Unet语义分割模型训练自定义过程中出现dice score趋近于0或者负值的解决方案（使用的仓库代码为——https://github.com/milesial/Pytorch-UNet）

使用仓库代码如下的时候：milesial/Pytorch-UNet: PyTorch implementation of the U-Net for image semantic segmentation with high quality images (github.com)https://github.com/milesial/Pytorch-UNet发现dice值无论和validation Dice始终为负值和趋近于（无限接近于0）0，具体情况如下图所示。

学习率为 learning_rate: float = 1e-5

对应的Val Dice值几乎无限趋近于0

修改学习率以后，各项Dice值正常了

参考——Dice coefficient no change during training,is always very close to 0 · Issue #173 · milesial/Pytorch-UNet (github.com)https://github.com/milesial/Pytorch-UNet/issues/173

因此解决方案为：

将train.py代码中的内容

def train_model(
        model,
        device,
        epochs: int = 8,
        batch_size: int = 16,
        learning_rate: float = 1e-5,
        val_percent: float = 0.1,
        save_checkpoint: bool = True,
        img_scale: float = 0.5,
        amp: bool = False,
        weight_decay: float = 1e-8,
        momentum: float = 0.999,
        gradient_clipping: float = 1.0,
):

改为

def train_model(
        model,
        device,
        epochs: int = 8,
        batch_size: int = 16,
        learning_rate: float =  0.0001,
        val_percent: float = 0.1,
        save_checkpoint: bool = True,
        img_scale: float = 0.5,
        amp: bool = False,
        weight_decay: float = 1e-8,
        momentum: float = 0.999,
        gradient_clipping: float = 1.0,
):

 learning_rate: float = 1e-5 ——>修改为——> learning_rate: float =  0.0001

再次尝试运行就可以正常输出得到正常的各项Dice值了

tips:一旦成功得到各项指标以后，再把学习率改回去，可能就无法复现之前的“趋近于0”的现象了。