Skip to content
TheChatArchiveThe Chat Archive
o1-preview-2024-09-12· OpenAIAi2 WildChat-4.8M (AI2)

class MetaMBConvBlock(MetaModule): def init(self, blockargs, globalparams,…

  1. User

    class MetaMBConvBlock(MetaModule):

    def __init__(self, block_args, global_params, image_size=None):
        super().__init__()
        self._block_args = block_args
        self._bn_mom = 1 - global_params.batch_norm_momentum # pytorch's difference from tensorflow
        self._bn_eps = global_params.batch_norm_epsilon
        self.has_se = (self._block_args.se_ratio is not None) and (0 < self._block_args.se_ratio <= 1)
        self.id_skip = block_args.id_skip  # whether to use skip connection and drop connect
    
        # Expansion phase (Inverted Bottleneck)
        inp = self._block_args.input_filters  # number of input channels
        oup = self._block_args.input_filters * self._block_args.expand_ratio  # number of output channels
        if self._block_args.expand_ratio != 1:
            Conv2d = get_same_padding_conv2d_meta(image_size=image_size)
            self._expand_conv = Conv2d(in_channels=inp, out_channels=oup, kernel_size=1, bias=False)
            self._bn0 = MetaBatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
            # image_size = calculate_output_image_size(image_size, 1) <-- this wouldn't modify image_size
    
        # Depthwise convolution phase
        k = self._block_args.kernel_size
        s = self._block_args.stride
        Conv2d = get_same_padding_conv2d_meta(image_size=image_size)
        self._depthwise_conv = Conv2d(
            in_channels=oup, out_channels=oup, groups=oup,  # groups makes it depthwise
            kernel_size=k, stride=s, bias=False)
        self._bn1 = MetaBatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
        
        image_size = calculate_output_image_size(image_size, s)
    
        # Squeeze and Excitation layer, if desired
        if self.has_se:
            Conv2d = get_same_padding_conv2d_meta(image_size=(1,1))
            num_squeezed_channels = max(1, int(self._block_args.input_filters * self._block_args.se_ratio))
            self._se_reduce = Conv2d(in_channels=oup, out_channels=num_squeezed_channels, kernel_size=1)
            self._se_expand = Conv2d(in_channels=num_squeezed_channels, out_channels=oup, kernel_size=1)
    
        # Pointwise convolution phase
        final_oup = self._block_args.output_filters
        Conv2d = get_same_padding_conv2d_meta(image_size=image_size)
        self._project_conv = Conv2d(in_channels=oup, out_channels=final_oup, kernel_size=1, bias=False)
        self._bn2 = MetaBatchNorm2d(num_features=final_oup, momentum=self._bn_mom, eps=self._bn_eps)
        
    
        self._swish = MemoryEfficientSwish()
    
    def forward(self, inputs, drop_connect_rate=None):
        """MBConvBlock's forward function.
    
        Args:
            inputs (tensor): Input tensor.
            drop_connect_rate (bool): Drop connect rate (float, between 0 and 1).
    
        Returns:
            Output of this block after processing.
        """
    
        # Expansion and Depthwise Convolution
        x = inputs
        if self._block_args.expand_ratio != 1:
            x = self._expand_conv(inputs)
            x = self._bn0(x)
            x = self._swish(x)
    
        x = self._depthwise_conv(x)
        x = self._bn1(x)
        
        x = self._swish(x)
    
        # Squeeze and Excitation
        if self.has_se:
            x_squeezed = F.adaptive_avg_pool2d(x, 1)
            x_squeezed = self._se_reduce(x_squeezed)
            x_squeezed = self._swish(x_squeezed)
            x_squeezed = self._se_expand(x_squeezed)
            x = torch.sigmoid(x_squeezed) * x
    
        # Pointwise Convolution
        x = self._project_conv(x)
        x = self._bn2(x)
        
        # Skip connection and drop connect
        input_filters, output_filters = self._block_args.input_filters, self._block_args.output_filters
        if self.id_skip and self._block_args.stride == 1 and input_filters == output_filters:
            # The combination of skip connection and drop connect brings about stochastic depth.
            if drop_connect_rate:
                x = drop_connect(x, p=drop_connect_rate, training=self.training)
            x = x + inputs  # skip connection
        return x
    
    def set_swish(self, memory_efficient=True):
        """Sets swish function as memory efficient (for training) or standard (for export).
    
