报告训练数据集中特定样本的训练损失，而不是训练过程中的平均损失（TensorFlow）

Question

我正在使用TensorFlow训练LSTM模型。我们知道，在训练过程中，是每个时期的loss和val_loss的报告，这是训练和测试数据集损失的平均值。我打算跟踪火车数据集中特定样本的丢失（特定日期）。另外，应该注意，我在fit函数中对火车数据进行混洗。

Answer 1

这里是用于跟踪单个样本损失的代码：

import tensorflow as tf 
import numpy as np 
import keras 

x = tf.Variable(initial_value=np.ndarray(shape=(10, 10), dtype=np.float32)) # your sample input 
y =np.random.randint(0, 9, size=(10, )) # your sample label
y_labels = keras.utils.to_categorical(y, 10)

loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=x, logits=y_labels)) # loss operation for that particular sample 
tf.summary.scalar('loss', loss) #logging loss op in summary 

print('loss op', loss)
merge = tf.summary.merge_all()
saver = tf.train.Saver()

with tf.Session() as sess: 
    sess.run(tf.global_variables_initializer())
    loss_val, merge_val = sess.run([loss, merge]) # no need to pass any feed_dict, loss value calculated is specific to that sample 
    print('loss val', loss_val)
    # merge val could be put in to tf summary writer