src/models/__init__.py

from .efficientnet import EfficientNetClassifier
from .resnet import ResNetClassifier
from .vit import VisionTransformerClassifier
from .mobilenetv3 import MobileNetV3SmallClassifier
from .shufflenetv2 import ShuffleNetV2Classifier

import torch
import torch.nn as nn


def create_model(config: dict) -> nn.Module:
    """根据配置创建模型实例。"""
    model_config = config['model']
    model_name = model_config['name']
    num_classes = model_config['num_classes']
    pretrained = model_config.get('pretrained', True)
    dropout = model_config.get('dropout', 0.2)

    model_name_lower = model_name.lower()


    if 'efficientnet' in model_name_lower:
        return EfficientNetClassifier(model_name, num_classes, pretrained, dropout)
    if 'resnet' in model_name_lower:
        return ResNetClassifier(model_name, num_classes, pretrained, dropout)
    if 'vit' in model_name_lower:
        return VisionTransformerClassifier(model_name, num_classes, pretrained, dropout)
    if 'mobilenetv3' in model_name_lower:
        return MobileNetV3SmallClassifier(num_classes=num_classes, pretrained=pretrained, dropout=dropout)
    if 'shufflenetv2' in model_name_lower:
        return ShuffleNetV2Classifier(num_classes=num_classes, pretrained=pretrained, dropout=dropout)

    raise ValueError(f'不支持的模型类型: {model_name}')


def count_parameters(model: nn.Module) -> int:
    """计算可训练参数数量。"""
    return sum(parameter.numel() for parameter in model.parameters() if parameter.requires_grad)


def model_size_mb(model: nn.Module) -> float:
    """估算模型参数与缓冲区占用的内存大小（MB）。"""
    param_size = 0
    buffer_size = 0

    for parameter in model.parameters():
        param_size += parameter.nelement() * parameter.element_size()

    for buffer in model.buffers():
        buffer_size += buffer.nelement() * buffer.element_size()

    return (param_size + buffer_size) / 1024 / 1024