        Args:
            memory_efficient (bool): Whether to use memory-efficient version of swish.
        """
        self._swish = MemoryEfficientSwish() if memory_efficient else Swish()
    

    class MetaEfficientNet(MetaModule):

    def __init__(self, blocks_args=None, global_params=None,in_channels=3):
        super().__init__()
    
        assert isinstance(blocks_args, list), 'blocks_args should be a list'
        assert len(blocks_args) > 0, 'block args must be greater than 0'
        self._global_params = global_params
        self._blocks_args = blocks_args
    
        # Batch norm parameters
        bn_mom = 1 - self._global_params.batch_norm_momentum
        bn_eps = self._global_params.batch_norm_epsilon
    
        # Get stem static or dynamic convolution depending on image size
        image_size = global_params.image_size
        Conv2d = get_same_padding_conv2d_meta(image_size=image_size)
    
        # Stem
        in_channels = in_channels  # rgb
        out_channels = round_filters(32, self._global_params)  # number of output channels
        self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)
        self._bn0 = MetaBatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
        image_size = calculate_output_image_size(image_size, 2)
    
        # Build blocks
        num_block = len(self._blocks_args)
        self._blocks = nn.ModuleList([])
        
        for i,block_args in enumerate(self._blocks_args):
    
            # Update block input and output filters based on depth multiplier.
            block_args = block_args._replace(
                input_filters=round_filters(block_args.input_filters, self._global_params),
                output_filters=round_filters(block_args.output_filters, self._global_params),
                num_repeat=round_repeats(block_args.num_repeat, self._global_params)
            )
    
            # The first block needs to take care of stride and filter size increase.
    
            self._blocks.append(MetaMBConvBlock(block_args, self._global_params, image_size=image_size))
            #self._blocks.append(MBConvBlockIN(block_args, self._global_params, image_size=image_size))
            image_size = calculate_output_image_size(image_size, block_args.stride)
            if block_args.num_repeat > 1: # modify block_args to keep same output size
                block_args = block_args._replace(input_filters=block_args.output_filters, stride=1)
            for _ in range(block_args.num_repeat - 1):
                self._blocks.append(MetaMBConvBlock(block_args, self._global_params, image_size=image_size))
        # Head
        in_channels = block_args.output_filters  # output of final block
        out_channels = round_filters(1280, self._global_params)
        Conv2d = get_same_padding_conv2d_meta(image_size=image_size)
        self._conv_head = Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
        self._bn1 = MetaBatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
    
        # Final linear layer
        self._avg_pooling = nn.AdaptiveAvgPool2d(1)
        self._dropout = nn.Dropout(self._global_params.dropout_rate)
        self._fc = MetaLinear(out_channels, self._global_params.num_classes)
        self._swish = MemoryEfficientSwish()
    
    def set_swish(self, memory_efficient=True):
        """Sets swish function as memory efficient (for training) or standard (for export).
    
        Args:
            memory_efficient (bool): Whether to use memory-efficient version of swish.
    
        """
        self._swish = MemoryEfficientSwish() if memory_efficient else Swish()
        for block in self._blocks:
            block.set_swish(memory_efficient)
    
    
    def extract_features(self, inputs):
        """use convolution layer to extract feature .
    
        Args:
            inputs (tensor): Input tensor.
    
        Returns:
            Output of the final convolution 
            layer in the efficientnet model.
        """
        # Stem
        x = self._swish(self._bn0(self._conv_stem(inputs)))
    
        # Blocks
        for idx, block in enumerate(self._blocks):
            drop_connect_rate = self._global_params.drop_connect_rate
            if drop_connect_rate:
                drop_connect_rate *= float(idx) / len(self._blocks) # scale drop connect_rate
            x = block(x, drop_connect_rate=drop_connect_rate)
        
        # Head
        x = self._swish(self._bn1(self._conv_head(x)))
    
        return x
    
    
    def forward(self, inputs):
        """EfficientNet's forward function.
           Calls extract_features to extract features, applies final linear layer, and returns logits.
    
