Question

我是机器学习和张量流的新手。而且我学到了一些相关材料。我完成了本教程中讲授的一些项目，例如mnist，图像识别。现在，我想对自己的数据集进行训练，该数据集是一组15秒的音乐剪辑。我使用librosa.mfcc提取了它们的功能。我将mfcc数组作为输入，使用RNN模型对其进行训练。但是在训练过程中，损失的结果是如此奇怪。

Step 0,train loss = 11.72
Step 50,train loss = 0.00
Step 100,train loss =0.72
Step 150,train loss =0.08
Step 200,train loss =0.64
Step 250,train loss =0.64
Step 300,train loss =0.00
Step 350,train loss =0.62
Step 400,train loss =0.61
Step 450,train loss = 115.77

我没关系。是输入数据预处理的问题吗？还是不适合使用RNN模型？

形状为（128,1293）的输入数据如下：

[[-2.40214356e+02 -2.54111029e+02 -2.81576989e+02 ... -3.26748334e+02
  -3.13127357e+02 -3.10083835e+02]
 [ 1.55226378e+02  1.88829858e+02  2.22116743e+02 ...  2.02720581e+02
   1.88478421e+02  1.71466354e+02]
 [-7.25124769e+01 -7.66927520e+01 -7.35990460e+01 ... -6.97141304e+01
  -8.91782486e+01 -1.01798663e+02]
 ...
 [-2.13188683e+00 -1.47389498e+00  4.32850268e-01 ... -8.17353566e-01
   1.74879699e-01  1.55565475e+00]
 [-1.18913985e+00 -1.75976975e+00 -5.36811511e-01 ... -1.70165869e+00
   1.08840259e+00  3.49373224e+00]
 [-1.80539142e-01 -4.37886115e-01 -5.02952858e-01 ... -1.91972103e+00
  -1.48080339e-01  9.51365549e-01]]

我的RNN模型是这样的：

def inference(input_mfcc, train):
with tf.device('/gpu:0'):
    with tf.variable_scope('conv1'):
        # 128*1293  conv1 29*294*32 ===> 100*1000*32
        # 100*1000*32  pool1 4*4  s4====>25*250*32
        conv1 = tf.layers.conv2d(inputs=input_mfcc,
                                filters=32,
                                kernel_size=[29,294],
                                padding='valid',
                                activation=tf.nn.relu)

        pool1 = tf.layers.max_pooling2d(inputs=conv1,pool_size=[4,4],strides=4)
        print("conv1:",conv1.get_shape().as_list())
        print("pool1:",pool1.get_shape().as_list())
    with tf.variable_scope('conv2'):
        # 25*250  conv2 6*51*64 ===> 20*200*64
        # 20*200*64  pool1 4*4  s4====> 5*50*64
        conv2 = tf.layers.conv2d(inputs=pool1,
                                filters=64,
                                kernel_size=[6,51],
                                padding='valid',
                                activation=tf.nn.relu)
        pool2 = tf.layers.max_pooling2d(inputs=conv2,pool_size=[4,4],strides=4)
        print("conv2:",conv2.get_shape().as_list())
        print("pool2:",pool2.get_shape().as_list())

    with tf.variable_scope('conv3'):
        #5*5*64
        conv3 = tf.layers.conv2d(inputs=pool2,
                                filters=64,
                                kernel_size=[1,46],
                                padding='valid',
                                activation=tf.nn.relu)
        print("conv3",conv3.get_shape().as_list())

    with tf.variable_scope('fc1'):
        pool2_flat = tf.reshape(pool2,[1,-1])
        print("pool2_flat",pool2_flat.get_shape().as_list())
        fc1 = tf.layers.dense(inputs=pool2_flat, units=1024, activation=tf.nn.relu)
        dropout1 = tf.layers.dropout(inputs=fc1, rate=0.4, training=train)
        print("dropout1",dropout1.get_shape().as_list())
    with tf.variable_scope('logits'):
        logits = tf.layers.dense(inputs=dropout1, units=2)
        #predit = tf.nn.softmax(logits=logits)
        print("logits",logits.get_shape().as_list())
        #print("predit",predit.get_shape().as_list())
return logits

其余代码为：

def losses(logits,labels):
    cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=labels,logits=logits,name='cross_entropy')
    cross_entropy_loss = tf.reduce_mean(cross_entropy)
    return cross_entropy_loss

def training(loss,learning_rate):

    with tf.name_scope("optimizer"):
        optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
        global_step = tf.Variable(0, name="global_step", trainable=False)
        train_op = optimizer.minimize(loss, global_step=global_step)
    return train_op
ckpt="./model/music/model.ckpt"
N_CLASSES = 2
MFCC_ROW = 128
MFCC_COL = 1293
INPUT_NODE = MFCC_ROW * MFCC_COL
BATCH_SIZE = 1
CAPACITY = 10
MAX_STEP = 500
learning_rate = 0.0001

def run_train():

    train_dir = ""
    logs_train_dir = ""
    mfcc, label= read_TFRecord()
    train_batch,train_labels_batch = tf.train.shuffle_batch([mfcc,label],batch_size=BATCH_SIZE,num_threads=1,capacity=CAPACITY,min_after_dequeue=5)
    print("train_batch",train_batch.get_shape().as_list())
    print("labels_batch",train_labels_batch.get_shape().as_list())
    train_logits = inference(train_batch,True)
    print("train_logits",train_logits.get_shape().as_list())
    train_loss = losses(train_logits, train_labels_batch)
    train_op = training(train_loss,learning_rate)
    #train_acc = evaluation(train_logits,train_labels_batch)

    with tf.Session() as sess:
        saver = tf.train.Saver()
        init_op = tf.group(tf.local_variables_initializer(),
                            tf.global_variables_initializer())
        sess.run(init_op)
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(coord=coord)
        try:
            for step in range(MAX_STEP):
                if coord.should_stop():
                    break;
                _,tra_loss = sess.run([train_op,train_loss])
        # print some 
                if step%50==0:
                    print('Step %d,train loss = %.4f'%(step,tra_loss))
        # 100 save
                if step % 100 ==0 or (step +1) == MAX_STEP:
                    saver.save(sess,ckpt,global_step = step)

        except tf.errors.OutOfRangeError:
            print('Done training epoch limit reached')
        finally:
            coord.request_stop()
            coord.join(threads)

主要： run_train（）

有人可以给我一些建议吗？非常感谢！

为什么RNN的火车损失很奇怪？

0 个答案: