Tensorflow相同的训练精度继续

Question

我在Tensorflow上卡在CNN模型上。 我的代码如下。

库

# -*- coding: utf-8 -*- 
import tensorflow as tf 
import time 
import json 
import numpy as np 
import matplotlib.pyplot as plt 
import random 
import multiprocessing as mp 
import glob 
import os 

型号

def inference(images_placeholder, keep_prob): 

    def weight_variable(shape): 
     initial = tf.truncated_normal(shape, stddev=0.1) 
     return tf.Variable(initial) 

    def bias_variable(shape): 
     initial = tf.constant(0.1, shape=shape) 
     return tf.Variable(initial) 

    # convolution 
    def conv2d(x, W): 
     return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') 

    # X2 pooling 
    def max_pool_2x128(x): 
     return tf.nn.max_pool(x, ksize=[1, 2, 1, 1],strides=[1, 2, 1, 1], padding='VALID') 
    # X4 pooling 
    def max_pool_4x128(x): 
     return tf.nn.max_pool(x, ksize=[1, 4, 1, 1],strides=[1, 4, 1, 1], padding='VALID') 

    x_image = tf.reshape(images_placeholder, [-1,599,1,128]) 

    #1st conv 
    with tf.name_scope('conv1') as scope: 
     W_conv1 = weight_variable([4, 1, 128, 256]) 
     b_conv1 = bias_variable([256]) 

     print "image変形後のshape" 
     print tf.Tensor.get_shape(x_image) 
     print "conv1の形" 
     print tf.Tensor.get_shape(conv2d(x_image, W_conv1)) 

     h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) 

    #1st pooling X4 
    with tf.name_scope('pool1') as scope: 
     h_pool1 = max_pool_4x128(h_conv1) 
     print "h_pool1の形" 
     print tf.Tensor.get_shape(h_pool1) 

    #2nd conv 
    with tf.name_scope('conv2') as scope: 
     W_conv2 = weight_variable([4, 1, 256, 256]) 
     b_conv2 = bias_variable([256]) 
     h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) 

    #2nd pooling X2 
    with tf.name_scope('pool2') as scope: 
     h_pool2 = max_pool_2x128(h_conv2) 
     print "h_pool2の形" 
     print tf.Tensor.get_shape(h_pool2) 

    #3rd conv 
    with tf.name_scope('conv3') as scope: 
     W_conv3 = weight_variable([4, 1, 256, 512]) 
     b_conv3 = bias_variable([512]) 
     h_conv3 = tf.nn.relu(conv2d(h_pool2, W_conv3) + b_conv3) 

    #3rd pooling X2 
    with tf.name_scope('pool3') as scope: 
     h_pool3 = max_pool_2x128(h_conv3) 
     print "h_pool3の形" 
     print tf.Tensor.get_shape(h_pool3) 

    #flatten + 1st fully connected 
    with tf.name_scope('fc1') as scope: 
     W_fc1 = weight_variable([37 * 1 * 512, 2048]) 
     b_fc1 = bias_variable([2048]) 
     h_pool3_flat = tf.reshape(h_pool3, [-1, 37 * 1 * 512]) 
     h_fc1 = tf.nn.relu(tf.matmul(h_pool3_flat, W_fc1) + b_fc1) 
     #ドロップ層の設定 
     h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) 

    #2nd fully connected 
    with tf.name_scope('fc2') as scope: 
     W_fc2 = weight_variable([2048, NUM_CLASSES]) 
     b_fc2 = bias_variable([NUM_CLASSES]) 

    #softmax output 
    with tf.name_scope('softmax') as scope: 
     y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2) 

    return y_conv 

损失

def loss(logits, labels): 
    # cross entropy 
    cross_entropy = -tf.reduce_sum(labels*tf.log(tf.clip_by_value(logits,1e-10,1.0))) 
    # TensorBoard 
    tf.scalar_summary("cross_entropy", cross_entropy) 
    return cross_entropy 

培训

def training(loss, learning_rate): 
    train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss) 
    return train_step 

精度

def accuracy(logits, labels): 
    correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1)) 
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) 
    tf.scalar_summary("accuracy", accuracy) 
    return accuracy 

主要

if __name__ == '__main__': 

    flags = tf.app.flags 
    FLAGS = flags.FLAGS 

    flags.DEFINE_string('train_dir', '/tmp/data', 'Directory to put the training data.') 
    flags.DEFINE_integer('max_steps', , 'Number of steps to run trainer.') 
    flags.DEFINE_integer('batch_size', 10, 'Batch size' 
         'Must divide evenly into the dataset sizes.') 
    flags.DEFINE_float('learning_rate', 1e-4, 'Initial learning rate.') 