        Args:
            inputs (tensor): Input tensor.
    
        Returns:
            Output of this model after processing.
        """
        bs = inputs.size(0)
    
        # Convolution layers
        feature = self.extract_features(inputs)
    
        # Pooling and final linear layer
        x = self._avg_pooling(feature)
        x = x.view(bs, -1)
        x = self._dropout(x)
        x = self._fc(x)
    
    
        return x,feature
    
    @classmethod
    def from_name(cls, model_name, in_channels=3, **override_params):
        """create an efficientnet model according to name.
    
        Args:
            model_name (str): Name for efficientnet.
            in_channels (int): Input data's channel number.
            override_params (other key word params): 
                Params to override model's global_params.
                Optional key:
                    'width_coefficient', 'depth_coefficient',
                    'image_size', 'dropout_rate',
                    'num_classes', 'batch_norm_momentum',
                    'batch_norm_epsilon', 'drop_connect_rate',
                    'depth_divisor', 'min_depth'
    
        Returns:
            An efficientnet model.
        """
        cls._check_model_name_is_valid(model_name)
        blocks_args, global_params = get_model_params(model_name, override_params)
        model = cls(blocks_args, global_params)
        model._change_in_channels(in_channels)
        return model
    
    @classmethod
    def from_pretrained(cls, model_name, weights_path=None, advprop=False, 
                        in_channels=3, num_classes=1000, **override_params):
        """create an efficientnet model according to name.
    
        Args:
            model_name (str): Name for efficientnet.
            weights_path (None or str): 
                str: path to pretrained weights file on the local disk.
                None: use pretrained weights downloaded from the Internet.
            advprop (bool): 
                Whether to load pretrained weights
                trained with advprop (valid when weights_path is None).
            in_channels (int): Input data's channel number.
            num_classes (int): 
                Number of categories for classification.
                It controls the output size for final linear layer.
            override_params (other key word params): 
                Params to override model's global_params.
                Optional key:
                    'width_coefficient', 'depth_coefficient',
                    'image_size', 'dropout_rate',
                    'num_classes', 'batch_norm_momentum',
                    'batch_norm_epsilon', 'drop_connect_rate',
                    'depth_divisor', 'min_depth'
    
        Returns:
            A pretrained efficientnet model.
        """
        model = cls.from_name(model_name, num_classes = num_classes, **override_params)
        load_pretrained_weights(model, model_name, weights_path=weights_path, load_fc=(num_classes == 1000), advprop=advprop)
        model._change_in_channels(in_channels)
        return model
    
    @classmethod
    def get_image_size(cls, model_name):
        """Get the input image size for a given efficientnet model.
    
        Args:
            model_name (str): Name for efficientnet.
    
        Returns:
            Input image size (resolution).
        """
        cls._check_model_name_is_valid(model_name)
        _, _, res, _ = efficientnet_params(model_name)
        return res
    
    @classmethod
    def _check_model_name_is_valid(cls, model_name):
        """Validates model name. 
    
        Args:
            model_name (str): Name for efficientnet.
    
        Returns:
            bool: Is a valid name or not.
        """
        valid_models = ['efficientnet-b'+str(i) for i in range(9)]
    
        # Support the construction of 'efficientnet-l2' without pretrained weights
        valid_models += ['efficientnet-l2']
    
        if model_name not in valid_models:
            raise ValueError('model_name should be one of: ' + ', '.join(valid_models))
    
    def _change_in_channels(self, in_channels):
        """Adjust model's first convolution layer to in_channels, if in_channels not equals 3.
    
