FAITH/models/transformer_SECA.py

import copy
from typing import Optional, List

import cv2
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn, Tensor
import pdb

# from .FrozenCLIPTextEncoder import FrozenCLIPEmbedder
from .attention_layer import GaussianMultiheadAttention
from pytorch_wavelets import DWTForward # (or import DWT, IDWT)
# from .FFT import FFT
# from .DCT import DCT

class Transformer(nn.Module):

    def __init__(self, config, d_model=512, nhead=8, num_encoder_layers=6,
                 num_decoder_layers=6, dim_feedforward=2048, dropout=0.1,
                 activation="relu", normalize_before=False,
                 return_intermediate_dec=False, smooth=8, dynamic_scale=True):
        super().__init__()

        encoder_layer = TransformerEncoderLayer(d_model, nhead, dim_feedforward,
                                                dropout, activation, normalize_before)
        encoder_norm = nn.LayerNorm(d_model) if normalize_before else None
        self.encoder = TransformerEncoder(encoder_layer, num_encoder_layers, encoder_norm)  # Transformer Encooder

        self.embeddings = DecoderEmbeddings(config)

        decoder_layers = []
        for layer_index in range(num_decoder_layers):
            decoder_layer = TransformerDecoderLayer(dynamic_scale, smooth, layer_index,
                                                    d_model, nhead, dim_feedforward, dropout,
                                                    activation, normalize_before)
            decoder_layers.append(decoder_layer)
        decoder_norm = nn.LayerNorm(d_model)
        self.decoder = TransformerDecoder(decoder_layers, num_decoder_layers, decoder_norm,
                                          return_intermediate=return_intermediate_dec)      # Transformer Decoder

        self._reset_parameters()
        # if dynamic_scale in ["type2", "type3", "type4"]:
        #     for layer_index in range(num_decoder_layers):
        #         nn.init.zeros_(self.decoder.layers[layer_index].point3.weight)      # 用全0初始化权重
        #         with torch.no_grad():
        #             nn.init.ones_(self.decoder.layers[layer_index].point3.bias)     # 用全1初始化偏差
        
        # nn.init.normal_(self.embeddings.projection_layer.weight, mean=0.05, std=0.01)

        self.d_model = d_model
        self.nhead = nhead

    def _reset_parameters(self):
        for p in self.parameters():
            if p.dim() > 1:
                nn.init.xavier_uniform_(p)      # 用均匀分布初始化参数

    def forward(self, src, mask, pos_embed, tgt, tgt_mask, h_w, images, img_path = None):
        # flatten NxCxHxW to HWxNxC
        '''
        src.shape: [bs, hidden_dim, 32, 32]
        mask.shape: [bs, 32, 32]
        pos_embed.shape: [bs, 512, 32, 32]
        tgt.shape: [bs, 5]
        tgt_mask.shape: [bs, 5]
        h_w.shape: [1, bs, 2]
        '''
        bs, c, h, w = src.shape
        grid_y, grid_x = torch.meshgrid(torch.arange(0, h), torch.arange(0, w))     # 返回一个h × w大小的网格，可用于生成坐标       # grid_y.shape: [32, 32], grid_x.shape: [32, 32]
        grid = torch.stack((grid_x, grid_y), 2).float().to(src.device)      # grid.shape: [32, 32, 2]
        grid = grid.reshape(-1, 2).unsqueeze(1).repeat(1, bs * self.nhead, 1)   # grid.shape: [1024, bs * self.nhead, 2]

        src = src.flatten(2).permute(2, 0, 1)       # src.shape: [1024, bs, hidden_dim]
        pos_embed = pos_embed.flatten(2).permute(2, 0, 1)       # pos_embed.shape: [1024, bs, 512]

        tgt = self.embeddings(tgt).permute(1, 0, 2)     # Tokenizer     tgt.shape: [5, bs, hidden_dim]
        query_embed = self.embeddings.position_embeddings.weight.unsqueeze(1)       # query_embed.shape: [5, 1, hidden_dim]
        query_embed = query_embed.repeat(1, bs, 1)     # query_embed.shape: [5, bs, hidden_dim]

        mask = mask.flatten(1)  # mask.shape: [bs, 1024]

