Implementation:PeterL1n BackgroundMattingV2 Load pretrained deeplabv3 state dict

Knowledge Sources	BackgroundMattingV2 DeepLabV3Plus-Pytorch
Domains	Transfer_Learning, Computer_Vision
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete tool for loading pretrained DeepLabV3 weights into the matting model provided by model/model.py.

Description

The load_pretrained_deeplabv3_state_dict method on the Base class converts and loads pretrained DeepLabV3 state dict weights into the matting model's encoder, ASPP, and decoder modules. It remaps the ASPP key names (classifier.classifier.0 → aspp) and uses load_matched_state_dict from model/utils.py to only load weights with matching keys and shapes, silently skipping mismatched parameters. For MobileNetV2 backbones, it temporarily restructures the backbone attributes to match DeepLabV3's naming convention.

Usage

Call this method once during initial training to initialize from pretrained segmentation weights. Not needed when resuming from a matting checkpoint — use load_state_dict() instead.

Code Reference

Source Location

• Repository: BackgroundMattingV2
• File: model/model.py
• Lines: 37-58

Signature

class Base(nn.Module):
    def load_pretrained_deeplabv3_state_dict(
        self,
        state_dict: Dict[str, Tensor],
        print_stats: bool = True
    ) -> None:
        """
        Converts and loads pretrained DeepLabV3 state_dict to match model structure.

        Args:
            state_dict: Pretrained DeepLabV3 state dict, typically loaded via
                        torch.load(path)['model_state']
            print_stats: Whether to print match statistics (matched/total keys)
        """

Import

from model import MattingBase
# or
from model.model import Base

I/O Contract

Inputs

Name	Type	Required	Description
state_dict	Dict[str, Tensor]	Yes	Pretrained DeepLabV3 state dict from torch.load(path)['model_state']
print_stats	bool	No	Print matched/total key counts (default True)

Outputs

Name	Type	Description
(in-place)	None	Model weights are modified in-place; matched keys loaded, unmatched skipped

Usage Examples

Loading Pretrained DeepLabV3 Weights

import torch
from model import MattingBase

# 1. Create model
model = MattingBase(backbone='resnet50')

# 2. Load pretrained DeepLabV3 checkpoint
deeplabv3_checkpoint = torch.load('best_deeplabv3_resnet50_voc_os16.pth')

# 3. Transfer matching weights
model.load_pretrained_deeplabv3_state_dict(deeplabv3_checkpoint['model_state'])
# Prints: "Loaded state_dict: 267/308 matched"

Related Pages

Implements Principle

• Principle:PeterL1n_BackgroundMattingV2_Pretrained_weight_transfer