models.py

# coding=utf-8
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import copy
import logging
import math
import torch
import torch.nn as nn
import torch.nn.functional as nnf
from torch.nn import Dropout, Softmax, Linear, Conv3d, LayerNorm
from torch.nn.modules.utils import _pair, _triple
import configs as configs
from torch.distributions.normal import Normal

logger = logging.getLogger(__name__)


ATTENTION_Q = "MultiHeadDotProductAttention_1/query"
ATTENTION_K = "MultiHeadDotProductAttention_1/key"
ATTENTION_V = "MultiHeadDotProductAttention_1/value"
ATTENTION_OUT = "MultiHeadDotProductAttention_1/out"
FC_0 = "MlpBlock_3/Dense_0"
FC_1 = "MlpBlock_3/Dense_1"
ATTENTION_NORM = "LayerNorm_0"
MLP_NORM = "LayerNorm_2"


def np2th(weights, conv=False):#转tensor
    """Possibly convert HWIO to OIHW."""
    if conv:
        weights = weights.transpose([3, 2, 0, 1])
    return torch.from_numpy(weights)


def swish(x):#swish激活函数，在深层模型上要优于传统的relu，具有无上界，有下界，光滑，非单调
    return x * torch.sigmoid(x)


ACT2FN = {"gelu": torch.nn.functional.gelu, "relu": torch.nn.functional.relu, "swish": swish}


class Attention(nn.Module):
    def __init__(self, config, vis):
        super(Attention, self).__init__()
        self.vis = vis
        self.num_attention_heads = config.transformer["num_heads"]#12
        self.attention_head_size = int(config.hidden_size / self.num_attention_heads)#252/12=21
        self.all_head_size = self.num_attention_heads * self.attention_head_size#252

        self.query = Linear(config.hidden_size, self.all_head_size)#(252,252)
        self.key = Linear(config.hidden_size, self.all_head_size)
        self.value = Linear(config.hidden_size, self.all_head_size)

        self.out = Linear(config.hidden_size, config.hidden_size)
        self.attn_dropout = Dropout(config.transformer["attention_dropout_rate"])
        self.proj_dropout = Dropout(config.transformer["attention_dropout_rate"])

        self.softmax = Softmax(dim=-1)
    #维度转换函数，输入尺寸：[B, S, H](32, 128, 768)， 输出尺寸：[B, N, S, H/N](32, 8, 128, 768/8)
    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)#[:-1]表示0到倒数第一个数
        x = x.view(*new_x_shape)#元素总数不变， 相当于resize
        return x.permute(0, 2, 1, 3)
    ## 前向传播函数，输入尺寸：[B, S, H](32, 128, 768)， 输出尺寸：[B, S, H](32, 128, 768)
    def forward(self, hidden_states):
        #Q K V 矩阵 尺寸为[B, S, H]
        mixed_query_layer = self.query(hidden_states)
        mixed_key_layer = self.key(hidden_states)
        mixed_value_layer = self.value(hidden_states)

        query_layer = self.transpose_for_scores(mixed_query_layer)
        key_layer = self.transpose_for_scores(mixed_key_layer)
        value_layer = self.transpose_for_scores(mixed_value_layer)

        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))# 将"query"和"key"点乘，得到未经处理注意力值
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        attention_probs = self.softmax(attention_scores)
        weights = attention_probs if self.vis else None
        attention_probs = self.attn_dropout(attention_probs)

        context_layer = torch.matmul(attention_probs, value_layer)
        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)
        attention_output = self.out(context_layer)
        attention_output = self.proj_dropout(attention_output)
        return attention_output, weights


class Mlp(nn.Module):#前向神经网络，多层感知机，也叫人工神经网络（ANN，Artificial Neural Network）有三层神经网络
    def __init__(self, config):
        super(Mlp, self).__init__()
        self.fc1 = Linear(config.hidden_size, config.transformer["mlp_dim"])
        self.fc2 = Linear(config.transformer["mlp_dim"], config.hidden_size)
        self.act_fn = ACT2FN["gelu"]
        self.dropout = Dropout(config.transformer["dropout_rate"])# 0.1

        self._init_weights()

    def _init_weights(self):
        nn.init.xavier_uniform_(self.fc1.weight)
        nn.init.xavier_uniform_(self.fc2.weight)
        nn.init.normal_(self.fc1.bias, std=1e-6)
        nn.init.normal_(self.fc2.bias, std=1e-6)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act_fn(x)
        x = self.dropout(x) # dropout可以避免过拟合，并增强模型的泛化能力。
        x = self.fc2(x)
        x = self.dropout(x)
        return x