        # original Transformer visual encoder
        memory = self.encoder(src, src_key_padding_mask=mask, pos=pos_embed)    # memory.shape: [1024, bs, hidden_dim]

        hs = self.decoder(grid, h_w, tgt, memory, memory_key_padding_mask=mask, tgt_key_padding_mask=tgt_mask,
                                  pos=pos_embed, query_pos=query_embed,
                                  tgt_mask=generate_square_subsequent_mask(len(tgt)).to(tgt.device), images=images, img_path = img_path)        # hs.shape: [max_position_embeddings, bs, hidden_dim]
        # print(hs)
        return hs


class TransformerEncoder(nn.Module):

    def __init__(self, encoder_layer, num_layers, norm=None):
        super().__init__()
        self.layers = _get_clones(encoder_layer, num_layers)
        self.num_layers = num_layers
        self.norm = norm

    def forward(self, src,
                mask: Optional[Tensor] = None,
                src_key_padding_mask: Optional[Tensor] = None,
                pos: Optional[Tensor] = None):
        '''
        src.shape, pos.shape: [1024, bs, hidden_dim]
        mask: None
        src_key_padding_mask.shape: [bs, 1024]
        '''
        output = src

        for layer in self.layers:
            output = layer(output, src_mask=mask,
                           src_key_padding_mask=src_key_padding_mask, pos=pos)

        if self.norm is not None:
            output = self.norm(output)

        return output


class TransformerDecoder(nn.Module):

    def __init__(self, decoder_layer, num_layers, norm=None, return_intermediate=False):
        super().__init__()
        self.layers = nn.ModuleList(decoder_layer)
        self.num_layers = num_layers
        self.norm = norm
        self.return_intermediate = return_intermediate

    def forward(self, grid, h_w, tgt, memory,
                tgt_mask: Optional[Tensor] = None,
                memory_mask: Optional[Tensor] = None,
                tgt_key_padding_mask: Optional[Tensor] = None,
                memory_key_padding_mask: Optional[Tensor] = None,
                pos: Optional[Tensor] = None,
                query_pos: Optional[Tensor] = None, images = None, img_path = None):
        output = tgt

        intermediate = []

        points = []
        point_sigmoid_ref = None
        for layer in self.layers:
            # output, point, point_sigmoid_ref = layer(
            #     grid, h_w, output, memory, tgt_mask=tgt_mask,
            #     memory_mask=memory_mask, tgt_key_padding_mask=tgt_key_padding_mask,
            #     memory_key_padding_mask=memory_key_padding_mask,
            #     pos=pos, query_pos=query_pos, point_ref_previous=point_sigmoid_ref
            # )

            output = layer(
                grid, h_w, output, memory, tgt_mask=tgt_mask,
                memory_mask=memory_mask, tgt_key_padding_mask=tgt_key_padding_mask,
                memory_key_padding_mask=memory_key_padding_mask,
                pos=pos, query_pos=query_pos, point_ref_previous=point_sigmoid_ref, images=images, img_path = img_path
            )
            # points.append(point)
            if self.return_intermediate:
                intermediate.append(self.norm(output))

        if self.norm is not None:
            output = self.norm(output)
            if self.return_intermediate:
                intermediate.pop()
                intermediate.append(output)

        if self.return_intermediate:
            return torch.stack(intermediate), points[0]

        return output


class TransformerEncoderLayer(nn.Module):

    def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1,
                 activation="relu", normalize_before=False):
        super().__init__()
        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
        # Implementation of Feedforward model
        self.linear1 = nn.Linear(d_model, dim_feedforward)
        self.dropout = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model)

        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)

        self.activation = _get_activation_fn(activation)
        self.normalize_before = normalize_before

    def with_pos_embed(self, tensor, pos: Optional[Tensor]):
        return tensor if pos is None else tensor + pos

    def forward_post(self,
                     src,
                     src_mask: Optional[Tensor] = None,
                     src_key_padding_mask: Optional[Tensor] = None,
                     pos: Optional[Tensor] = None):
        q = k = self.with_pos_embed(src, pos)
        src2 = self.self_attn(q, k, value=src, attn_mask=src_mask,
                              key_padding_mask=src_key_padding_mask)[0]
        src = src + self.dropout1(src2)
        src = self.norm1(src)
        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
        src = src + self.dropout2(src2)
        src = self.norm2(src)
        return src

    def forward_pre(self, src,
                    src_mask: Optional[Tensor] = None,
                    src_key_padding_mask: Optional[Tensor] = None,
                    pos: Optional[Tensor] = None):
        '''
        src.shape, pos.shape: [1024, bs, hidden_dim]
        src_mask: None
        src_key_padding_mask.shape: [bs, 1024]
        '''
        src2 = self.norm1(src)  # .shape: [1024, bs, hidden_dim]
        q = k = self.with_pos_embed(src2, pos)      # 将pos添加到feature map上
        src2 = self.self_attn(q, k, value=src2, attn_mask=src_mask,
                              key_padding_mask=src_key_padding_mask)[0]     # .shape: [1024, bs, hidden_dim]
        src = src + self.dropout1(src2)
        src2 = self.norm2(src)
        src2 = self.linear2(self.dropout(self.activation(self.linear1(src2))))
        src = src + self.dropout2(src2)
        return src