        Args:
            in_channels (int): Input data's channel number.
        """
        if in_channels != 3:
            Conv2d = get_same_padding_conv2d(image_size = self._global_params.image_size)
            out_channels = round_filters(32, self._global_params)
            self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)用中文详细解释这段代码
    
  2. Assistant

    这个代码是一个基于 PyTorch 实现的 EfficientNet 模型的变体,名为 MetaEfficientNet,它继承自 MetaModule。MetaModule 通常用于元学习(Meta-Learning)场景,如 MAML(Model-Agnostic Meta-Learning),允许在梯度计算中对模型的参数进行高阶求导。

    以下是代码的详细解释:


    MetaMBConvBlock 类

    这个类实现了 EfficientNet 中的基础构建块,即 Mobile Inverted Residual Bottleneck Convolutional Block,简称 MBConvBlock。这个块具有残差连接、扩展卷积、深度可分离卷积以及 Squeeze-and-Excitation(SE)模块。

    初始化方法 __init__

    • 参数初始化:

      • block_args:包含块的配置参数,如输入通道数、输出通道数、卷积核大小、步幅、扩展系数等。
      • global_params:包含全局配置参数,如批归一化的动量和 epsilon 值等。
      • image_size:输入图像的尺寸。
    • 成员变量设置:

      • _block_args:保存块的参数。
      • _bn_mom 和 _bn_eps:批归一化的动量和 epsilon,用于 BatchNorm 层。
      • has_se:判断是否使用 SE 模块。
      • id_skip:判断是否使用跳跃连接(skip connection)。
    • 卷积层定义:

      1. 扩展卷积层 _expand_conv(Inverted Bottleneck):
        • 当扩展系数 expand_ratio 不等于 1 时,使用 1x1 卷积将通道数从 inp 扩展到 oup,其中 oup = inp * expand_ratio。
        • 后接批归一化层 _bn0 和激活函数(Swish)。
      2. 深度可分离卷积层 _depthwise_conv:
        • 使用深度可分离卷积,实现计算高效的卷积操作。
        • 卷积核大小为 k,步幅为 s,分组数等于通道数 oup。
        • 后接批归一化层 _bn1 和激活函数。
      3. SE 模块(如果使用的话):
        • 包含一个全局平均池化,两个 1x1 卷积层 _se_reduce 和 _se_expand,以及 Swish 激活函数和 Sigmoid 函数,实现通道上的注意力机制。
      4. 投影卷积层 _project_conv:
        • 使用 1x1 卷积将通道数从 oup 投影到目标输出通道数 final_oup。
        • 后接批归一化层 _bn2。
    • 激活函数:

      • 使用内存高效的 Swish 激活函数 MemoryEfficientSwish。

    前向传播方法 forward

    • 流程:
      1. 扩展阶段:
        • 如果 expand_ratio 不等于 1,首先通过 _expand_conv 扩展通道数,后接批归一化和激活函数。
      2. 深度可分离卷积:
        • 输入数据通过 _depthwise_conv 进行空间上的卷积操作,后接批归一化和激活函数。
      3. SE 模块(如果使用):
        • 通过全局平均池化将特征图变为 1x1 大小。
        • 经过 _se_reduce 和 _se_expand 两个 1x1 卷积层,形成通道注意力权重。
        • 使用 Sigmoid 激活函数生成权重,元素级乘法作用在特征图上,实现通道重标定。
      4. 投影阶段:
        • 通过 _project_conv 将通道数投影到目标输出通道数,后接批归一化。
      5. 跳跃连接和 Drop Connect:
        • 如果满足条件(使用跳跃连接,步幅为 1,输入和输出通道数相同),则进行残差连接。
        • 如果设置了 drop_connect_rate,则在残差连接前应用 Drop Connect。
      6. 输出:
        • 返回经过上述操作后的张量。

    方法 set_swish

    • 用于设置 Swish 激活函数是否使用内存高效的版本(训练时)或标准版本(导出或推理时)。

    MetaEfficientNet 类

    该类是整个 EfficientNet 模型的实现,包含了模型的各个部分:Stem、Blocks、Head 以及分类器。

    初始化方法 __init__

    • 参数初始化:

      • blocks_args:包含每个 MBConvBlock 的配置信息的列表。
      • global_params:包含全局配置参数。
    • 批归一化参数:

      • 计算批归一化层的动量和 epsilon 值。
    • 输入卷积层(Stem):

      • 使用 3x3 卷积,将输入通道数(通常为 3)映射到设定的通道数(通常为 32),步幅为 2,实现初始的降采样。
      • 后接批归一化层 _bn0 和激活函数。
    • 构建 Blocks:

