02-model-soup.py

import torch
from timm import create_model

# 加载模型
def load_model(model_path):
    model = create_model('tf_efficientnet_b3_ns', num_classes=1605, pretrained=False)
    model.load_state_dict(torch.load(model_path))
    return model

# 模型融合
def model_soup(models, weights):
    if len(models) != len(weights):
        raise ValueError("Number of models and weights must match")
    
    # Normalize weights
    weights = [w / sum(weights) for w in weights]
    
    # Initialize the fused model with the structure of the first model
    fused_model = create_model('tf_efficientnet_b3_ns', num_classes=1605, pretrained=False)
    fused_model_dict = fused_model.state_dict()
    
    for key in fused_model_dict.keys():
        fused_model_dict[key] = sum(weight * models[i].state_dict()[key] for i, weight in enumerate(weights))
    
    fused_model.load_state_dict(fused_model_dict)
    return fused_model

# 模型权重路径
model_paths = [
    '/data/cjm/FungiCLEF2024/EfficientNet/output/trick_1.4.3/efficientnet_b3_epoch_28.pth',  # 权重 4
    '/data/cjm/FungiCLEF2024/EfficientNet/output/trick_1.4.3.2/efficientnet_b3_epoch_28.pth',  # 权重 2.6
    '/data/cjm/FungiCLEF2024/EfficientNet/output/trick_1.4.1/efficientnet_b3_epoch_23.pth',  # 权重 2.4
    '/data/cjm/FungiCLEF2024/EfficientNet/output/trick_1.5.2/efficientnet_b3_epoch_21.pth', # 权重 1
]

# 加载模型
models = [load_model(path) for path in model_paths]

# 模型权重
weights = [4, 2.6, 2.4, 1]

# 进行模型融合
fused_model = model_soup(models, weights)

# 保存融合后的模型权重
fused_model_path = '/data/cjm/FungiCLEF2024/EfficientNet/output/fused_model_soup.pth'
torch.save(fused_model.state_dict(), fused_model_path)

print(f"Fused model saved to {fused_model_path}")