class Embeddings(nn.Module):
    """Construct the embeddings from patch, position embeddings.
    """
    def __init__(self, config, img_size):
        super(Embeddings, self).__init__()
        self.config = config
        down_factor = config.down_factor# 2
        patch_size = _triple(config.patches["size"])#(8, 8, 8)
        n_patches = int((img_size[0]/2**down_factor// patch_size[0]) * (img_size[1]/2**down_factor// patch_size[1]) * (img_size[2]/2**down_factor// patch_size[2]))
        # 5*6*7=210,/2**down_factor是因为经过2层卷积后，初始图像尺寸变了
        self.hybrid_model = CNNEncoder(config, n_channels=2)
        in_channels = config['encoder_channels'][-1]#32
        self.patch_embeddings = Conv3d(in_channels=in_channels,
                                       out_channels=config.hidden_size,#252
                                       kernel_size=patch_size,
                                       stride=patch_size)
        self.position_embeddings = nn.Parameter(torch.zeros(1, n_patches, config.hidden_size))#(1, 210, 252)
        # nn.Parameter将一个固定不可训练的tensor转换成可以训练的类型parameter

        self.dropout = Dropout(config.transformer["dropout_rate"])#0.1

    def forward(self, x):
        x, features = self.hybrid_model(x)#x:[2, 32, 40, 48, 56]

        x = self.patch_embeddings(x)  # (B, hidden. n_patches^(1/2), n_patches^(1/2))[2, 252, 5, 6, 7]
        x = x.flatten(2)
        x = x.transpose(-1, -2)  # (B, n_patches, hidden)转置
        embeddings = x + self.position_embeddings#[2, 210, 252]
        embeddings = self.dropout(embeddings)


        return embeddings, features


class Block(nn.Module):
    def __init__(self, config, vis):
        super(Block, self).__init__()
        self.hidden_size = config.hidden_size # 252
        self.attention_norm = LayerNorm(config.hidden_size, eps=1e-6)#LayerNorm 层归一化，train()和eval()对LayerNorm没有影响
        self.ffn_norm = LayerNorm(config.hidden_size, eps=1e-6)
        self.ffn = Mlp(config)
        self.attn = Attention(config, vis)

    def forward(self, x):
        h = x

        x = self.attention_norm(x)
        x, weights = self.attn(x)
        x = x + h

        h = x
        x = self.ffn_norm(x)
        x = self.ffn(x)
        x = x + h
        return x, weights

class Encoder(nn.Module):
    def __init__(self, config, vis):
        super(Encoder, self).__init__()
        self.vis = vis
        self.layer = nn.ModuleList()#在创建 ModuleList 的时候传入一个 module 的 列表，还可以使用extend 函数和 append 函数来添加模型
        self.encoder_norm = LayerNorm(config.hidden_size, eps=1e-6)
        for _ in range(config.transformer["num_layers"]):
            layer = Block(config, vis)
            self.layer.append(copy.deepcopy(layer))#和list 的append 方法一样，将 一个 Module 添加到ModuleList
            #deepcopy是因为layer里有数组，每次数组指向同一个数组，所以不能用copy

    def forward(self, hidden_states):
        attn_weights = []
        for layer_block in self.layer:
            hidden_states, weights = layer_block(hidden_states)
            if self.vis:
                attn_weights.append(weights)
        encoded = self.encoder_norm(hidden_states)
        return encoded, attn_weights


class Transformer(nn.Module):
    def __init__(self, config, img_size, vis):
        super(Transformer, self).__init__()
        self.embeddings = Embeddings(config, img_size=img_size)
        self.encoder = Encoder(config, vis)

    def forward(self, input_ids):
        embedding_output, features = self.embeddings(input_ids)#embedding_output=[2, 210, 252]
        encoded, attn_weights = self.encoder(embedding_output)  # (B, n_patch, hidden)encoded=[2, 210, 252]
        return encoded, attn_weights, features


class Conv3dReLU(nn.Sequential):
    def __init__(
            self,
            in_channels,
            out_channels,
            kernel_size,
            padding=0,
            stride=1,
            use_batchnorm=True,
    ):
        conv = nn.Conv3d(
            in_channels,
            out_channels,
            kernel_size,
            stride=stride,
            padding=padding,
            bias=not (use_batchnorm),
        )
        relu = nn.LeakyReLU(inplace=True)