      • 遍历 blocks_args,根据每个块的配置,创建对应数量的 MetaMBConvBlock。
      • 如果某个块需要重复(num_repeat 大于 1),除了第一个块外,其余重复的块需要将步幅修改为 1,以保持特征图尺寸不变。
    • 头部卷积层(Head):

      • 使用 1x1 卷积,将最后一个块的输出通道数映射到固定的通道数(通常为 1280)。
      • 后接批归一化层 _bn1 和激活函数。
    • 分类器:

      • 使用自适应平均池化将特征图大小缩放到 1x1。
      • 展平后通过 Dropout 层。
      • 最后通过全连接层 _fc,输出类别数的 logits。
    • 激活函数:

      • 同样使用内存高效的 Swish 激活函数。

    方法 set_swish

    • 用于设置模型中所有 Swish 激活函数是否使用内存高效的版本。

    方法 extract_features

    • 实现特征提取部分的前向传播。
    • 输入经过 Stem 卷积层、批归一化和激活函数。
    • 依次通过所有的 Blocks(MBConvBlock)。
    • 经过 Head 卷积层、批归一化和激活函数。
    • 返回最终的特征图。

    前向传播方法 forward

    • 输入经过特征提取部分,得到特征图 feature。
    • 经过自适应平均池化,将特征图缩放到 1x1。
    • 展平后通过 Dropout 层。
    • 经过全连接层 _fc,得到分类结果 x。
    • 同时返回分类结果 x 和特征图 feature。

    类方法 from_name

    • 根据给定的模型名称(如 "efficientnet-b0"),创建模型实例。
    • 获取对应的 blocks_args 和 global_params。
    • 可接受一些覆盖参数 override_params,用于修改模型的默认配置。
    • 调用私有方法 _change_in_channels,根据输入通道数调整模型的输入层。

    类方法 from_pretrained

    • 创建预训练模型的实例。
    • 可以指定预训练权重的路径、输入通道数、类别数等。
    • 如果类别数不是 1000,需要重新初始化全连接层。
    • 调用 _change_in_channels 方法调整输入通道数。

    方法 _change_in_channels

    • 这是您在代码末尾要求详细解释的部分。

    • 功能:

      • 当输入通道数 in_channels 不等于 3 时,调整模型的输入卷积层 _conv_stem,使其适应新的输入通道数。
    • 代码解析:

      def _change_in_channels(self, in_channels):
          if in_channels != 3:
              Conv2d = get_same_padding_conv2d(image_size = self._global_params.image_size)
              out_channels = round_filters(32, self._global_params)
              self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)
      
      • 步骤说明:
        1. 判断输入通道数是否不等于 3:

          • 如果输入通道数不是默认的 3(即非 RGB 图像,如灰度图像或多光谱图像),需要调整输入层。
        2. 获取合适的卷积层构造函数 Conv2d:

          • 使用函数 get_same_padding_conv2d,根据模型的输入图像尺寸 image_size,获取一个具有适当填充方式的卷积层类或函数。
        3. 计算输出通道数 out_channels:

          • 使用函数 round_filters 将基准通道数 32 依据模型的宽度系数进行调整,得到实际的输出通道数。
        4. 重新定义输入卷积层 _conv_stem:

          • 使用新的 in_channels(用户指定的输入通道数)和计算得到的 out_channels,创建一个新的卷积层。
          • 卷积核大小为 3x3,步幅为 2,不使用偏置。
          • 这一步替换了原有的 _conv_stem 层,使模型能够处理具有不同通道数的输入。
    • 作用:

      • 通过这个方法,模型可以适应不同类型的输入数据,例如:
        • 单通道的灰度图像(in_channels=1)。
        • 多通道的遥感图像或医学图像(in_channels>3)。
      • 这在迁移学习或特定领域的应用中非常有用。

    总结

    这段代码实现了一个可定制的 EfficientNet 模型,允许用户根据需要调整输入通道数、类别数等参数。通过使用元学习模块(MetaModule 等),模型可以方便地应用于元学习场景。


    如果您有任何关于代码的具体问题,或者需要进一步的解释,请随时提问!