如何在Tensorflow2.x 2.x中实现自定义特定度量？

Question

我正在尝试实现一个自定义度量，为我的语义分割问题计算特异性。当我用这个度量来训练模型的时候，我不断地得到这个错误。当模型显示培训度量时，此度量很好地工作，但对于验证度量则停止，如下所示。

我的实施情况如下：

class Specificity(tf.keras.metrics.Metric):
def __init__(self, **kwargs):
    super().__init__(**kwargs)
    self.tn = tf.keras.metrics.TrueNegatives()
    self.fp = tf.keras.metrics.FalsePositives()

def update_state(self, y_true, y_pred, sample_weight=None):
    self.tn.update_state(y_true, y_pred)
    self.fp.update_state(y_true, y_pred)
    
def result(self):
    tn = self.tn.result()
    fp = self.fp.result()
    return tf.expand_dims(tf.divide(tn, tn + fp), axis=-1)

输出信息：

Epoch 1/200
2022-07-26 10:16:25.238678: I tensorflow/stream_executor/cuda/cuda_dnn.cc:384] Loaded cuDNN version 8302
2022-07-26 10:16:26.083645: I tensorflow/core/platform/default/subprocess.cc:304] Start cannot spawn child process: No such file or directory
100/100 [==============================] - ETA: 0s - loss: 0.7601 - iou_score: 0.5211 - precision: 0.6069 - recall: 0.7806 - f1-score: 0.6732 - auc: 0.9403 - specificity: 0.8967Traceback (most recent call last):
  File "/home/bhattrai/corneal_neovascularization_tf2/test_statistics.py", line 892, in <module>
    history = model.fit(my_generator, validation_data=validation_datagen, steps_per_epoch=100, validation_steps=100,
  File "/home/bhattrai/.local/lib/python3.9/site-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/home/bhattrai/.local/lib/python3.9/site-packages/keras/backend.py", line 4028, in batch_set_value
    x.assign(np.asarray(value, dtype=dtype_numpy(x)))
ValueError: Cannot assign value to variable ' accumulator:0': Shape mismatch.The variable shape (1,), and the assigned value shape () are incompatible.
2022-07-26 10:16:39.142482: W tensorflow/core/kernels/data/generator_dataset_op.cc:108] Error occurred when finalizing GeneratorDataset iterator: FAILED_PRECONDITION: Python interpreter state is not initialized. The process may be terminated.
         [[{{node PyFunc}}]]

我试着用另一种方式实现它，如下所示，但是当我在模型培训期间使用它时，它总是显示0.000。

import tensorflow as tf

class Specificity(tf.keras.metrics.Metric):
    def __init__(self, name='specificity', **kwargs):
        super().__init__(name=name, **kwargs)
        self.tn = self.add_weight(name='tn', initializer='zeros')
        self.fp = self.add_weight(name='fp', initializer='zeros')
        self.tnr = self.add_weight(name='tnr', initializer='zeros')

    def update_state(self, y_true, y_pred, sample_weight=None):
        y_true = tf.cast(y_true, tf.bool)
        y_pred = tf.cast(y_pred, tf.bool)

        # Getting the True Negatives
        values = tf.logical_and(tf.equal(y_true, False), tf.equal(y_pred, False))
        values = tf.cast(values, self.dtype)
        self.tn.assign_add(tf.reduce_sum(values))

        # Getting the False Positives
        values = tf.logical_and(tf.equal(y_true, False), tf.equal(y_pred, True))
        values = tf.cast(values, self.dtype)
        self.fp.assign_add(tf.reduce_sum(values))

        self.tnr.assign_add(tf.divide(self.tn, tf.add(self.tn, self.fp)))

    def result(self):
        return self.tnr

输出：

m = Specificity()
m.update_state([0, 1, 0, 0], [0, 1, 0, 0])
X = m.result().numpy()
1.0

但当我用它训练模特时：

X = tf.random.normal(shape=(100, 256, 256, 3))
Y = tf.random.uniform(minval=0, maxval=2, shape=(100, 256, 256, 1), dtype=tf.int32)
dataset = tf.data.Dataset.from_tensor_slices((X,Y))
train = dataset.take(80)
val = dataset.skip(80)
train = train.cache().shuffle(1000).batch(32).prefetch(tf.data.experimental.AUTOTUNE)
val = val.cache().shuffle(1000).batch(32).prefetch(tf.data.experimental.AUTOTUNE)

model = tf.keras.Sequential([
                         tf.keras.layers.Input(shape=(256,256,3)),
                         tf.keras.layers.Conv2D(16, 3, padding='same'),
                         tf.keras.layers.MaxPooling2D(),
                         tf.keras.layers.Conv2D(32, 3, padding='same'),
                         tf.keras.layers.MaxPooling2D(),
                         tf.keras.layers.Conv2D(64, 3, padding='same'),
                         tf.keras.layers.MaxPooling2D(),
                         tf.keras.layers.Conv2DTranspose(64, strides=2, padding='same', kernel_size=3),
                         tf.keras.layers.Conv2DTranspose(32, strides=2, padding='same', kernel_size=3),
                         tf.keras.layers.Conv2DTranspose(32, strides=2, padding='same', kernel_size=3),
                         tf.keras.layers.Conv2D(1, 3, activation='sigmoid', padding="same")
])

model.compile(loss=tf.keras.losses.BinaryCrossentropy(),
          optimizer='adam',
          metrics=[Specificity(), 'accuracy'])

model.fit(train, validation_data=val, epochs=2)

Epoch 1/2
3/3 [==============================] - 16s 5s/step - loss: 0.6939 - specificity: 0.0000e+00 - accuracy: 0.5003 - val_loss: 0.6934 - val_specificity: 0.0000e+00 - val_accuracy: 0.4999
Epoch 2/2
3/3 [==============================] - 15s 5s/step - loss: 0.6933 - specificity: 0.0000e+00 - accuracy: 0.5004 - val_loss: 0.6933 - val_specificity: 0.0000e+00 - val_accuracy: 0.5002
<keras.callbacks.History at 0x7f9fdb2c1250>

Answer 1

我现在还不能测试它，但是我总是将自定义度量定义为函数(而不是子类)。也许你可以试试。另外，在这个文章中，我发现了一个具体性度量的实现，也许您可以根据您的需要对它进行调整。

这是代码：

import numpy as np
import tensorflow as tf
from keras import backend as K

def specificity(y_true, y_pred):
    tn = K.sum(K.round(K.clip((1 - y_true) * (1 - y_pred), 0, 1)))
    fp = K.sum(K.round(K.clip((1 - y_true) * y_pred, 0, 1)))
    return tn / (tn + fp + K.epsilon())

model = tf.keras.Sequential(...)
model.compile(
    optimizer="adam",
    loss=tf.keras.losses.BinaryCrossentropy(),
    metrics=[
        "accuracy",
        specificity
    ]
)
model.fit(train, validation_data=val, epochs=2)

更新：

在您的子类版本中，问题可能与缺少reset_states函数有关(在每个时代之后清除状态)。尝试将以下内容添加到您的课堂中：

def reset_states(self):
    self.self.tn.assign(0)
    self.self.tp.assign(0)
    self.self.tnr.assign(0)