transfg

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

import copy
import logging
import math

from os.path import join as pjoin

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np

from torch.nn import CrossEntropyLoss, Dropout, Softmax, Linear, Conv2d, LayerNorm
from torch.nn.modules.utils import _pair
from scipy import ndimage

# import models.configs as configs
import configs as configs
from torchkeras import summary


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

def swish(x):
    return x * torch.sigmoid(x)

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

class LabelSmoothing(nn.Module):
    """
    NLL loss with label smoothing.
    """
    def __init__(self, smoothing=0.0):
        """
        Constructor for the LabelSmoothing module.
        :param smoothing: label smoothing factor
        """
        super(LabelSmoothing, self).__init__()
        self.confidence = 1.0 - smoothing
        self.smoothing = smoothing

    def forward(self, x, target):
        logprobs = torch.nn.functional.log_softmax(x, dim=-1)

        nll_loss = -logprobs.gather(dim=-1, index=target.unsqueeze(1))
        nll_loss = nll_loss.squeeze(1)
        smooth_loss = -logprobs.mean(dim=-1)
        loss = self.confidence * nll_loss + self.smoothing * smooth_loss
        return loss.mean()

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

        self.query = Linear(config.hidden_size, self.all_head_size)
        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)

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(self, hidden_states):
        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))
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        attention_probs = self.softmax(attention_scores)
        weights = attention_probs
        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):
    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"])

        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)
        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, in_channels=3):
        super(Embeddings, self).__init__()
        self.hybrid = None
        img_size = _pair(img_size)

        patch_size = _pair(config.patches["size"])
        if config.split == 'non-overlap':
            n_patches = (img_size[0] // patch_size[0]) * (img_size[1] // patch_size[1]) #计算非重叠补丁的数量。这是通过将图像尺寸除以补丁尺寸并相乘得到的
            self.patch_embeddings = Conv2d(in_channels=in_channels, #创建一个 Conv2d 层，用于将输入图像的每个补丁转换为嵌入向量。
                                       out_channels=config.hidden_size,  # 这个层的输出通道数由 config.hidden_size 指定，卷积核大小等于补丁大小，步长也等于补丁大小。
                                       kernel_size=patch_size,
                                       stride=patch_size)
        elif config.split == 'overlap':
            #计算重叠补丁的数量。这是通过将图像尺寸减去补丁大小，然后除以滑动步长 config.slide_step 并相乘得到的。
            n_patches = ((img_size[0] - patch_size[0]) // config.slide_step + 1) * ((img_size[1] - patch_size[1]) // config.slide_step + 1)
            #与非重叠情况类似，创建一个 Conv2d 层，但步长设置为 (config.slide_step, config.slide_step) 以允许重叠。
            self.patch_embeddings = Conv2d(in_channels=in_channels,
                                        out_channels=config.hidden_size,
                                        kernel_size=patch_size,
                                        stride=(config.slide_step, config.slide_step))
        #创建一个位置嵌入张量，它是一个可学习的参数，初始化为全零。这个张量的大小是 (1, n_patches+1, config.hidden_size)，其中 n_patches+1 表示包括一个额外的分类标记（class token）在内的总补丁数。
        self.position_embeddings = nn.Parameter(torch.zeros(1, n_patches+1, config.hidden_size))
        #创建一个分类标记（class token）的嵌入，它也是一个可学习的参数，初始化为全零。
        self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))

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


    def forward(self, x):
        B = x.shape[0]
        cls_tokens = self.cls_token.expand(B, -1, -1)

        if self.hybrid:
            x = self.hybrid_model(x)
        x = self.patch_embeddings(x)
        x = x.flatten(2) #将 x 的第二维和第三维展平。
        x = x.transpose(-1, -2) #交换 x 的最后两个维度
        x = torch.cat((cls_tokens, x), dim=1) #分类标记和展平后的 x 在第一维拼接起来。

        embeddings = x + self.position_embeddings
        embeddings = self.dropout(embeddings)

        return embeddings

class Block(nn.Module):
    def __init__(self, config):
        super(Block, self).__init__()
        self.hidden_size = config.hidden_size
        self.attention_norm = LayerNorm(config.hidden_size, eps=1e-6)
        self.ffn_norm = LayerNorm(config.hidden_size, eps=1e-6)
        self.ffn = Mlp(config)
        self.attn = Attention(config)