    def forward(self, src,
                src_mask: Optional[Tensor] = None,
                src_key_padding_mask: Optional[Tensor] = None,
                pos: Optional[Tensor] = None):
        if self.normalize_before:
            return self.forward_pre(src, src_mask, src_key_padding_mask, pos)
        return self.forward_post(src, src_mask, src_key_padding_mask, pos)


class TransformerDecoderLayer(nn.Module):

    def __init__(self, dynamic_scale, smooth, layer_index,
                 d_model, nhead, dim_feedforward=2048, dropout=0.1,
                 activation="relu", normalize_before=False):
        super().__init__()
        self.wavelettransformer = WaveletTransformer()
        # self.fft = FFT()
        # self.dct = DCT()
        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
        self.multihead_attn = GaussianMultiheadAttention(d_model, nhead, dropout=dropout)
        # Implementation of Feedforward model
        self.linear1 = nn.Linear(d_model, dim_feedforward)
        self.dropout = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model)

        self.smooth = smooth
        self.dynamic_scale = dynamic_scale

        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        self.norm3 = nn.LayerNorm(d_model)
        # self.norm4 = nn.LayerNorm(d_model)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)
        self.dropout3 = nn.Dropout(dropout)

        # if layer_index == 0:
        #     self.point1 = MLP(d_model, d_model, 2, 3)
        #     self.point2 = nn.Linear(d_model, 2*nhead)
        # else:
        #     self.point2 = nn.Linear(d_model, 2*nhead)
        self.layer_index = layer_index
        # if self.dynamic_scale == "type2":
        #     self.point3 = nn.Linear(d_model, nhead)
        # elif self.dynamic_scale == "type3":
        #     self.point3 = nn.Linear(d_model, 2*nhead)
        # elif self.dynamic_scale == "type4":
        #     self.point3 = nn.Linear(d_model, 3*nhead)

        self.activation = _get_activation_fn(activation)
        self.normalize_before = normalize_before

        self.nhead = nhead

    def with_pos_embed(self, tensor, pos: Optional[Tensor]):
        return tensor if pos is None else tensor + pos

    def forward_post(self, grid, h_w, tgt, memory,
                     tgt_mask: Optional[Tensor] = None,
                     memory_mask: Optional[Tensor] = None,
                     tgt_key_padding_mask: Optional[Tensor] = None,
                     memory_key_padding_mask: Optional[Tensor] = None,
                     pos: Optional[Tensor] = None,
                     query_pos: Optional[Tensor] = None,
                     point_ref_previous: Optional[Tensor] = None):
        tgt_len = tgt.shape[0]

        out = self.norm4(tgt + query_pos)
        point_sigmoid_offset = self.point2(out)

        q = k = self.with_pos_embed(tgt, query_pos)
        tgt2 = self.self_attn(q, k, value=tgt, attn_mask=tgt_mask,
                              key_padding_mask=tgt_key_padding_mask)[0]
        tgt = tgt + self.dropout1(tgt2)
        tgt = self.norm1(tgt)

        if self.layer_index == 0:
            point_sigmoid_ref_inter = self.point1(out)
            point_sigmoid_ref = point_sigmoid_ref_inter.sigmoid()
            point_sigmoid_ref = (h_w - 0) * point_sigmoid_ref / int((h_w.max().item()/int(grid.max()+1)))
            point_sigmoid_ref = point_sigmoid_ref.repeat(1, 1, self.nhead)
        else:
            point_sigmoid_ref = point_ref_previous

        point = point_sigmoid_ref + point_sigmoid_offset
        point = point.view(tgt_len, -1, 2)
        distance = (point.unsqueeze(1) - grid.unsqueeze(0)).pow(2)

