为什么我的 IoU 在使用 tensorflow / keras 的训练中不断减少？

Question

我正在为语义分割训练一个类似于 U-net 的模型，但 IoU 在一个时期之后不断减少时期。 这是我的 IoU 和 IoU 损失函数。我的输入和输出掩码是一个带有dtype=np.bool 的numpy 数组，所以我将它转换为float32 以计算IoU。 我不知道是什么问题？我的指标函数或我的模型。我真的需要有人帮助我。

def iou(y_true, y_pred):
    y_true = tf.keras.backend.flatten(y_true)
    y_pred = tf.keras.backend.flatten(y_pred)
    y_true_f = tf.cast(y_true, tf.float32)
    y_pred_f = tf.cast(y_pred, tf.float32)
    intersection = tf.keras.backend.sum(y_true_f * y_pred_f)
    union = tf.keras.backend.sum(y_true_f) + tf.keras.backend.sum(y_pred_f) - intersection
    return (intersection + 1e-7) / (union + 1e-7)

def iou_loss(y_true, y_pred):
    return 1.0 - iou(y_true, y_pred)

# Compile model
metrics = [iou_loss, iou, 'accuracy']
model.compile(optimizer=Adam(learning_rate), loss=iou, metrics=[metrics], run_eagerly=True)

这是我的训练结果

Epoch 2/100
34/34 [==============================] - 3s 89ms/step - loss: 0.0186 - iou_loss: 0.9814 - iou: 0.0186 - accuracy: 0.9022 - val_loss: 0.0358 - val_iou_loss: 0.9647 - val_iou: 0.0353 - val_accuracy: 0.9460

Epoch 00002: val_loss improved from 0.03619 to 0.03579, saving model to /content/gdrive/MyDrive/model_ccnet_iris.h5
Epoch 3/100
34/34 [==============================] - 3s 89ms/step - loss: 0.0158 - iou_loss: 0.9843 - iou: 0.0157 - accuracy: 0.8972 - val_loss: 0.0352 - val_iou_loss: 0.9652 - val_iou: 0.0348 - val_accuracy: 0.9071

Epoch 00003: val_loss improved from 0.03579 to 0.03525, saving model to /content/gdrive/MyDrive/model_ccnet_iris.h5
Epoch 4/100
34/34 [==============================] - 3s 88ms/step - loss: 0.0132 - iou_loss: 0.9868 - iou: 0.0132 - accuracy: 0.8910 - val_loss: 0.0348 - val_iou_loss: 0.9656 - val_iou: 0.0344 - val_accuracy: 0.8690

Epoch 00004: val_loss improved from 0.03525 to 0.03485, saving model to /content/gdrive/MyDrive/model_ccnet_iris.h5
Epoch 5/100
34/34 [==============================] - 3s 87ms/step - loss: 0.0112 - iou_loss: 0.9888 - iou: 0.0112 - accuracy: 0.8842 - val_loss: 0.0345 - val_iou_loss: 0.9659 - val_iou: 0.0341 - val_accuracy: 0.8411

Epoch 00005: val_loss improved from 0.03485 to 0.03455, saving model to /content/gdrive/MyDrive/model_ccnet_iris.h5
Epoch 6/100
34/34 [==============================] - 3s 85ms/step - loss: 0.0096 - iou_loss: 0.9904 - iou: 0.0096 - accuracy: 0.8740 - val_loss: 0.0343 - val_iou_loss: 0.9662 - val_iou: 0.0338 - val_accuracy: 0.8216

Answer 1

优化器的功能是最小化损失函数
您将IoU 设置为损失函数，这就是它减少的原因。