modeling_my_segformer.py

import torch
import torch.nn as nn
from transformers import PreTrainedModel

from segformer_plusplus.model.backbone.mit import MixVisionTransformer
from mix_vision_transformer_config import MySegformerConfig
from segformer_plusplus.model.head.segformer_head import SegformerHead


class MySegformerForSemanticSegmentation(PreTrainedModel):
    config_class = MySegformerConfig
    base_model_prefix = "my_segformer"

    def __init__(self, config):
        super().__init__(config)

        # Backbone (MixVisionTransformer)
        self.backbone = MixVisionTransformer(
            embed_dims=config.embed_dims[0],
            num_stages=config.num_stages,
            num_layers=config.num_layers,
            num_heads=config.num_heads,
            patch_sizes=config.patch_sizes,
            strides=config.strides,
            sr_ratios=config.sr_ratios,
            mlp_ratio=config.mlp_ratio,
            qkv_bias=config.qkv_bias,
            drop_rate=config.drop_rate,
            attn_drop_rate=config.attn_drop_rate,
            drop_path_rate=config.drop_path_rate,
            out_indices=config.out_indices
        )

        # Head initialization
        in_channels = config.decode_head["in_channels"]

        self.segmentation_head = SegformerHead(
            in_channels=in_channels,
            in_index=list(config.out_indices),
            channels=config.decode_head["channels"],
            dropout_ratio=config.decode_head["dropout_ratio"],
            align_corners=config.decode_head["align_corners"],
            interpolate_mode=config.decode_head["interpolate_mode"],
            use_conv_bias_in_convmodules=False
        )

        self.post_init()

    def forward(self, x):
        features = self.backbone(x)

        # Debug: Ausgabe der Shapes der Backbone-Features
        for i, f in enumerate(features):
            print(f"Feature {i}: shape = {f.shape}")

        logits = self.segmentation_head(features)

        return logits