        if self.dynamic_scale == "type1":
            scale = 1
            distance = distance.sum(-1) * scale
        elif self.dynamic_scale == "type2":
            scale = self.point3(out)
            scale = scale * scale
            scale = scale.reshape(tgt_len, -1).unsqueeze(1)
            distance = distance.sum(-1) * scale
        elif self.dynamic_scale == "type3":
            scale = self.point3(out)
            scale = scale * scale
            scale = scale.reshape(tgt_len, -1, 2).unsqueeze(1)
            distance = (distance * scale).sum(-1)
        elif self.dynamic_scale == "type4":
            scale = self.point3(out)
            scale = scale * scale
            scale = scale.reshape(tgt_len, -1, 3).unsqueeze(1)
            distance = torch.cat([distance, torch.prod(distance, dim=-1, keepdim=True)], dim=-1)
            distance = (distance * scale).sum(-1)

        gaussian = -(distance - 0).abs() / self.smooth

        tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt, query_pos),
                                   key=self.with_pos_embed(memory, pos),
                                   value=memory, attn_mask=memory_mask,
                                   key_padding_mask=memory_key_padding_mask,
                                   gaussian=[gaussian])[0]
        tgt = tgt + self.dropout2(tgt2)
        tgt = self.norm2(tgt)
        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
        tgt = tgt + self.dropout3(tgt2)
        tgt = self.norm3(tgt)
        if self.layer_index == 0:
            return tgt, point_sigmoid_ref_inter, point_sigmoid_ref
        else:
            return tgt, None, point_sigmoid_ref

    def forward_pre(self, grid, h_w, tgt, memory,
                    tgt_mask: Optional[Tensor] = None,
                    memory_mask: Optional[Tensor] = None,
                    tgt_key_padding_mask: Optional[Tensor] = None,
                    memory_key_padding_mask: Optional[Tensor] = None,
                    pos: Optional[Tensor] = None,
                    query_pos: Optional[Tensor] = None,
                    point_ref_previous: Optional[Tensor] = None,
                    idx = -1, images=None, img_path = None):
        '''
        grid.shape: [1024, bs * nhead, 2]
        tgt.shape, query_pos.shape: [5, bs, hidden_dim]
        '''

        # tgt_len = tgt.shape[0]
        # out = self.norm4(tgt + query_pos)
        # point_sigmoid_offset = self.point2(out)    # Δt .shape: [5, bs, 2 * nhead]

        tgt2 = self.norm1(tgt)
        q = k = self.with_pos_embed(tgt2, query_pos)
        tgt2 = self.self_attn(q, k, value=tgt2, attn_mask=tgt_mask,
                              key_padding_mask=tgt_key_padding_mask)[0]     # .shape: [5, bs, hidden_dim]
        tgt = tgt + self.dropout1(tgt2)

        # if self.layer_index == 0:
        #     point_sigmoid_ref_inter = self.point1(out)      # .shape: [5, bs, 2]
        #     point_sigmoid_ref = point_sigmoid_ref_inter.sigmoid()
        #     point_sigmoid_ref = (h_w - 0) * point_sigmoid_ref / int((h_w.max().item()/int(grid.max()+1)))
        #     point_sigmoid_ref = point_sigmoid_ref.repeat(1, 1, self.nhead)  # [5, bs, 2 * nhead]
        # else:
        #     point_sigmoid_ref = point_ref_previous  # t .shape: [5, bs, 2 * nhead]

        # point = point_sigmoid_ref + point_sigmoid_offset
        # point = point.view(tgt_len, -1, 2)      # .shape: [5, bs * nhead, 2]
        # distance = (point.unsqueeze(1) - grid.unsqueeze(0)).pow(2)  # distance.shape: [5, 1024, bs * nhead, 2] = [5, 1, bs * nhead, 2] - [1, 1024, bs * nhead, 2]

        # if self.dynamic_scale == "type1":
        #     scale = 1
        #     distance = distance.sum(-1) * scale
        # elif self.dynamic_scale == "type2":
        #     scale = self.point3(out)
        #     scale = scale * scale
        #     scale = scale.reshape(tgt_len, -1).unsqueeze(1)
        #     distance = distance.sum(-1) * scale
        # elif self.dynamic_scale == "type3":
        #     scale = self.point3(out)    # .shape: [5, bs, 2 * nhead]
        #     scale = scale * scale
        #     scale = scale.reshape(tgt_len, -1, 2).unsqueeze(1)  # .shape: [5, 1, bs * nhead, 2]
        #     distance = (distance * scale).sum(-1)   # distance.shape = [5, 1024, bs * nhead]
        # elif self.dynamic_scale == "type4":
        #     scale = self.point3(out)
        #     scale = scale * scale
        #     scale = scale.reshape(tgt_len, -1, 3).unsqueeze(1)
        #     distance = torch.cat([distance, torch.prod(distance, dim=-1, keepdim=True)], dim=-1)
        #     distance = (distance * scale).sum(-1)

