-5, weight_decay=0):
        #parameter checks
        if not 0.0 <= alpha <= 1.0:
            raise ValueError(f'Invalid slow update rate: {alpha}')
        if not 1 <= k:
            raise ValueError(f'Invalid lookahead steps: {k}')
        if not lr > 0:
            raise ValueError(f'Invalid Learning Rate: {lr}')
        if not eps > 0:
            raise ValueError(f'Invalid eps: {eps}')

        #parameter comments:
        # beta1 (momentum) of .95 seems to work better than .90...
        #N_sma_threshold of 5 seems better in testing than 4.
        #In both cases, worth testing on your dataset (.90 vs .95, 4 vs 5) to make sure which works best for you.

        #prep defaults and init torch.optim base
        defaults = dict(lr=lr, alpha=alpha, k=k, step_counter=0, betas=betas, N_sma_threshhold=N_sma_threshhold, eps=eps, weight_decay=weight_decay)
        super().__init__(params,defaults)

        #adjustable threshold
        self.N_sma_threshhold = N_sma_threshhold

        #now we can get to work...
        #removed as we now use step from RAdam...no need for duplicate step counting
        #for group in self.param_groups:
        #    group["step_counter"] = 0
            #print("group step counter init")

        #look ahead params
        self.alpha = alpha
        self.k = k 

        #radam buffer for state
        self.radam_buffer = [[None,None,None] for ind in range(10)]

        #self.first_run_check=0

        #lookahead weights
        #9/2/19 - lookahead param tensors have been moved to state storage.  
        #This should resolve issues with load/save where weights were left in GPU memory from first load, slowing down future runs.

        #self.slow_weights = [[p.clone().detach() for p in group['params']]
        #                     for group in self.param_groups]

        #don't use grad for lookahead weights
        #for w in it.chain(*self.slow_weights):
        #    w.requires_grad = False

    def __setstate__(self, state):
        print("set state called")
        super(Ranger, self).__setstate__(state)


    def step(self, closure=None):
        loss = None
        #note - below is commented out b/c I have other work that passes back the loss as a float, and thus not a callable closure.  
        #Uncomment if you need to use the actual closure...

        #if closure is not None:
            #loss = closure()

        #Evaluate averages and grad, update param tensors
        for group in self.param_groups:

            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad.data.float()
                if grad.is_sparse:
                    raise RuntimeError('Ranger optimizer does not support sparse gradients')

                p_data_fp32 = p.data.float()

                state = self.state[p]  #get state dict for this param

                if len(state) == 0:   #if first time to run...init dictionary with our desired entries
                    #if self.first_run_check==0:
                        #self.first_run_check=1
                        #print("Initializing slow buffer...should not see this at load from saved model!")
                    state['step'] = 0
                    state['exp_avg'] = torch.zeros_like(p_data_fp32)
                    state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)

                    #look ahead weight storage now in state dict 
                    state['slow_buffer'] = torch.empty_like(p.data)
                    state['slow_buffer'].copy_(p.data)

                else:
                    state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
                    state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)

                #begin computations 
                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
                beta1, beta2 = group['betas']

                #compute variance mov avg
                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
                #compute mean moving avg
                exp_avg.mul_(beta1).add_(1 - beta1, grad)

                state['step'] += 1


                buffered = self.radam_buffer[int(state['step'] % 10)]
                if state['step'] == buffered[0]:
                    N_sma, step_size = buffered[1], buffered[2]
                else:
                    buffered[0] = state['step']
                    beta2_t = beta2 ** state['step']
                    N_sma_max = 2 / (1 - beta2) - 1
                    N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
                    buffered[1] = N_sma
                    if N_sma > self.N_sma_threshhold:
                        step_size = math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
                    else:
                        step_size = 1.0 / (1 - beta1 ** state['step'])
                    buffered[2] = step_size

                if group['weight_decay'] != 0:
                    p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)

                if N_sma > self.N_sma_threshhold:
                    denom = exp_avg_sq.sqrt().add_(group['eps'])
                    p_data_fp32.addcdiv_(-step_size * group['lr'], exp_avg, denom)
                else:
                    p_data_fp32.add_(-step_size * group['lr'], exp_avg)

                p.data.copy_(p_data_fp32)

                #integrated look ahead...
                #we do it at the param level instead of group level
                if state['step'] % group['k'] == 0:
                    slow_p = state['slow_buffer'] #get access to slow param tensor
                    slow_p.add_(self.alpha, p.data - slow_p)  #(fast weights - slow weights) * alpha
                    p.data.copy_(slow_p)  #copy interpolated weights to RAdam param tensor

        return loss