    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

    def load_from(self, weights, n_block):
        ROOT = f"Transformer/encoderblock_{n_block}"
        with torch.no_grad():
            query_weight = np2th(weights[pjoin(ROOT, ATTENTION_Q, "kernel")]).view(self.hidden_size, self.hidden_size).t()
            key_weight = np2th(weights[pjoin(ROOT, ATTENTION_K, "kernel")]).view(self.hidden_size, self.hidden_size).t()
            value_weight = np2th(weights[pjoin(ROOT, ATTENTION_V, "kernel")]).view(self.hidden_size, self.hidden_size).t()
            out_weight = np2th(weights[pjoin(ROOT, ATTENTION_OUT, "kernel")]).view(self.hidden_size, self.hidden_size).t()

            query_bias = np2th(weights[pjoin(ROOT, ATTENTION_Q, "bias")]).view(-1)
            key_bias = np2th(weights[pjoin(ROOT, ATTENTION_K, "bias")]).view(-1)
            value_bias = np2th(weights[pjoin(ROOT, ATTENTION_V, "bias")]).view(-1)
            out_bias = np2th(weights[pjoin(ROOT, ATTENTION_OUT, "bias")]).view(-1)

            self.attn.query.weight.copy_(query_weight)
            self.attn.key.weight.copy_(key_weight)
            self.attn.value.weight.copy_(value_weight)
            self.attn.out.weight.copy_(out_weight)
            self.attn.query.bias.copy_(query_bias)
            self.attn.key.bias.copy_(key_bias)
            self.attn.value.bias.copy_(value_bias)
            self.attn.out.bias.copy_(out_bias)

            mlp_weight_0 = np2th(weights[pjoin(ROOT, FC_0, "kernel")]).t()
            mlp_weight_1 = np2th(weights[pjoin(ROOT, FC_1, "kernel")]).t()
            mlp_bias_0 = np2th(weights[pjoin(ROOT, FC_0, "bias")]).t()
            mlp_bias_1 = np2th(weights[pjoin(ROOT, FC_1, "bias")]).t()

            self.ffn.fc1.weight.copy_(mlp_weight_0)
            self.ffn.fc2.weight.copy_(mlp_weight_1)
            self.ffn.fc1.bias.copy_(mlp_bias_0)
            self.ffn.fc2.bias.copy_(mlp_bias_1)

            self.attention_norm.weight.copy_(np2th(weights[pjoin(ROOT, ATTENTION_NORM, "scale")]))
            self.attention_norm.bias.copy_(np2th(weights[pjoin(ROOT, ATTENTION_NORM, "bias")]))
            self.ffn_norm.weight.copy_(np2th(weights[pjoin(ROOT, MLP_NORM, "scale")]))
            self.ffn_norm.bias.copy_(np2th(weights[pjoin(ROOT, MLP_NORM, "bias")]))


#通过计算输入张量x的加权和，然后找到最大值及其索引。
# 输入张量的尺寸为(batch_size, num_channels, height, width)，输出的最大值及其索引的尺寸为(batch_size, height-1, width-1)。
class Part_Attention(nn.Module):
    def __init__(self):
        super(Part_Attention, self).__init__()

    def forward(self, x):
        length = len(x) #获取输入张量x的长度。
        last_map = x[0] #将输入张量x的第一个元素赋值给last_map
        for i in range(1, length): #遍历输入张量x的剩余元素。
            last_map = torch.matmul(x[i], last_map) #将当前元素与last_map进行矩阵乘法，并将结果赋值给last_map。
        last_map = last_map[:,:,0,1:]  #对last_map进行切片操作，保留第0个维度和第1个维度，以及第3个维度从第1个元素开始的所有元素。

