Tensorflow中用于学习身份功能的怪异损失开发

Question

我尝试在Tensorflow中学习从0到200的数字的身份函数，它具有一个简单的完全连接的网络，该网络由一个具有512个神经元的隐藏层和一个具有200个神经元的输出层组成（虽然杀伤力强，但仅用于测试）。使用亚当和softmax交叉熵作为损失函数，损失从刚初始化的权重的16个迅速减少到约500步（约200个示例），经过约500个步骤，并停留在那里或仅非常缓慢地移动。

培训过程：

即使模型过于复杂，通常也不会仅仅拟合数据吗？我没有进行标准化，但是我认为这没有太大的不同，不是吗？ 这种行为是否有潜在原因，还是我的实现有错误（tensorflow 1.9）？

import tensorflow as tf
import numpy as np
from random import randint


def model(features, labels, mode, params):
    hidden = tf.layers.dense(features, 512, activation=tf.nn.relu)
    logits = tf.layers.dense(hidden, 201, activation=tf.nn.relu)

    if mode == tf.estimator.ModeKeys.PREDICT:
        return tf.estimator.EstimatorSpec(mode, predictions={'logits': logits})

    loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=labels)
    mean_loss = tf.reduce_mean(loss)

    learning_rate = params['learning_rate']
    train_op = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(mean_loss, global_step=tf.train.get_global_step())

    return tf.estimator.EstimatorSpec(mode, loss=mean_loss, train_op=train_op)


def train_input_fn(features, labels, batch_size):
    dataset = tf.data.Dataset.from_tensor_slices((features, labels))

    if batch_size is None:
        dataset = dataset.batch(len(labels))
    else:
        dataset = dataset.shuffle(1000).repeat().batch(batch_size)

    return dataset

我像这样开始训练和预测：

if __name__ == '__main__':
    feat = [[float(i)] for i in range(200)]
    lab = [int(i) for sublist in feat for i in sublist]

    learning_rate = 0.0001
    classifier = tf.estimator.Estimator(
        model_fn=model,
        model_dir='/home/test_dir/{}'.format(learning_rate),
        params={'learning_rate': learning_rate}
    )

    classifier.train(input_fn=lambda: train_input_fn(feat, lab, 200), max_steps=800000)

    # the rest is just testing the prediction process
    input = [[float(randint(0, 200))] for i in range(10)]
    input_labels = np.array([int(i) for sublist in input for i in sublist])
    predictions = classifier.predict(input_fn=lambda: train_input_fn(input, input_labels, None))

    logits = []
    for p in predictions:
        logits.append(p['logits'])
    tensor_logits = tf.convert_to_tensor(np.array(logits))
    tensor_label_ids = tf.convert_to_tensor(input_labels)
    loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=tensor_logits, labels=tensor_label_ids)
    mean_loss = tf.reduce_mean(loss)
    preds = tf.argmax(tensor_logits, axis=-1)

    with tf.Session() as sess:
        print(input_labels)
        print(sess.run(preds))
        print(sess.run(loss))
        print(sess.run(mean_loss))