        # gaussian = -(distance - 0).abs() / self.smooth  # .shape: [5, 1024, bs * nhead]
        
        gaussian = self.wavelettransformer(images).permute(1, 0).repeat(5, 1, self.nhead)       # .shape: [5, 1024, bs * nhead]
        # gaussian = self.fft(images).permute(1, 0).repeat(5, 1, self.nhead)
        # gaussian = self.dct(images).permute(1, 0).repeat(5, 1, self.nhead)

        # gaussian[gaussian < (torch.max(gaussian) + torch.min(gaussian)) / 2] = 0
        # pdb.set_trace()

        tgt2 = self.norm2(tgt)
        tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt2, query_pos),
                                   key=self.with_pos_embed(memory, pos),
                                   value=memory, attn_mask=memory_mask,
                                   key_padding_mask=memory_key_padding_mask,
                                   gaussian=[gaussian],
                                   idx = idx, images=images, img_path = img_path)[0]
            
        # 生成attention map
        # attn_out_weights_ = attn_out_weights[0][0][0]

        tgt = tgt + self.dropout2(tgt2)
        tgt2 = self.norm3(tgt)
        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
        tgt = tgt + self.dropout3(tgt2)
        
        # if self.layer_index == 0:
        #     return tgt, point_sigmoid_ref_inter, point_sigmoid_ref
        # else:
        #     return tgt, None, point_sigmoid_ref
        return tgt

    def forward(self, grid, h_w, tgt, memory,
                tgt_mask: Optional[Tensor] = None,
                memory_mask: Optional[Tensor] = None,
                tgt_key_padding_mask: Optional[Tensor] = None,
                memory_key_padding_mask: Optional[Tensor] = None,
                pos: Optional[Tensor] = None,
                query_pos: Optional[Tensor] = None,
                point_ref_previous: Optional[Tensor] = None, images=None, img_path = None):
        if self.normalize_before:
            return self.forward_pre(grid, h_w, tgt, memory, tgt_mask, memory_mask,
                                    tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos,
                                    point_ref_previous, self.layer_index, images=images, img_path = img_path)
        return self.forward_post(grid, h_w, tgt, memory, tgt_mask, memory_mask,
                                 tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos,
                                 point_ref_previous)


# 定义一个包含卷积和池化的网络
class ConvNet(nn.Module):
    def __init__(self):
        super(ConvNet, self).__init__()
        
        # 第一个卷积层，输入9个通道，输出64个通道，卷积核大小3x3，步幅1，填充1
        self.conv1 = nn.Conv2d(9, 64, kernel_size=3, stride=1, padding=1)  
        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)  # 池化操作，每次池化减半空间尺寸
        
        # 第二个卷积层，输入64个通道，输出128个通道，卷积核大小3x3，步幅1，填充1
        self.conv2 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
        
        # 第三个卷积层，输入128个通道，输出1个通道，卷积核大小3x3，步幅1，填充1
        self.conv3 = nn.Conv2d(128, 1, kernel_size=3, stride=1, padding=1)

    def forward(self, x):
        x = self.pool(self.conv1(x))  # (1, 9, 256, 256) -> (1, 64, 256, 256) -> (1, 64, 128, 128)
        x = self.pool(self.conv2(x))  # (1, 64, 128, 128) -> (1, 128, 128, 128) -> (1, 128, 64, 64)
        x = self.pool(self.conv3(x))  # (1, 128, 64, 64) -> (1, 1, 64, 64) -> (1, 1, 32, 32)

        return x.flatten(0)    # (1024,)
    

class WaveletTransformer(nn.Module):
    def __init__(self):
        super().__init__()
        # self.ConvNet = ConvNet()
        self.xfm = DWTForward(J=1, mode='zero', wave='haar').cuda()  # Accepts all wave types available to PyWavelets
        self.xfm.requires_grad_ = False
        self.conv = nn.Conv2d(3, 1, kernel_size=3, stride=1, padding=1) 
        self.pool = nn.AdaptiveAvgPool2d((32, 32))

    # def forward(self, images):
    #     # images.shapes: [bs, 3, 512, 512]
    #     batch_size = images.shape[0]    
    #     # 预先分配输出张量
    #     coeffs = torch.zeros(batch_size, 1024).cuda()
    #     for i in range(batch_size):
    #         # pdb.set_trace()
    #         wave = self.wavelet(images[i])  
    #         # coeffs[i] = self.ConvNet(wave)
    #         coeffs[i] = self.pool(self.conv(wave)).flatten(0)
    #     # pdb.set_trace()
    #     return coeffs       # .shape: [bs, 1024]