        _, max_inx = last_map.max(2) #在last_map的第2个维度上找到最大值，并返回最大值及其索引。
        return _, max_inx

class Encoder(nn.Module):
    def __init__(self, config):
        super(Encoder, self).__init__()
        self.layer = nn.ModuleList()
        #循环 12- 1 次，因为 self.layer 将包含 num_layers - 1 个 Block 层（最后一个层已经在 self.part_layer 中处理）。
        for _ in range(config.transformer["num_layers"] - 1):
            layer = Block(config)
            #创建的 Block 层深拷贝后添加到 self.layer 列表中。这样做是为了确保每个层都是独立的，修改一个层不会影响其他层。
            self.layer.append(copy.deepcopy(layer))
        self.part_select = Part_Attention()
        self.part_layer = Block(config)
        self.part_norm = LayerNorm(config.hidden_size, eps=1e-6)

    def forward(self, hidden_states):
        #创建一个空列表 attn_weights，用于存储每一层的注意力权重。
        attn_weights = []
        #遍历 self.layer 列表中的每个 Block 层。
        for layer in self.layer:
            #输入的隐藏状态 hidden_states 传递给当前的 Block 层，并获取输出的隐藏状态和注意力权重。
            hidden_states, weights = layer(hidden_states)
            attn_weights.append(weights)
        #使用 Part_Attention 层从注意力权重中选择重要的部分
        part_num, part_inx = self.part_select(attn_weights)
        part_inx = part_inx + 1
        parts = []
        #获取 part_inx 的形状，其中 B 是批量大小，num 是选择部分的数量。
        B, num = part_inx.shape
        #遍历每个批量样本。
        for i in range(B):
            #从每个样本的隐藏状态中选择 part_inx[i] 指定的部分，并添加到 parts 列表中。
            parts.append(hidden_states[i, part_inx[i,:]])
        # parts 列表中的张量堆叠起来，并移除中间的维度（假设这个维度是选择的部分数量）。
        parts = torch.stack(parts).squeeze(1)
        #将第一个隐藏状态（通常是分类标记）和选择的部分 parts 在第二维拼接起来。
        concat = torch.cat((hidden_states[:,0].unsqueeze(1), parts), dim=1)
        #将拼接后的张量传递给 part_layer，获取输出的隐藏状态和注意力权重。
        part_states, part_weights = self.part_layer(concat)
        part_encoded = self.part_norm(part_states)   

        return part_encoded

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

    def forward(self, input_ids):
        embedding_output = self.embeddings(input_ids)
        part_encoded = self.encoder(embedding_output)
        return part_encoded

class VisionTransformer(nn.Module):
    def __init__(self, config, img_size=224, num_classes=21843, smoothing_value=0, zero_head=False):
        super(VisionTransformer, self).__init__()
        self.num_classes = num_classes
        self.smoothing_value = smoothing_value
        self.zero_head = zero_head
        self.classifier = config.classifier
        self.transformer = Transformer(config, img_size)
        self.part_head = Linear(config.hidden_size, num_classes)

    def forward(self, x, labels=None):
        part_tokens = self.transformer(x)
        part_logits = self.part_head(part_tokens[:, 0])

        if labels is not None:
            if self.smoothing_value == 0:
                loss_fct = CrossEntropyLoss()
            else:
                loss_fct = LabelSmoothing(self.smoothing_value)
            part_loss = loss_fct(part_logits.view(-1, self.num_classes), labels.view(-1))
            contrast_loss = con_loss(part_tokens[:, 0], labels.view(-1))
            loss = part_loss + contrast_loss
            return loss, part_logits
        else:
            return part_logits

