Keras 中 ADR 和 FAR 的自定义指标

Question

我需要在带有 tensorflow 后端的 keras 中定义一个自定义指标，以获取 ADR 和 FAR 指标，其中：

• ADR 是正确检测到的第 1 类元素数与第 1 类元素总数之比。

• FAR 是被错误分类为 1 类元素的 0 类元素的数量与 0 类元素的总数之间的比率。

目前，我已经能够计算出每个类的元素总数：

def false_rates(y_true, y_pred):
    total_1 = K.sum(tf.cast(tf.equal(y_true, 1), 'int32'))
    total_0 = K.sum(tf.cast(tf.equal(y_true, 0), 'int32'))
    # [...]

现在，对于张量中的每个索引i，我需要计算有多少：

• y_true[i] == 1 and y_pred[i] == 1 用于 ADR
• y_true[i] == 0 and y_pred[i] == 1 代表 FAR

但我不知道如何使用 tensorflow 操作来做到这一点，我得到的最接近的是：

adr = K.sum(
        tf.cast(
            tf.equal(
                tf.add(
                    tf.cast(tf.equal(y_true, 1), 'int32'),
                    tf.cast(tf.equal(y_pred, 1), 'int32')), 2), 'int32'))

# far = ...?

但它似乎只返回 0。

感谢您的帮助！

---

编辑以包含完整的代码：

# coding: utf-8
import csv
import numpy as np
import os

os.environ["KERAS_BACKEND"] = "tensorflow"

from keras import backend as K
from keras.callbacks import ModelCheckpoint, EarlyStopping, TensorBoard
from keras.layers import Convolution2D
from keras.models import Sequential
from keras.optimizers import RMSprop, SGD
from keras.layers.core import Dense, Dropout
from keras.layers.recurrent import LSTM
from keras.layers.embeddings import Embedding
from keras.preprocessing.sequence import pad_sequences
from keras.preprocessing.text import one_hot

def build_model(input_shape, n1, dropout_p, n_symbols, n_sequence):
    model = Sequential()
    model.add(Embedding(n_symbols+1,
                        output_dim=n1,
                        input_length=n_sequence))
    model.add(LSTM(n1))
    model.add(Dropout(dropout_p))
    model.add(Dense(1))
    return model

def load_sequences(path):
    X, y = [], []
    with open(path,'r') as file:
        reader = csv.reader(file)
        for line in reader:
                X.append(line[0])
                y.append(int(line[1]))
    return np.array(X), np.array(y)

def adr_and_far(y_true, y_pred):
    total_adr = K.sum(K.tf.cast(K.tf.equal(y_true, 1), 'int32'))
    total_far = K.sum(K.tf.cast(K.tf.equal(y_true, 0), 'int32'))

    adr_idx = K.tf.equal(y_true, 1) & K.tf.equal(y_pred, 1)
    adr_idx = K.tf.reshape(K.tf.where(adr_idx), [-1])

    far_idx = K.tf.equal(y_true, 0) & K.tf.equal(y_pred, 1)
    far_idx = K.tf.reshape(K.tf.where(far_idx), [-1])

    num_adr = K.tf.shape(adr_idx)[0]  # 3
    num_far = K.tf.shape(far_idx)[0]  # 2

    return { 'total_adr': total_adr,
             'adr': num_adr / total_adr,
             'total_far': total_far,
             'far': num_far / total_far
           }

simbols = ['1','a','A','r','2','b','B','s','3','c','C','t','4','d','D','u','5','e','E','v','6','f','F','w','7','g','G','x','8','h','H','y','9','i','I','z', ',', '.', '*', '+', '0']
n_symbols = len(simbols) # dimensionality of your word vectors

max_len_sequence = 20
input_shape = (max_len_sequence, n_symbols)
n1 = 128 # From Paper
dropout_p = 0.1
loss_method = 'binary_crossentropy'

optimizer = 'adam'
model = build_model(input_shape, n1, dropout_p, n_symbols, max_len_sequence)
model.compile(loss=loss_method, optimizer=optimizer, metrics=['accuracy', adr_and_far])

还有错误：

 Using TensorFlow backend.
Traceback (most recent call last):
  File "1%29+Training+RNN+Single+Model.py", line 68, in <module>
    model.compile(loss=loss_method, optimizer=optimizer, metrics=['accuracy', adr_and_far])
  File "/usr/local/lib/python2.7/site-packages/keras/models.py", line 594, in compile
    **kwargs)
  File "/usr/local/lib/python2.7/site-packages/keras/engine/training.py", line 716, in compile
    metric_result = metric_fn(y_true, y_pred)
  File "1%29+Training+RNN+Single+Model.py", line 43, in adr_and_far
    adr_idx = K.tf.reshape(K.tf.where(adr_idx), [-1])
TypeError: select() takes at least 3 arguments (1 given)

Answer 1

你只需要数数吗？

y_true = tf.constant([1, 1, 1, 0, 0, 0, 1, 1, 1])
y_pred = tf.constant([1, 1, 0, 1, 1, 0, 1, 0, 0])

adr_idx = tf.equal(y_true, y_pred) & tf.equal(y_true, 1)
adr_idx = tf.reshape(tf.where(adr_idx), [-1]) # [0, 1, 6]

far_idx = tf.equal(y_true, 0) & tf.equal(y_pred, 1) 
far_idx = tf.reshape(tf.where(far_idx), [-1])  # [3, 4]

num_adr = tf.shape(adr_idx)[0]  # 3
num_far = tf.shape(far_idx)[0]  # 2

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    a, b = sess.run([num_adr, num_far])

使用 keras 1.2.1 和 tensorflow 版本 0.11
我认为原因是形状不同
y_true 的形状是 (?, ?) 和 y_pred 的形状是 (?, 1) ？

def adr_and_far(y_true, y_pred):
    total_adr = K.sum(K.tf.cast(K.tf.equal(y_true, 1), 'int32'))
    total_far = K.sum(K.tf.cast(K.tf.equal(y_true, 0), 'int32'))

    adr_idx = K.tf.equal(y_true, 1) & K.tf.equal(y_pred, 1)
    adr_idx = K.sum(K.tf.cast(K.reshape(adr_idx, [-1]), 'int32'))

    far_idx = K.tf.equal(y_true, 0) & K.tf.equal(y_pred, 1)
    far_idx = K.sum(K.tf.cast(K.reshape(far_idx, [-1]), 'int32'))


    return { 'total_adr': total_adr,
             'adr': adr_idx / total_adr,
             'total_far': total_far,
             'far': far_idx / total_far
           }