Python 训练uNet模型预测结果仅为黑色_Python_Python 3.x_Keras_Conv Neural Network_Training Data

Python 训练uNet模型预测结果仅为黑色

python python-3.x keras

Python 训练uNet模型预测结果仅为黑色,python,python-3.x,keras,conv-neural-network,training-data,Python,Python 3.x,Keras,Conv Neural Network,Training Data,我正在训练一个用于细分的uNet模型。训练模型后，输出都是零，我不明白为什么我看到建议我应该使用一个特定的损失函数，所以我使用了骰子损失函数。这是因为黑色区域（0）比白色区域（1）大得多我做错什么了吗我的型号是： Layer (type) Output Shape Param # Connected to ==============================================================

我正在训练一个用于细分的uNet模型。训练模型后，输出都是零，我不明白为什么

我看到建议我应该使用一个特定的损失函数，所以我使用了骰子损失函数。这是因为黑色区域（0）比白色区域（1）大得多

我做错什么了吗

我的型号是：

Layer (type) Output Shape Param # Connected to ================================================================================================== input_1 (InputLayer) (None, 80, 80, 1) 0 __________________________________________________________________________________________________ conv2d_1 (Conv2D) (None, 80, 80, 64) 640 input_1[0][0] __________________________________________________________________________________________________ conv2d_2 (Conv2D) (None, 80, 80, 64) 36928 conv2d_1[0][0] __________________________________________________________________________________________________ max_pooling2d_1 (MaxPooling2D) (None, 40, 40, 64) 0 conv2d_2[0][0] __________________________________________________________________________________________________ conv2d_3 (Conv2D) (None, 40, 40, 128) 73856 max_pooling2d_1[0][0] __________________________________________________________________________________________________ conv2d_4 (Conv2D) (None, 40, 40, 128) 147584 conv2d_3[0][0] __________________________________________________________________________________________________ max_pooling2d_2 (MaxPooling2D) (None, 20, 20, 128) 0 conv2d_4[0][0] __________________________________________________________________________________________________ conv2d_5 (Conv2D) (None, 20, 20, 256) 295168 max_pooling2d_2[0][0] __________________________________________________________________________________________________ conv2d_6 (Conv2D) (None, 20, 20, 256) 590080 conv2d_5[0][0] __________________________________________________________________________________________________ max_pooling2d_3 (MaxPooling2D) (None, 10, 10, 256) 0 conv2d_6[0][0] __________________________________________________________________________________________________ conv2d_7 (Conv2D) (None, 10, 10, 512) 1180160 max_pooling2d_3[0][0] __________________________________________________________________________________________________ conv2d_8 (Conv2D) (None, 10, 10, 512) 2359808 conv2d_7[0][0] __________________________________________________________________________________________________ dropout_1 (Dropout) (None, 10, 10, 512) 0 conv2d_8[0][0] __________________________________________________________________________________________________ max_pooling2d_4 (MaxPooling2D) (None, 5, 5, 512) 0 dropout_1[0][0] __________________________________________________________________________________________________ conv2d_9 (Conv2D) (None, 5, 5, 1024) 4719616 max_pooling2d_4[0][0] __________________________________________________________________________________________________ conv2d_10 (Conv2D) (None, 5, 5, 1024) 9438208 conv2d_9[0][0] __________________________________________________________________________________________________ dropout_2 (Dropout) (None, 5, 5, 1024) 0 conv2d_10[0][0] __________________________________________________________________________________________________ conv2d_transpose_1 (Conv2DTrans (None, 10, 10, 512) 2097664 dropout_2[0][0] __________________________________________________________________________________________________ concatenate_1 (Concatenate) (None, 10, 10, 1024) 0 dropout_1[0][0] conv2d_transpose_1[0][0] __________________________________________________________________________________________________ conv2d_11 (Conv2D) (None, 10, 10, 512) 4719104 concatenate_1[0][0] __________________________________________________________________________________________________ conv2d_12 (Conv2D) (None, 10, 10, 512) 2359808 conv2d_11[0][0] __________________________________________________________________________________________________ conv2d_transpose_2 (Conv2DTrans (None, 20, 20, 256) 524544 conv2d_12[0][0] __________________________________________________________________________________________________ concatenate_2 (Concatenate) (None, 20, 20, 512) 0 conv2d_6[0][0] conv2d_transpose_2[0][0] __________________________________________________________________________________________________ conv2d_13 (Conv2D) (None, 20, 20, 256) 1179904 concatenate_2[0][0] __________________________________________________________________________________________________ conv2d_14 (Conv2D) (None, 20, 20, 256) 590080 conv2d_13[0][0] __________________________________________________________________________________________________ conv2d_transpose_3 (Conv2DTrans (None, 40, 40, 128) 131200 conv2d_14[0][0] __________________________________________________________________________________________________ concatenate_3 (Concatenate) (None, 40, 40, 256) 0 conv2d_4[0][0] conv2d_transpose_3[0][0] __________________________________________________________________________________________________ conv2d_15 (Conv2D) (None, 40, 40, 128) 295040 concatenate_3[0][0] __________________________________________________________________________________________________ conv2d_16 (Conv2D) (None, 40, 40, 128) 147584 conv2d_15[0][0] __________________________________________________________________________________________________ conv2d_transpose_4 (Conv2DTrans (None, 80, 80, 64) 32832 conv2d_16[0][0] __________________________________________________________________________________________________ concatenate_4 (Concatenate) (None, 80, 80, 128) 0 conv2d_2[0][0] conv2d_transpose_4[0][0] __________________________________________________________________________________________________ conv2d_17 (Conv2D) (None, 80, 80, 64) 73792 concatenate_4[0][0] __________________________________________________________________________________________________ conv2d_18 (Conv2D) (None, 80, 80, 64) 36928 conv2d_17[0][0] __________________________________________________________________________________________________ conv2d_19 (Conv2D) (None, 80, 80, 2) 1154 conv2d_18[0][0] __________________________________________________________________________________________________ conv2d_20 (Conv2D) (None, 80, 80, 1) 3 conv2d_19[0][0] ==================================================================================================
损失函数

def dice_loss_v2（y_true，y_pred）：分子=2*tf.减少总和（y_真*y_pred，轴=（1,2,3））分母=tf.减少总和（y_真+y_pred，轴=（1,2,3））返回1-分子/分母
激活

model.compile（优化器='adam'，损失=骰子损失v2，指标=[“准确度”，iou损失（核心]）
预定义的学习率为LR=0.001
额外信息：

datagen=ImageDataGenerator( 旋转范围=10，宽度\偏移\范围=0.1，高度位移范围=0.1，缩放（范围=0.1）数据发生器安装（X_系列）模型拟合发生器（数据生成流（X列，y列，批量大小=100），每历元步长=len（X列）， epochs=4，验证\数据=（X \ u检验，y \ u检验）
可能的原因之一是你的面具。可能是将背景视为遮罩中的目标对象。这就是为什么它正在学习检测0（黑色）的背景。检查你的面具。我遇到了同样的问题，然后我检查了我的面具。我猜不是。你可以看到，上面是输入，第二个是输入的掩码，预测是这样的…你可以这样做：为了改善结果，你可以做以下事情：1）使用相册进行数据扩充2）创建自己的自定义数据集并添加预处理（标准化/标准化图像）加快收敛速度。3）将焦点损失用于此背景-前景不平衡。4）使用预训练编码器。我认为你是在增加你的输入，而不是你的掩码，可能会有所帮助