    #num output 
    NUM_CLASSES = 5 
    #num frame 
    IMAGE_SIZE = 599 
    #tensor shape 
    IMAGE_PIXELS = IMAGE_SIZE*1*128 

    ################## 
    #modify the data # 
    ################## 

    #number of training data 
    train_num = 70 
    #loading data limit 
    data_limit = 100 

    flatten_data = [] 
    flatten_label = [] 

    # データの整形 
    filenames = glob.glob(os.path.join('/Users/kosukefukui/Qosmo/WASABEAT/song_features/*.json')) 
    filenames = filenames[0:data_limit] 
    print "----loading data---" 
    for file_path in filenames: 
     data = json.load(open(file_path)) 
     data = np.array(data) 

     for_flat = np.array(data) 
     assert for_flat.flatten().shape == (IMAGE_PIXELS,) 
     flatten_data.append(for_flat.flatten().tolist()) 

    # ラベルの整形 
    f2 = open("id_information.txt") 
    print "---loading labels----" 

    for line in f2: 
     line = line.rstrip() 
     l = line.split(",") 
     tmp = np.zeros(NUM_CLASSES) 
     tmp[int(l[4])] = 1 
     flatten_label.append(tmp) 

    flatten_label = flatten_label[0:data_limit] 

    print "データ数 %s" % len(flatten_data) 
    print "ラベルデータ数 %s" % len(flatten_label) 

    #train data 
    train_image = np.asarray(flatten_data[0:train_num], dtype=np.float32) 
    train_label = np.asarray(flatten_label[0:train_num],dtype=np.float32) 

    print "訓練データ数 %s" % len(train_image) 

    #test data 
    test_image = np.asarray(flatten_data[train_num:data_limit], dtype=np.float32) 
    test_label = np.asarray(flatten_label[train_num:data_limit],dtype=np.float32) 

    print "テストデータ数 %s" % len(test_image) 

    print "599×128 = " 
    print len(train_image[0]) 

    f2.close() 

    if 1==1: 
     # Image Tensor 
     images_placeholder = tf.placeholder("float", shape=(None, IMAGE_PIXELS)) 
     # Label Tensor 
     labels_placeholder = tf.placeholder("float", shape=(None, NUM_CLASSES)) 
     # dropout Tensor 
     keep_prob = tf.placeholder("float") 

     # construct model 
     logits = inference(images_placeholder, keep_prob) 
     # calculate loss 
     loss_value = loss(logits, labels_placeholder) 
     # training 
     train_op = training(loss_value, FLAGS.learning_rate) 
     # accuracy 
     acc = accuracy(logits, labels_placeholder) 

     saver = tf.train.Saver() 
     sess = tf.Session() 
     sess.run(tf.initialize_all_variables()) 
     # for TensorBoard 
     summary_op = tf.merge_all_summaries() 
     summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph_def) 

     # Training 
     for step in range(FLAGS.max_steps): 
      for i in range(len(train_image)/FLAGS.batch_size): 
       # train for batch_size 
       batch = FLAGS.batch_size*i 
       sess.run(train_op, feed_dict={ 
        images_placeholder: train_image[batch:batch+FLAGS.batch_size], 
        labels_placeholder: train_label[batch:batch+FLAGS.batch_size], 
        keep_prob: 0.5}) 

      # calculate accuracy at each step 
      train_accuracy = sess.run(acc, feed_dict={ 
       images_placeholder: train_image, 
       labels_placeholder: train_label, 
       keep_prob: 1.0}) 
      print "step %d, training accuracy %g"%(step, train_accuracy) 

      # add value for Tensorboard at each step 
      summary_str = sess.run(summary_op, feed_dict={ 
       images_placeholder: train_image, 
       labels_placeholder: train_label, 
       keep_prob:1.0}) 
      summary_writer.add_summary(summary_str, step) 

    # show accuracy for test data 
    print "test accuracy %g"%sess.run(acc, feed_dict={ 
     images_placeholder: test_image, 
     labels_placeholder: test_label, 
     keep_prob: 1.0}) 
    # save the last model 
    save_path = saver.save(sess, "model.ckpt") 

但是，我得到了相同的训练精度。如何解决这个问题？

step 0, training accuracy 0.142857 
step 1, training accuracy 0.142857 
step 2, training accuracy 0.142857 
step 3, training accuracy 0.142857 
step 4, training accuracy 0.142857 
step 5, training accuracy 0.142857 
step 6, training accuracy 0.142857 
step 7, training accuracy 0.142857 
step 8, training accuracy 0.142857 
step 9, training accuracy 0.142857 
test accuracy 0.133333 

我提到了下面的模型和我的张量板如下。

Answer 1

难道你不减少正确的张量？ 您正在最小化cross_entropy，但应该是cross_entropy_mean（代码中的准确性）。

cross_entropy = tf.nn.softmax_cross_entropy_with_logits（ logits，ground_truth_placeholder）

cross_entropy_mean = tf.reduce_mean（cross_entropy）

train_step = tf.train.GradientDescentOptimizer：

与以下逻辑

基本上（FLAGS.learning_rate）.minimize（ cross_entropy_mean）