这是代码
num_epochs = 100total_series_length = 50000truncated_backprop_length = 15state_size = 4num_classes = 2echo_step = 3batch_size = 5num_batches = total_series_length//batch_size//truncated_backprop_length生成数据{…}
batchX_placeholder = tf.placeholder(tf.int32, [batch_size, truncated_backprop_length])batchY_placeholder = tf.placeholder(tf.int32, [batch_size, truncated_backprop_length])#以及一个用于RNN状态的，5,4 init_state = tf.placeholder(tf.float32, [batch_size, state_size])batchX_placeholder = tf.one_hot(batchX_placeholder, num_classes)inputs_series = tf.unstack(batchX_placeholder, axis=1)cell = tf.contrib.rnn.BasicRNNCell(state_size)rnn_outputs, final_state = tf.contrib.rnn.static_rnn(cell, inputs_series, initial_state=init_state)一些优化代码{….}然后创建图形
#第3步 训练网络with tf.Session() as sess: #我们每次都必须这样做，它应该只是知道 #我们已经初始化了这些变量。v2版的家伙们，v2.. sess.run(tf.initialize_all_variables()) #交互模式 plt.ion() #初始化图形 plt.figure() #显示图形 plt.show() #显示损失减少 loss_list = [] for epoch_idx in range(num_epochs): #在每个epoch生成数据，批次在epoch中运行 x,y = generateData() #初始化一个空的隐藏状态 _current_state = np.zeros((batch_size, state_size)) print("新数据，epoch", epoch_idx) #每个批次 for batch_idx in range(num_batches): #每个批次的开始和结束点 #因为权重在每个时间层重复 #这些层不会被展开到时间的开始， #那将是太高的计算成本，因此在有限的时间步内被截断 start_idx = batch_idx * truncated_backprop_length end_idx = start_idx + truncated_backprop_length batchX = x[:,start_idx:end_idx] batchY = y[:,start_idx:end_idx] #运行计算图，给它我们之前计算的值 _total_loss, _train_step, _final_state, _predictions_series = sess.run( [total_loss, train_step, final_state, predictions], feed_dict={ batchX_placeholder:batchX, batchY_placeholder:batchY, init_state:_current_state }) loss_list.append(_total_loss) if batch_idx%100 == 0: print("步骤",batch_idx, "损失", _total_loss) plot(loss_list, _predictions_series, batchX, batchY)plt.ioff()plt.show()这是错误信息：
ValueError Traceback (most recent call last)<ipython-input-9-7c3d1289d16b> in <module>() 40 batchX_placeholder:batchX, 41 batchY_placeholder:batchY,---> 42 init_state:_current_state 43 }) 44 /home/pranshu_44/anaconda3/lib/python3.5/site-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata) 765 try: 766 result = self._run(None, fetches, feed_dict, options_ptr,--> 767 run_metadata_ptr) 768 if run_metadata: 769 proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)/home/pranshu_44/anaconda3/lib/python3.5/site-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata) 942 'Cannot feed value of shape %r for Tensor %r, ' 943 'which has shape %r'--> 944 % (np_val.shape, subfeed_t.name, str(subfeed_t.get_shape()))) 945 if not self.graph.is_feedable(subfeed_t): 946 raise ValueError('Tensor %s may not be fed.' % subfeed_t)ValueError: Cannot feed value of shape (5, 15) for Tensor 'one_hot:0', which has shape '(5, 15, 2)'我查看了文档，但那一点也不helpful，如果有其他简单的方法也会很有帮助
回答：
您将占位符变量转换为one-hot表示，但没有将您在训练期间实际输入网络的数据转换。尝试在输入之前将batchX转换为one-hot表示。这段代码将矩阵转换为其one-hot表示：
# 假设y包含从0到N-1的值，其中N是类别数batchX = (np.arange(max(batchX.flatten())+1) == batchX[:,:,None]).astype(int)