        IN = nn.InstanceNorm3d(out_channels)

        super(Conv3dReLU, self).__init__(conv, IN, relu)


class DecoderBlock(nn.Module):
    def __init__(
            self,
            in_channels,
            out_channels,
            skip_channels=0,
            use_batchnorm=True,
    ):
        super().__init__()
        self.conv1 = Conv3dReLU(
            in_channels + skip_channels,
            out_channels,
            kernel_size=3,
            padding=1,
            use_batchnorm=use_batchnorm,
        )
        self.conv2 = Conv3dReLU(
            out_channels,
            out_channels,
            kernel_size=3,
            padding=1,
            use_batchnorm=use_batchnorm,
        )
        self.up = nn.Upsample(scale_factor=2, mode='trilinear', align_corners=False)
        #mode是可使用的上采样算法，scale_factor输出为输入的多少倍数， trilinear是三线性插值

    def forward(self, x, skip=None):
        x = self.up(x)
        if skip is not None:
            x = torch.cat([x, skip], dim=1)
        x = self.conv1(x)
        x = self.conv2(x)
        return x

class DecoderCup(nn.Module):
    def __init__(self, config, img_size):
        super().__init__()
        self.config = config
        self.down_factor = config.down_factor
        head_channels = config.conv_first_channel
        self.img_size = img_size
        self.conv_more = Conv3dReLU(
            config.hidden_size,
            head_channels,
            kernel_size=3,
            padding=1,
            use_batchnorm=True,
        )
        decoder_channels = config.decoder_channels
        in_channels = [head_channels] + list(decoder_channels[:-1])
        out_channels = decoder_channels
        self.patch_size = _triple(config.patches["size"])
        skip_channels = self.config.skip_channels#(32, 32, 32, 32, 16)
        blocks = [
            DecoderBlock(in_ch, out_ch, sk_ch) for in_ch, out_ch, sk_ch in zip(in_channels, out_channels, skip_channels)
        ]
        self.blocks = nn.ModuleList(blocks)

    def forward(self, hidden_states, features=None):
        B, n_patch, hidden = hidden_states.size()  # [2, 210, 252]reshape from (B, n_patch, hidden) to (B, h, w, hidden)
        l, h, w = (self.img_size[0]//2**self.down_factor//self.patch_size[0]), (self.img_size[1]//2**self.down_factor//self.patch_size[0]), (self.img_size[2]//2**self.down_factor//self.patch_size[0])
        #5 * 6 * 7 = 210
        x = hidden_states.permute(0, 2, 1)#[2, 252, 210]
        x = x.contiguous().view(B, hidden, l, h, w)#[2, 252, 5, 6, 7]view只能用在contiguous的variable上。如果在view之前用了transpose, permute等，需要用contiguous()来返回一个contiguous copy
        x = self.conv_more(x)
        for i, decoder_block in enumerate(self.blocks):
            if features is not None:
                skip = features[i] if (i < self.config.n_skip) else None
                #print(skip.shape)
            else:
                skip = None
            x = decoder_block(x, skip=skip)
        return x

class SpatialTransformer(nn.Module):
    """
    N-D Spatial Transformer

    Obtained from https://github.com/voxelmorph/voxelmorph
    """

    def __init__(self, size, mode='bilinear'):
        super().__init__()

        self.mode = mode

        # create sampling grid
        vectors = [torch.arange(0, s) for s in size]
        grids = torch.meshgrid(vectors)
        grid = torch.stack(grids)
        grid = torch.unsqueeze(grid, 0)
        grid = grid.type(torch.FloatTensor)

        # registering the grid as a buffer cleanly moves it to the GPU, but it also
        # adds it to the state dict. this is annoying since everything in the state dict
        # is included when saving weights to disk, so the model files are way bigger
        # than they need to be. so far, there does not appear to be an elegant solution.
        # see: https://discuss.pytorch.org/t/how-to-register-buffer-without-polluting-state-dict
        self.register_buffer('grid', grid)

    def forward(self, src, flow):
        # new locations
        #print("self.grid.shape", self.grid.shape)
        #print( "flow.shape", flow.shape )
        new_locs = self.grid + flow
        shape = flow.shape[2:]

        # need to normalize grid values to [-1, 1] for resampler
        for i in range(len(shape)):
            new_locs[:, i, ...] = 2 * (new_locs[:, i, ...] / (shape[i] - 1) - 0.5)

