Tensorflow - 奇怪的验证丢失行为

Question

My validation loss is behaving in a very weird way

What does this behaviour mean?

My network is a simple NN with two hidden layers of 100 nodes each, an input layer with 3 units and an output layer of 50 units. It is used to learn a mapping between a set of 3 coordinates to a time series of 50 time components. There is one and one only time series possible for each triplet of coordinates. The aim is to train the network with a few samples, so that the network learns how to predict the time series associated to a generic triplet of coordinates. This is the network

c = tf.placeholder(tf.float32, shape=[None, 3])    
act_f = getattr(tf.nn, 'leaky_relu')

def latent_P(coord):
    h2 = tf.layers.dense(coord, 100, activation=act_f)
    h1 = tf.layers.dense(h2, 100, activation=act_f)
    logits = tf.layers.dense(h1, 50)
    return logits

z_nn_samples = latent_P(c)

recon_loss_nn = tf.keras.losses.MAE(z, z_nn_samples)

loss_nn = tf.reduce_mean(recon_loss_nn)

solver_nn = tf.train.AdamOptimizer().minimize(loss_nn)

My training set is made of 3000 time series (with their respective 3000 triplets of coordinates). The validation set is made of 500 time series/ coordinates.

Answer 1

训练和验证数据的分布不同和/或您过度拟合。