python - 用 tensorflow 训练模型，但损失就是不会下降？

Question

我正在训练一个对 62 个字母数字字符进行分类的模型。但是，损失值在前几批中首先急剧下降，然后落地，再也不会下降。我不知道哪里出了问题，也不知道如何调试模型.

这是训练日志的快照:

以下是训练数据的示例:

我使用 4 个 conv 层，后跟 1 个 fc 层，adam 优化器来最小化对数损失。我仔细检查了图像标签是否正确。所以我不知道哪里还有问题。

代码如下:

import numpy as np
import tensorflow as tf
import os
from PIL import Image
import shutil
import time

input = temp= tf.placeholder(dtype='float32', shape=(None,32,32,1), name='input')#(None,62)
label = tf.placeholder(dtype='float32',shape=(None,62))#(None,62)

temp = tf.layers.conv2d(inputs=temp,filters=32,kernel_size=(3,3),padding="SAME",activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.he_normal())#(None,32,32,32)
#temp = tf.layers.dropout(inputs=temp,rate=0.5)
temp = tf.layers.max_pooling2d(temp,pool_size=[2, 2], strides=2)#(None,16,16,32)

temp = tf.layers.conv2d(inputs=temp,filters=64,kernel_size=(3,3),padding="SAME",activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.he_normal())#(None,16,16,64)
#temp = tf.layers.dropout(inputs=temp,rate=0.2)
temp = tf.layers.max_pooling2d(temp,pool_size=[2, 2], strides=2)#(None,8,8,64)

temp = tf.layers.conv2d(inputs=temp,filters=128,kernel_size=(3,3),padding="SAME",activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.he_normal())#(None,8,8,128)
temp = tf.layers.dropout(inputs=temp,rate=0.2)
temp = tf.layers.max_pooling2d(temp,pool_size=[2, 2], strides=2)#(None,4,4,128)

temp = tf.layers.conv2d(inputs=temp,filters=256,kernel_size=(3,3),padding="SAME",activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.he_normal())#(None,4,4,256)
temp = tf.layers.dropout(inputs=temp,rate=0.2)
temp = tf.layers.max_pooling2d(temp,pool_size=[2, 2], strides=2)#(None,2,2,256)

temp = tf.layers.conv2d(inputs=temp,filters=62,kernel_size=(2,2),padding="VALID",kernel_initializer=tf.keras.initializers.he_normal())#fc (None,1,1,62)
output = temp = tf.layers.flatten(temp)#(None,62)
output = tf.nn.softmax(output)

#loss
output_clip = tf.clip_by_value(output,1e-7,1-1e-7)
loss = tf.reduce_mean(tf.reduce_sum(-label*tf.log(output_clip)-(1-label)*tf.log(1-output_clip),axis=-1))#scaler
optimizer = tf.train.AdamOptimizer().minimize(loss)

#accuracy
indexoutput = tf.argmax(output,axis=-1)#(None,)
labelindex = tf.argmax(label,axis=-1)#(None,)
equals = tf.equal(indexoutput,labelindex)#(None,)
equals = tf.reduce_sum(tf.cast(equals,dtype='int8'),axis=-1)#scaler
acc = tf.cast(equals,dtype='float32')/tf.cast(tf.shape(output)[0],dtype='float32')#scaler

def train(epochs):
    saver = tf.train.Saver()
    lossrec=[]
    accrec = []
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        #saver.restore(sess, os.path.join(os.getcwd(),'model_logloss','captchabreak.ckpt'))
        valimg,vallabel = next(validategenerator(os.path.join(os.getcwd(),'tests')))
        for i in range(epochs):
            for j, (trainimg, trainlabel) in enumerate(traingenerator(os.path.join(os.getcwd(),'pics'),32)):
                _, trainacc, trainloss = sess.run([optimizer, acc, loss],feed_dict={input: trainimg, label: trainlabel});
                valacc, valloss = sess.run([acc, loss], feed_dict={input: np.array(valimg), label: np.array(vallabel)})
                print("epoch:{} batch:{} trainloss:{:.4f} validateloss:{:.4f} trainacc:{:.2f} validateacc:{:.2f}"
                      .format(i, j, trainloss, valloss, trainacc, valacc))
            #keep some logs
            lossrec.append(valloss)
            accrec.append(valacc)
            print(lossrec);
            print(accrec)
            if len(lossrec) >= 3 and valloss >= lossrec[-2] and valloss >= lossrec[-3]: break;
            shutil.rmtree("model_logloss")
            saver.save(sess, "model_logloss/captchabreak.ckpt")

def traingenerator(path,batch_size):
    fs = os.listdir(path);
    fs=np.random.permutation(fs)
    if batch_size == 0:
        batch_size = len(fs)
    offset=0
    while offset<len(fs):
        yield fetch(fs[offset:offset+batch_size],path)
        offset+=batch_size

def validategenerator(path):
    fs = os.listdir(path);
    yield fetch(fs,path)

def fetch(fs,path):
    imgs = []
    labels = []
    for i, fname in enumerate(fs):
        fp = os.path.join(path, fname)
        imp = Image.open(fp).resize((32, 32));
        imp = imp.convert('L')
        imp = imp.point(lambda p: p > 210 and 255)
        im = np.array(imp)
        im = np.expand_dims(im, axis=-1)
        imp.close()

        c = fname[0]
        lb = np.zeros((62))
        if ord(c) >= 48 and ord(c) <= 57:
            lb[ord(c) - 48] = 1
        if ord(c) >= 65 and ord(c) <= 90:
            lb[ord(c) - 65 + 10] = 1
        if ord(c) >= 97 and ord(c) <= 122:
            lb[ord(c) - 97 + 36] = 1
        imgs.append(im)
        labels.append(lb)
    return np.array(imgs), np.array(labels)

if __name__ == "__main__":
    train(30)

Answer 1

已经解决了。我忘记将像素值除以 255。