        # move channels dim to last position
        # also not sure why, but the channels need to be reversed
        if len(shape) == 2:
            new_locs = new_locs.permute(0, 2, 3, 1)
            new_locs = new_locs[..., [1, 0]]
        elif len(shape) == 3:
            new_locs = new_locs.permute(0, 2, 3, 4, 1)
            new_locs = new_locs[..., [2, 1, 0]]

        return nnf.grid_sample(src, new_locs, align_corners=True, mode=self.mode)

class DoubleConv(nn.Module): #双层卷积
    """(convolution => [BN] => ReLU) * 2"""

    def __init__(self, in_channels, out_channels, mid_channels=None):
        super().__init__()
        if not mid_channels:
            mid_channels = out_channels
        self.double_conv = nn.Sequential(
            nn.Conv3d(in_channels, mid_channels, kernel_size=3, padding=1),
            nn.LeakyReLU(inplace=True),
            nn.InstanceNorm3d( out_channels ),
            nn.Conv3d(mid_channels, out_channels, kernel_size=3, padding=1),
            nn.LeakyReLU(inplace=True),
            nn.InstanceNorm3d(out_channels)
        )

    def forward(self, x):
        return self.double_conv(x)


class Down(nn.Module):
    """Downscaling with maxpool then double conv"""

    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.maxpool_conv = nn.Sequential(
            nn.MaxPool3d(2),
            DoubleConv(in_channels, out_channels)


        )

    def forward(self, x):
        return self.maxpool_conv(x)

class CNNEncoder(nn.Module):
    def __init__(self, config, n_channels=2):
        super(CNNEncoder, self).__init__()
        self.n_channels = n_channels
        decoder_channels = config.decoder_channels
        encoder_channels = config.encoder_channels#(16, 32, 32)
        self.down_num = config.down_num # 2
        self.inc = DoubleConv(n_channels, encoder_channels[0])
        self.down1 = Down(encoder_channels[0], encoder_channels[1])
        self.down2 = Down(encoder_channels[1], encoder_channels[2])
        self.width = encoder_channels[-1]
    def forward(self, x):
        features = []
        x1 = self.inc(x)#第一层卷积
        features.append(x1)
        x2 = self.down1(x1)#第二层卷积
        features.append(x2)
        feats = self.down2(x2)#第三层卷积
        features.append(feats)
        feats_down = feats
        for i in range(self.down_num):
            feats_down = nn.MaxPool3d(2)(feats_down)
            features.append(feats_down)
        #print("features0:", features[0].size())#[2, 16, 160, 192, 224]
        #print( "features1;", features[1].size() )#[2, 32, 80, 96, 112]
        #print( "features2:", features[2].size() )#[2, 32, 40, 48, 56]
        #print( "features3:", features[3].size() )#[2, 32, 20, 24, 28]
        #print( "features4:", features[4].size() )#[2, 32, 10, 12, 14]
        return feats, features[::-1] # [::-1]所有元素反向

class RegistrationHead(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, upsampling=1):
        conv3d = nn.Conv3d(in_channels, out_channels, kernel_size=kernel_size, padding=kernel_size // 2)
        conv3d.weight = nn.Parameter(Normal(0, 1e-5).sample(conv3d.weight.shape))
        conv3d.bias = nn.Parameter(torch.zeros(conv3d.bias.shape))
        super().__init__(conv3d)

class ViTVNet(nn.Module):
    def __init__(self, config, img_size=(64, 256, 256), int_steps=7, vis=False, mode='bilinear'):
        super(ViTVNet, self).__init__()
        self.transformer = Transformer(config, img_size, vis)
        self.decoder = DecoderCup(config, img_size)
        self.reg_head = RegistrationHead(
            in_channels=config.decoder_channels[-1],#16
            out_channels=config['n_dims'],#3
            kernel_size=3,
        )
        self.spatial_trans = SpatialTransformer(img_size, mode)
        self.config = config
        #self.integrate = VecInt(img_size, int_steps)
    def forward(self, x):
        #x:[2, 2, 160, 192, 224]
        moving = x[:,0:1,:,:]#[2, 1, 160, 192, 224]
        x, attn_weights, features = self.transformer(x)  # (B, n_patch, hidden),x=([2, 210, 252])
        x = self.decoder(x, features)#([2, 16, 160, 192, 224])
        flow = self.reg_head(x)#[2, 3, 160, 192, 224]
        #flow = self.integrate(x)
        #img = x[0].cuda( )
        #flow = x[1].cuda( )
        
        out = self.spatial_trans(moving, flow)#[2, 1, 160, 192, 224])
        #out = self.spatial_trans( source, flow )
        return out, flow