    def load_from(self, weights):
        with torch.no_grad():
            self.transformer.embeddings.patch_embeddings.weight.copy_(np2th(weights["embedding/kernel"], conv=True))
            self.transformer.embeddings.patch_embeddings.bias.copy_(np2th(weights["embedding/bias"]))
            self.transformer.embeddings.cls_token.copy_(np2th(weights["cls"]))
            self.transformer.encoder.part_norm.weight.copy_(np2th(weights["Transformer/encoder_norm/scale"]))
            self.transformer.encoder.part_norm.bias.copy_(np2th(weights["Transformer/encoder_norm/bias"]))

            posemb = np2th(weights["Transformer/posembed_input/pos_embedding"])
            posemb_new = self.transformer.embeddings.position_embeddings
            if posemb.size() == posemb_new.size():
                self.transformer.embeddings.position_embeddings.copy_(posemb)
            else:
                logger.info("load_pretrained: resized variant: %s to %s" % (posemb.size(), posemb_new.size()))
                ntok_new = posemb_new.size(1)

                if self.classifier == "token":
                    posemb_tok, posemb_grid = posemb[:, :1], posemb[0, 1:]
                    ntok_new -= 1
                else:
                    posemb_tok, posemb_grid = posemb[:, :0], posemb[0]

                gs_old = int(np.sqrt(len(posemb_grid)))
                gs_new = int(np.sqrt(ntok_new))
                print('load_pretrained: grid-size from %s to %s' % (gs_old, gs_new))
                posemb_grid = posemb_grid.reshape(gs_old, gs_old, -1)

                zoom = (gs_new / gs_old, gs_new / gs_old, 1)
                posemb_grid = ndimage.zoom(posemb_grid, zoom, order=1)
                posemb_grid = posemb_grid.reshape(1, gs_new * gs_new, -1)
                posemb = np.concatenate([posemb_tok, posemb_grid], axis=1)
                self.transformer.embeddings.position_embeddings.copy_(np2th(posemb))

            for bname, block in self.transformer.encoder.named_children():
                if bname.startswith('part') == False:
                    for uname, unit in block.named_children():
                        unit.load_from(weights, n_block=uname)

            if self.transformer.embeddings.hybrid:
                self.transformer.embeddings.hybrid_model.root.conv.weight.copy_(np2th(weights["conv_root/kernel"], conv=True))
                gn_weight = np2th(weights["gn_root/scale"]).view(-1)
                gn_bias = np2th(weights["gn_root/bias"]).view(-1)
                self.transformer.embeddings.hybrid_model.root.gn.weight.copy_(gn_weight)
                self.transformer.embeddings.hybrid_model.root.gn.bias.copy_(gn_bias)

                for bname, block in self.transformer.embeddings.hybrid_model.body.named_children():
                    for uname, unit in block.named_children():
                        unit.load_from(weights, n_block=bname, n_unit=uname) 

def con_loss(features, labels):
    B, _ = features.shape
    features = F.normalize(features)
    cos_matrix = features.mm(features.t())
    pos_label_matrix = torch.stack([labels == labels[i] for i in range(B)]).float()
    neg_label_matrix = 1 - pos_label_matrix
    pos_cos_matrix = 1 - cos_matrix
    neg_cos_matrix = cos_matrix - 0.4
    neg_cos_matrix[neg_cos_matrix < 0] = 0
    loss = (pos_cos_matrix * pos_label_matrix).sum() + (neg_cos_matrix * neg_label_matrix).sum()
    loss /= (B * B)
    return loss

CONFIGS = {
    'ViT-B_16': configs.get_b16_config(),
    'ViT-B_32': configs.get_b32_config(),
    'ViT-L_16': configs.get_l16_config(),
    'ViT-L_32': configs.get_l32_config(),
    'ViT-H_14': configs.get_h14_config(),
    'testing': configs.get_testing(),
}

if __name__ == "__main__":
    config = CONFIGS['ViT-B_32']
    config.split = 'overlap'
    net = VisionTransformer(config, num_classes=200)
    x = torch.rand((2, 3, 224, 224)).type((torch.float32))
    y = net(x)
    print(net)
    print(y.shape)
    # print(summary(net, x))