GAN训练结果 D loss: nan, acc.: 50% G loss: nan

Question

我正在尝试实现一个 GAN 来生成网络流量 .csv 数据集（表格 GAN），我的训练结果继续显示 [D loss: nan, acc.: 50%] [G loss: nan]。我想这是因为我的数据集在预处理后有 NaN 值，所以我使用了代码“assert not np.any(np.isnan(x))”，我得到了下面的错误。我需要帮助...

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-30-5e46f39aeea1> in <module>
      5 #Training the GAN model chosen: Vanilla GAN, CGAN, DCGAN, etc.
      6 synthesizer = model_1(gan_args)
----> 7 synthesizer.train(dataset, train_args)

<ipython-input-26-65296d00d312> in train(self, data, train_arguments)
     72             #  Train Discriminator
     73             # ---------------------
---> 74             batch_data = self.get_data_batch(data, self.batch_size)
     75             noise = tf.random.normal((self.batch_size, self.noise_dim))
     76 

<ipython-input-26-65296d00d312> in get_data_batch(self, train, batch_size, seed)
     56         train_ix = list(train_ix) + list(train_ix)  # duplicate to cover ranges past the end 
of the set
     57         x = train.loc[train_ix[start_i: stop_i]].values
---> 58         assert not np.any(np.isnan(x))
     59         return np.reshape(x, (batch_size, -1))
     60 

AssertionError: `

Answer 1

我终于明白了。在删除不需要的列后使用 .dropna(how='any', inplace = True) 它解决了问题。现在我的结果以 93.57% 的准确率生成。

GAN训练结果 D loss: nan, acc.: 50% G loss: nan

GAN training result D loss: nan, acc.: 50% G loss: nan

python

dataset

tabular

generative-adversarial-network