class vm2(nn.Module):
    def __init__(self, dim=3, enc_nf = [16, 32, 32, 32], dec_nf = [32, 32, 32, 32, 32, 16, 16], bn=None, full_size=True, img_size=(64, 256, 256), mode='bilinear'):
        super(vm2, self).__init__()
        #vm1是[32, 32, 32, 32, 8, 8]
        #vm2是[32, 32, 32, 32, 32, 16, 16]
        self.bn = bn
        self.dim = dim
        self.enc_nf = enc_nf
        self.full_size = full_size
        self.vm2 = len(dec_nf) == 7
        # Encoder functions
        self.enc = nn.ModuleList()
        for i in range(len(enc_nf)):
            prev_nf = 2 if i == 0 else enc_nf[i - 1]
            self.enc.append(self.conv_block(dim, prev_nf, enc_nf[i], 4, 2, batchnorm=bn))
        # Decoder functions
        self.dec = nn.ModuleList()
        self.dec.append(self.conv_block(dim, enc_nf[-1], dec_nf[0], batchnorm=bn))  # 1
        self.dec.append(self.conv_block(dim, dec_nf[0] * 2, dec_nf[1], batchnorm=bn))  # 2
        self.dec.append(self.conv_block(dim, dec_nf[1] * 2, dec_nf[2], batchnorm=bn))  # 3
        self.dec.append(self.conv_block(dim, dec_nf[2] + enc_nf[0], dec_nf[3], batchnorm=bn))  # 4
        self.dec.append(self.conv_block(dim, dec_nf[3], dec_nf[4], batchnorm=bn))  # 5

        if self.full_size:
            self.dec.append(self.conv_block(dim, dec_nf[4] + 2, dec_nf[5], batchnorm=bn))
        if self.vm2:
            self.vm2_conv = self.conv_block(dim, dec_nf[5], dec_nf[6], batchnorm=bn)
        self.upsample = nn.Upsample(scale_factor=2, mode='nearest')

        # One conv to get the flow field
        conv_fn = getattr(nn, 'Conv%dd' % dim)
        self.flow = conv_fn(dec_nf[-1], dim, kernel_size=3, padding=1)
        # Make flow weights + bias small. Not sure this is necessary.
        nd = Normal(0, 1e-5)
        self.flow.weight = nn.Parameter(nd.sample(self.flow.weight.shape))
        self.flow.bias = nn.Parameter(torch.zeros(self.flow.bias.shape))
        self.batch_norm = getattr(nn, "BatchNorm{0}d".format(dim))(3)
        self.spatial_trans = SpatialTransformer( img_size, mode )

    def conv_block(self, dim, in_channels, out_channels, kernel_size=3, stride=1, padding=1, batchnorm=False):
        conv_fn = getattr(nn, "Conv{0}d".format(dim))
        bn_fn = getattr(nn, "BatchNorm{0}d".format(dim))
        if batchnorm:
            layer = nn.Sequential(
                conv_fn(in_channels, out_channels, kernel_size, stride=stride, padding=padding),
                bn_fn(out_channels),
                nn.LeakyReLU(0.2))
        else:
            layer = nn.Sequential(
                conv_fn(in_channels, out_channels, kernel_size, stride=stride, padding=padding),
                nn.LeakyReLU(0.2))
        return layer

    def forward(self, x):
        moving = x[:, 0:1, :, :]#[1, 1, 192, 160, 192]tensor
        # Get encoder activations
        x_enc = [x]
        for i, l in enumerate(self.enc):
            x = l(x_enc[-1])
            x_enc.append(x)
        # Three conv + upsample + concatenate series
        y = x_enc[-1]
        for i in range(3):
            y = self.dec[i](y)
            y = self.upsample(y)
            y = torch.cat([y, x_enc[-(i + 2)]], dim=1)
        # Two convs at full_size/2 res
        y = self.dec[3](y)
        y = self.dec[4](y)
        # Upsample to full res, concatenate and conv
        if self.full_size:
            y = self.upsample(y)
            y = torch.cat([y, x_enc[0]], dim=1)
            y = self.dec[5](y)
        # Extra conv for vm2
        if self.vm2:
            y = self.vm2_conv(y)
        flow = self.flow(y)
        if self.bn:
            flow = self.batch_norm(flow)

        out = self.spatial_trans( moving, flow )  # [2, 1, 160, 192, 224])
        return out, flow

class VecInt(nn.Module):
    """
    Integrates a vector field via scaling and squaring.

    Obtained from https://github.com/voxelmorph/voxelmorph
    """

    def __init__(self, inshape, nsteps):
        super().__init__()

        assert nsteps >= 0, 'nsteps should be >= 0, found: %d' % nsteps
        self.nsteps = nsteps
        self.scale = 1.0 / (2 ** self.nsteps)
        self.transformer = SpatialTransformer(inshape)

    def forward(self, vec):
        vec = vec * self.scale
        for _ in range(self.nsteps):
            vec = vec + self.transformer(vec, vec)
        return vec

CONFIGS = {
    'ViT-V-Net': configs.get_3DReg_config(),
}