Mask2Former Code Dataflow

1. Entrance
    Mask2Former/mask2former/maskformer_model.py
    features = self.backbone(images.tensor)
    outputs = self.sem_seg_head(features)
    
2.1
    Mask2Former/mask2former/modeling/meta_arch/mask_former_head.py
    mask_features, transformer_encoder_features, multi_scale_features = self.pixel_decoder.forward_features(features)
    predictions = self.predictor(multi_scale_features, mask_features, mask)
    
2.2    self.pixel_decoder.forward_features(features)
    Mask2Former/mask2former/modeling/pixel_decoder/msdeformattn.py
    def forward_features(self, features): # refer to 2.
        y, spatial_shapes, level_start_index = self.transformer(srcs, pos)
        
2.2.1
    Mask2Former/mask2former/modeling/pixel_decoder/msdeformattn.py
    self.transformer = MSDeformAttnTransformerEncoderOnly(
    
    
2.3    predictions = self.predictor(multi_scale_features, mask_features, mask)
    Mask2Former/mask2former/modeling/transformer_decoder/mask2former_transformer_decoder.py
    @TRANSFORMER_DECODER_REGISTRY.register()
    class MultiScaleMaskedTransformerDecoder(nn.Module):
        def forward(self, x, mask_features, mask = None):
            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