一种通用的图像分割模型

    def forward_features(self, features):
        srcs = []
        pos = []
        # Reverse feature maps into top-down order (from low to high resolution)
        # 将其通道维数全部转变为256
        for idx, f in enumerate(self.transformer_in_features[::-1]):
            x = features[f].float()  # deformable detr does not support half precision
            srcs.append(self.input_proj[idx](x))
            pos.append(self.pe_layer(x)) # 存放有关像素分辨率的sine位置编码不可学习
        # y: [1,43008,256] 将不同大小的特征图进行拼接
        y, spatial_shapes, level_start_index = self.transformer(srcs, pos)
        bs = y.shape[0]

        split_size_or_sections = [None] * self.transformer_num_feature_levels
        
        for i in range(self.transformer_num_feature_levels):
            if i < self.transformer_num_feature_levels - 1:
                split_size_or_sections[i] = level_start_index[i + 1] - level_start_index[i]
            else:
                split_size_or_sections[i] = y.shape[1] - level_start_index[i]
        y = torch.split(y, split_size_or_sections, dim=1)

        out = []
        multi_scale_features = []
        num_cur_levels = 0
        for i, z in enumerate(y):
            # z:[1,2048,256]->[1,256,2048]->[1,256,32,64]
            out.append(z.transpose(1, 2).view(bs, -1, spatial_shapes[i][0], spatial_shapes[i][1]))
        # append `out` with extra FPN levels
        # Reverse feature maps into top-down order (from low to high resolution)
        for idx, f in enumerate(self.in_features[:self.num_fpn_levels][::-1]):
            x = features[f].float()
            # x: [1,256,256,512] cur_fpn:[1,256,256,512]
            lateral_conv = self.lateral_convs[idx]
            output_conv = self.output_convs[idx]
            cur_fpn = lateral_conv(x)
            # Following FPN implementation, we use nearest upsampling here
            # y:[1,256,256,512] 将最后一个128,256进行拼接
            y = cur_fpn + F.interpolate(out[-1], size=cur_fpn.shape[-2:], mode="bilinear", align_corners=False)
            y = output_conv(y)
            out.append(y)
        # 此时将3个pixel从低到高分辨率的特征图像提取出来了
        for o in out:
            if num_cur_levels < self.maskformer_num_feature_levels:
                multi_scale_features.append(o)
                num_cur_levels += 1
        # 用最大的的特征图来确定是否是模型所关心的前景区域，out[0]表示第一个特征图的情况
        # multi_scale_features表示不同的特征
        return self.mask_features(out[-1]), out[0], multi_scale_features

def forward(self, x, mask_features, mask = None):
        # x is a list of multi-scale feature
        assert len(x) == self.num_feature_levels
        src = []
        pos = []
        size_list = []

        # disable mask, it does not affect performance
        del mask

        for i in range(self.num_feature_levels):
            # 获取不同特征图的大小
            size_list.append(x[i].shape[-2:])
            # 将不同的特征图添加位置编码，并且展平处理 [1,256,2048]
            pos.append(self.pe_layer(x[i], None).flatten(2))
            # 将图像通道统一，并且获得不同的尺度可学习编码
            src.append(self.input_proj[i](x[i]).flatten(2) + self.level_embed.weight[i][None, :, None])

            # flatten NxCxHxW to HWxNxC
            pos[-1] = pos[-1].permute(2, 0, 1)
            src[-1] = src[-1].permute(2, 0, 1)

        _, bs, _ = src[0].shape

        # QxNxC [100,1,256] output作为查询提供信息
        query_embed = self.query_embed.weight.unsqueeze(1).repeat(1, bs, 1)
        output = self.query_feat.weight.unsqueeze(1).repeat(1, bs, 1)

        predictions_class = []
        predictions_mask = []

        # prediction heads on learnable query features
        # 使用注意力内的区域来进行预测
        outputs_class, outputs_mask, attn_mask = self.forward_prediction_heads(output, mask_features, attn_mask_target_size=size_list[0])
        predictions_class.append(outputs_class)
        predictions_mask.append(outputs_mask)

        # Decoder架构，先进行交叉注意力机制，之后进行自注意力机制，最后进行FFN结构
        for i in range(self.num_layers):
            level_index = i % self.num_feature_levels
            attn_mask[torch.where(attn_mask.sum(-1) == attn_mask.shape[-1])] = False
            # attention: cross-attention first
            output = self.transformer_cross_attention_layers[i](
                output, src[level_index],
                memory_mask=attn_mask,
                memory_key_padding_mask=None,  # here we do not apply masking on padded region
                pos=pos[level_index], query_pos=query_embed
            )

            output = self.transformer_self_attention_layers[i](
                output, tgt_mask=None,
                tgt_key_padding_mask=None,
                query_pos=query_embed
            )
            
            # FFN
            output = self.transformer_ffn_layers[i](
                output
            )

            outputs_class, outputs_mask, attn_mask = self.forward_prediction_heads(output, mask_features, attn_mask_target_size=size_list[(i + 1) % self.num_feature_levels])
            predictions_class.append(outputs_class)
            predictions_mask.append(outputs_mask)

        assert len(predictions_class) == self.num_layers + 1

        out = {
            'pred_logits': predictions_class[-1],
            'pred_masks': predictions_mask[-1],
            'aux_outputs': self._set_aux_loss(
                predictions_class if self.mask_classification else None, predictions_mask
            )
        }
        return out

conda create --name mask2former python=3.8 -y
conda activate mask2former
conda install pytorch==1.9.0 torchvision==0.10.0 cudatoolkit=11.1 -c pytorch -c nvidia
pip install -U opencv-python

# under your working directory
git clone git@github.com:facebookresearch/detectron2.git
cd detectron2
pip install -e .
pip install git+https://github.com/cocodataset/panopticapi.git
pip install git+https://github.com/mcordts/cityscapesScripts.git

cd ..
git clone git@github.com:facebookresearch/Mask2Former.git
cd Mask2Former
pip install -r requirements.txt
cd mask2former/modeling/pixel_decoder/ops
sh make.sh