GAN实现MNIST图片生成——2.tensorflow图都搞不定

实现想法

我的想法是，将G的输入增加对应数字项；
在D的输出中，分为real和fake的判别输出，以及num的值输出；
用对真输入的判别和估值作为D的loss；
用对假输入的判别和估值作为G的loss；
这些思路都与原来变化不大，改动量较少；
进一步想法，用判别的loss和估值的loss分别进化网络不同层的参数，那是下一步想法。

修改input函数

在input函数中增加noise_num和real_num的输出项。由于我目前还不会将两个张量作为输入，只能在input中将noise的img和num合并再输入到G中了。

def get_inputs(real_size, noise_size, noise_n):
    
    real_img = tf.placeholder(tf.float32, [None, real_size], name='real_img')
    real_num = tf.placeholder(tf.float32, [None, 1], name='real_num')
    noise_img_o = tf.placeholder(tf.float32, [None, noise_size], name='noise_img_o')
    noise_num = tf.placeholder(tf.float32, [None, 1], name='noise_num')
    
    #real_img = tf.concat([real_img_o,real_num], 1, name='real_img')
    noise_img = tf.concat([noise_img_o, noise_num], 1, name='noise_img')
    
    return real_img, noise_img, real_num, noise_num

修改D的网络

G网络基本没有变化，只是输入宽度增加了1，其他没有变化；
D网络在网络末端进行了修改；

leaky ReLU

softmax

全连接层

sigmoid

图像数据

全连接层

全连接层

logits_max

outputs_num

logits_sig

输出判别结果

def get_discriminator(img, n_units, reuse=False, alpha=0.01):
    
    with tf.variable_scope("discriminator", reuse=reuse):
        hidden1 = tf.layers.dense(img, n_units)
        hidden1 = tf.maximum(alpha * hidden1, hidden1)
        
        logits_max = tf.layers.dense(hidden1, 10)
        outputs_num = tf.nn.softmax(logits_max)
        logits_sig = tf.layers.dense(logits_max, 1)
        outputs = tf.sigmoid(logits_sig)
        
        return logits_sig, logits_max, outputs_num, outputs

修改loss函数

这里，就按照上边想法的方式进行输出。但是，我的天！！！我不知道为什么，在d_loss = tf.add(d_loss_real, d_loss_fake, d_loss_num_real)这句就一直都会报错！！！
这三个值应该都是tensor类型啊，都是tf.reduce_mean求出来。我真的不懂了，弄了一整天！！！
这里我真的要放弃tensorflow了，虽然刚刚领悟到他绘图以及sess的好，但是绘图过程中出现这个问题，真的不知道怎么调整啊。目前我感觉应该是这三个tensor求和处，各个tensor维度不同导致的问题。但是，由于没到session呢，用tfdbg都看不到tensor值啊。我不知道怎么办了，感觉老手应该秒解决这个问题。**我放弃了！**我去看pytorch了。

  File "", line 1, in 
    runfile('E:/nuts/20180502_personal/learn/DeepLearning/mnist_GAN/try_Mnist_GAN.py', wdir='E:/nuts/20180502_personal/learn/DeepLearning/mnist_GAN')

  File "d:\programdata\anaconda3\envs\l_tf_36\lib\site-packages\spyder_kernels\customize\spydercustomize.py", line 827, in runfile
    execfile(filename, namespace)

  File "d:\programdata\anaconda3\envs\l_tf_36\lib\site-packages\spyder_kernels\customize\spydercustomize.py", line 110, in execfile
    exec(compile(f.read(), filename, 'exec'), namespace)

  File "E:/nuts/20180502_personal/learn/DeepLearning/mnist_GAN/try_Mnist_GAN.py", line 94, in 
    d_loss = tf.add(d_loss_real, d_loss_fake, d_loss_num_real)

  File "d:\programdata\anaconda3\envs\l_tf_36\lib\site-packages\tensorflow\python\ops\gen_math_ops.py", line 374, in add
    "Add", x=x, y=y, name=name)

  File "d:\programdata\anaconda3\envs\l_tf_36\lib\site-packages\tensorflow\python\framework\op_def_library.py", line 394, in _apply_op_helper
    with g.as_default(), ops.name_scope(name) as scope:

  File "d:\programdata\anaconda3\envs\l_tf_36\lib\site-packages\tensorflow\python\framework\ops.py", line 6088, in __enter__
    return self._name_scope.__enter__()

  File "d:\programdata\anaconda3\envs\l_tf_36\lib\contextlib.py", line 81, in __enter__
    return next(self.gen)

  File "d:\programdata\anaconda3\envs\l_tf_36\lib\site-packages\tensorflow\python\framework\ops.py", line 3980, in name_scope
    if name:

  File "d:\programdata\anaconda3\envs\l_tf_36\lib\site-packages\tensorflow\python\framework\ops.py", line 653, in __bool__
    raise TypeError("Using a `tf.Tensor` as a Python `bool` is not allowed. "

TypeError: Using a `tf.Tensor` as a Python `bool` is not allowed. Use `if t is not None:` instead of `if t:` to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the value of a tensor.

# discriminator
d_logits_sig_real, d_logits_max_real, d_num_real, d_outputs_real = get_discriminator(real_img, d_units)
d_logits_sig_fake, d_logits_max_fake, d_num_fake, d_outputs_fake = get_discriminator(g_outputs, d_units, reuse=True)


# discriminator的loss
# 识别真实图片
d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_sig_real, labels=tf.ones_like(d_logits_sig_real)) * (1 - smooth))
# 识别生成的图片
d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_sig_fake, labels=tf.zeros_like(d_logits_sig_fake)))

# 识别真图值
d_loss_num_real = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=d_logits_max_real, labels=tf.cast(tf.reduce_sum(real_num, 1), dtype=tf.int32)))

# 总体loss
d_loss = tf.add(d_loss_real, d_loss_fake, d_loss_num_real)

# generator的loss
g_loss_realfake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_sig_fake, labels=tf.ones_like(d_logits_sig_fake)) * (1 - smooth))

# g图值
g_loss_num = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=d_logits_max_fake, labels=tf.cast(tf.reduce_sum(noise_num, 1), dtype=tf.int32)))

# 总体loss
g_loss = tf.add(g_loss_realfake, g_loss_num)