    # def wavelet(self, image):
    #     #J为分解的层次数,wave表示使用的变换方法
    #     _, Yh = self.xfm(image.unsqueeze(0))
    #     # pdb.set_trace()
        
    #     # h = torch.cat((Yh[0][:,:,0,:,:], Yh[0][:,:,1,:,:], Yh[0][:,:,2,:,:]), dim=1)
    #     h = Yh[0][:, :, 2, :, :]    # .shape: [1, 3, 256, 256]

    #     return h

    def forward(self, images):
    # 输入 images.shape: [B, 3, 512, 512]
    
        # 批量小波变换
        wave = self.wavelet_batch(images)  # 输出 [B, 3, 256, 256]
        
        # 并行卷积和池化
        coeffs = self.pool(self.conv(wave)).flatten(1)  # 直接处理整个batch
        
        return coeffs  # 输出 [B, 1024]

    def wavelet_batch(self, images):
        # 批量处理版本
        _, Yh = self.xfm(images)          # 保持输入为完整batch
        
        # 直接批量索引操作
        return Yh[0][:, :, 2, :, :]      # 输出 [B, 3, 256, 256]


class MLP(nn.Module):
    """ Very simple multi-layer perceptron (also called FFN)"""

    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
        super().__init__()
        self.num_layers = num_layers
        h = [hidden_dim] * (num_layers - 1)
        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))

    def forward(self, x):
        for i, layer in enumerate(self.layers):
            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
        return x


def _get_clones(module, N):
    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])


class DecoderEmbeddings(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.word_embeddings = nn.Embedding(
            config.vocab_size, config.hidden_dim, padding_idx=config.PAD_token_id)
        # self.clip_text_embeddings = FrozenCLIPEmbedder(max_length=10)

        self.position_embeddings = nn.Embedding(
            config.max_position_embeddings, config.hidden_dim
        )
        # self.projection_layer = nn.Linear(768, config.hidden_dim)
        # self.clip_layernorm = nn.LayerNorm(config.hidden_dim, eps=config.layer_norm_eps)

        self.LayerNorm = torch.nn.LayerNorm(
            config.hidden_dim, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, x):
        # pdb.set_trace()
        input_shape = x.size()  # x.size(): [bs, 5]
        # seq_length = input_shape[1]
        # device = x.device

        position_ids = torch.arange(
            input_shape[1], dtype=torch.long, device=x.device)
        position_ids = position_ids.unsqueeze(0).expand(input_shape)

        input_embeds = self.word_embeddings(x)      # .shape: [bs, 5, hidden_dim]
        # input_embeds = self.clip_text_embeddings(x)     # .shape: [bs, 5, 768]
        # TODO: init
        # input_embeds = self.projection_layer(input_embeds) # .shape: [bs, 5, hidden_dim]
        # input_embeds = self.clip_layernorm(input_embeds)

        position_embeds = self.position_embeddings(position_ids)    # .shape: [bs, 5, hidden_dim]
        # position_embeds = self.LayerNorm(position_embeds)

        embeddings = input_embeds + position_embeds
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)

        return embeddings


# 得到下半部分全为0，上半部分全为'-inf'的矩阵
def generate_square_subsequent_mask(sz):
    r"""Generate a square mask for the sequence. The masked positions are filled with float('-inf').
        Unmasked positions are filled with float(0.0).
    """
    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)    # triu: 获取矩阵的上三角部分
    mask = mask.float().masked_fill(mask == 0, float(
        '-inf')).masked_fill(mask == 1, float(0.0))     
    return mask     # .shape: [sz, sz]


def build_transformer(config):
    return Transformer(
        config,
        d_model=config.hidden_dim,
        dropout=config.dropout,
        nhead=config.nheads,
        dim_feedforward=config.dim_feedforward,
        num_encoder_layers=config.enc_layers,
        num_decoder_layers=config.dec_layers,
        normalize_before=config.pre_norm,
        return_intermediate_dec=False,
        smooth=config.smooth,
        dynamic_scale=config.dynamic_scale,
    )


def _get_activation_fn(activation):
    """Return an activation function given a string"""
    if activation == "relu":
        return F.relu
    if activation == "gelu":
        return F.gelu
    if activation == "glu":
        return F.glu
    raise RuntimeError(F"activation should be relu/gelu, not {activation}.")