def add_eval_step(self, result_tensor, ground_truth_tensor):
with tf.name_scope('correlations'):
with tf.name_scope('Yaw_correlation'):
yaw_correlation, yaw_update_op = streaming_pearson_correlation(
predictions=result_tensor[0],
labels=ground_truth_tensor[0])
tf.summary.scalar('Yaw_correlation', yaw_update_op)
with tf.name_scope('Roll_correlation'):
roll_correlation, roll_update_op = streaming_pearson_correlation(
predictions=result_tensor[1],
labels=ground_truth_tensor[1])
tf.summary.scalar('Roll_correlation', roll_update_op)
with tf.name_scope('Pitch_correlation'):
pitch_correlation, pitch_update_op = streaming_pearson_correlation(
predictions=result_tensor[2],
labels=ground_truth_tensor[2])
tf.summary.scalar('Pitch_correlation', pitch_update_op)
final_correlations = [
yaw_correlation, roll_correlation, pitch_correlation
]
streaming_correlations = [
yaw_update_op, roll_update_op, pitch_update_op
]
return streaming_correlations, final_correlations
def metric_fn(per_example_loss, label_ids, logits,
is_real_example):
"""Compute Pearson correlations for STS-B."""
# Display labels and predictions
concat1 = contrib_metrics.streaming_concat(logits)
concat2 = contrib_metrics.streaming_concat(label_ids)
# Compute Pearson correlation
pearson = contrib_metrics.streaming_pearson_correlation(
logits, label_ids, weights=is_real_example)
# Compute MSE
# mse = tf.metrics.mean(per_example_loss)
mse = tf.metrics.mean_squared_error(
label_ids, logits, weights=is_real_example)
loss = tf.metrics.mean(values=per_example_loss,
weights=is_real_example)
return {
"pred": concat1,
"label_ids": concat2,
"pearson": pearson,
"MSE": mse,
"eval_loss": loss,
}
def _define_metrics(self):
mse = tf.reduce_mean(tf.square(self.prediction - self.label), name='mse')
mae = tf.reduce_mean(tf.abs(self.prediction - self.label), name='mae')
pcc = streaming_pearson_correlation(self.prediction, self.label, name='pcc')
self.metrics['mse'] = mse
self.metrics['mae'] = mae
self.metrics['pcc'] = pcc
def rnn_framework(features, labels, mode):
#输入层
input_layer = tf.reshape(features, [-1, 256, 128])
#LSTM神经元
lstm_cell = rnn.BasicLSTMCell(50)
#RNN
_, Rnn = dynamic_rnn(lstm_cell, input_layer, dtype=tf.float64)
logits = tf.layers.dense(inputs=Rnn.h, units=8)
#生成预测（字典,分别为单值预测和概率预测）
predictions = {
'classes': tf.argmax(input=logits, axis=1),
'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
}
#预测功能：当mode处于预测模式时，返回字典predictions作为对样本的分类预测结果
if mode == tensorflow_estimator.estimator.ModeKeys.PREDICT:
return tensorflow_estimator.estimator.EstimatorSpec(
mode=mode, predictions=predictions)
#损失计算(mode处于训练模式)，使用交叉熵作为损失指标
loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)
#配置训练操作
if mode == tensorflow_estimator.estimator.ModeKeys.TRAIN:
optimizer = tf.train.AdagradOptimizer(learning_rate=0.001)
op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
return tensorflow_estimator.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=op)
# 建立评估指标
accuracy = tf.metrics.accuracy(labels=labels,
predictions=predictions['classes'])
precision = tf.metrics.precision(labels=labels,
predictions=predictions['classes'])
recall = tf.metrics.recall(labels=labels,
predictions=predictions['classes'])
coe = streaming_pearson_correlation(predictions=tf.to_float(
predictions['classes']),
labels=tf.to_float(labels))
evaluate = {
'accuracy': accuracy,
'precision': precision,
'recall': recall,
'coef': coe
}
return tensorflow_estimator.estimator.EstimatorSpec(
mode=mode, loss=loss, eval_metric_ops=evaluate)
def pearson(y_true, y_pred):
result = streaming_pearson_correlation(y_true, y_pred)
return result[0]
def cnn_framework(features, labels, mode):
#输入层
input_layer = tf.reshape(features, [-1, 256, 128, 1])
#卷积层1
conv1 = tf.layers.conv2d(inputs=input_layer,
filters=32,
kernel_size=[5, 128],
padding='valid',
activation=tf.nn.tanh)
#输出张量:[252,1,32]
#池化层1
pool1 = tf.layers.max_pooling2d(inputs=conv1,
pool_size=[2, 1],
strides=[2, 1])
#输出张量:[126,1,32]
#全连接层
#将pool1扁平化
flat = tf.reshape(pool1, [-1, 126 * 1 * 32])
#全连接层
dense = tf.layers.dense(inputs=flat, units=1024, activation=tf.nn.tanh)
#为改善模型效果，对全连接层应用丢弃正则化
dropout = tf.layers.dropout(
inputs=dense,
rate=0.4,
training=mode == tensorflow_estimator.estimator.ModeKeys.TRAIN)
#对数层
logits = tf.layers.dense(inputs=dropout, units=8)
#生成预测（字典,分别为单值预测和概率预测）
predictions = {
'classes': tf.argmax(input=logits, axis=1),
'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
}
#预测功能：当mode处于预测模式时，返回字典predictions作为对样本的分类预测结果
if mode == tensorflow_estimator.estimator.ModeKeys.PREDICT:
return tensorflow_estimator.estimator.EstimatorSpec(
mode=mode, predictions=predictions)
#损失计算(mode处于训练模式)，使用交叉熵作为损失指标
loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)
#配置训练操作
if mode == tensorflow_estimator.estimator.ModeKeys.TRAIN:
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
return tensorflow_estimator.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=op)
accuracy = tf.metrics.accuracy(labels=labels,
predictions=predictions['classes'])
precision = tf.metrics.precision(labels=labels,
predictions=predictions['classes'])
recall = tf.metrics.recall(labels=labels,
predictions=predictions['classes'])
# labels64=tf.constant(labels,name='labels64',dtype=tf.int64)
# predictions64=tf.constant(predictions['classes'],name='predictions64',dtype=tf.int64)
coe = streaming_pearson_correlation(
predictions=tf.to_float(predictions['classes']),
# labels=labels
labels=tf.to_float(labels))
# f_score=tf.contrib.metrics.f1_score(
# labels=labels,
# predictions=predictions['classes']
# )
#建立评估指标
evaluate = {
'accuracy': accuracy,
'precision': precision,
'recall': recall,
'coef': coe
# 'f_score':f_score
}
return tensorflow_estimator.estimator.EstimatorSpec(
mode=mode, loss=loss, eval_metric_ops=evaluate)
def metric_fn(label_ids, predictions, is_real_example):
from tensorflow.contrib.metrics import streaming_pearson_correlation
eval_pearson_cosine = streaming_pearson_correlation(predictions, label_ids)
return {
"eval_pearson_cosine": eval_pearson_cosine
}
