Question

Tensorflow 1.14给我问题“无法为Tensor'占位符_1：0'提供形状（100，）的值，它具有形状'（？，1）'”，如问题中所述。此错误发生在代码的会话段中[代码的最后一个单元格]。我该如何解决？我无法理解错误所在。该代码为iPython格式。提前致谢。

代码如下：

#!/usr/bin/env python
# coding: utf-8

# In[1]:
import tensorflow as tf
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.metrics import confusion_matrix
import random
import math
import time
import h5py

# In[2]:
if len(tf.config.experimental.list_physical_devices('GPU')) > 0:
    tf.debugging.set_log_device_placement(True)

# In[3]:
df = pd.read_hdf('../Dataset/dataset.h5',key='df')

# In[4]:
df.head()

# In[5]:
df.pop('App')
df_y = df.pop('Label')
df_x = df

# In[6]:
trn_x, val_x, trn_y, val_y = train_test_split(df_x, df_y, test_size=0.3)

# In[7]:
trn_x.head()

# In[8]:
trn_y.head()

# In[9]:
val_x.head()

# In[10]:
val_y.head()

# In[11]:
train_size,num_features = trn_x.shape
print(train_size)

# In[12]:
x = tf.placeholder("float", shape=[None, num_features])
y = tf.placeholder("float", shape=[None,1])

# In[13]:
w = tf.Variable(np.random.rand(num_features,1), dtype = np.float32)
b = tf.Variable(np.random.rand(1,1), dtype = np.float32)

# In[14]:
C = 1
batch = 100
epochs = 100

# In[15]:
y_cap = tf.matmul(x,w)+b

# In[16]:
reg_loss = 0.5*tf.math.sqrt(tf.reduce_sum(tf.square(w)))
hinge_loss = C*tf.reduce_sum(tf.square(tf.maximum(tf.zeros([batch,1]),1-y*y_cap)))
loss = reg_loss + hinge_loss

# In[17]:
train_step = tf.train.GradientDescentOptimizer(0.01).minimize(loss)

# In[18]:
label = tf.sign(y_cap)

# In[19]:
c_pred = tf.equal(y,label)

# In[20]:
accuracy = tf.reduce_sum(tf.cast(c_pred, "float"))

# In[23]:
with tf.Session() as s:
    tf.initialize_all_variables().run()
    total_steps = epochs*train_size//batch
    for step in range(total_steps):
        print("step "+str(step)+"/"+str(total_steps)+":",end="\t")
        offset = (step*batch)%train_size
        b_data = trn_x[offset: (offset+batch)]
        b_label = trn_y[offset: (offset+batch)]
        train_step.run(feed_dict={x:b_data, y:b_label})
        print("loss:"+loss.eval(feed_dict={x:b_data, y:b_label})

详细错误日志：

ValueError Traceback (most recent call last)
<ipython-input-23-57d6ca3d1893> in <module>
      7         b_data = trn_x[offset: (offset+batch)]
      8         b_label = trn_y[offset: (offset+batch)]
----> 9         train_step.run(feed_dict={x:b_data, y:b_label})
     10         print("loss:"+loss.eval(feed_dict={x:b_data, y:b_label}))

/usr/local/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in run(self, feed_dict, session)
   2677         none, the default session will be used.
   2678     """
-> 2679     _run_using_default_session(self, feed_dict, self.graph, session)
   2680 
   2681 _gradient_registry = registry.Registry("gradient")

/usr/local/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in _run_using_default_session(operation, feed_dict, graph, session)
   5612                        "the operation's graph is different from the session's "
   5613                        "graph.")
-> 5614   session.run(operation, feed_dict)
   5615 
   5616 

/usr/local/lib/python3.7/site-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)
    948     try:
    949       result = self._run(None, fetches, feed_dict, options_ptr,
--> 950                          run_metadata_ptr)
    951       if run_metadata:
    952         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)

/usr/local/lib/python3.7/site-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)
   1147                              'which has shape %r' %
   1148                              (np_val.shape, subfeed_t.name,
-> 1149                               str(subfeed_t.get_shape())))
   1150           if not self.graph.is_feedable(subfeed_t):
   1151             raise ValueError('Tensor %s may not be fed.' % subfeed_t)

ValueError: Cannot feed value of shape (100,) for Tensor 'Placeholder_1:0', which has shape '(?, 1)'

Answer 1

正如错误的说法，您应该匹配占位符和输入数据。我认为您的第二个占位符“ y”具有[？，1]形状，但是“ b_label”仅具有[100]形状。您应该在运行网络之前添加此代码。

b_label = np.expand_dims(b_label,axis=-1)

此代码使b_label的形状为[100,1]，这是与占位符匹配的大小。

Answer 2

这是因为您正在尝试为占位符提供形状为(batch,)的数据，而不是预期形状为(None, 1)的数据

因此，请尝试将trn_y重塑为（batch，1）的形状，如下所示，

# In[23]:
with tf.Session() as s:
    tf.initialize_all_variables().run()
    total_steps = epochs*train_size//batch
    for step in range(total_steps):
        print("step "+str(step)+"/"+str(total_steps)+":",end="\t")
        offset = (step*batch)%train_size
        b_data = trn_x[offset: (offset+batch)]
        b_label = trn_y[offset: (offset+batch)].values.reshape(batch, 1)
        train_step.run(feed_dict={x:b_data, y:b_label})
        print("loss:"+loss.eval(feed_dict={x:b_data, y:b_label})

Answer 3

y = tf.placeholder("float", shape=[None])

此行解决了我的问题。

无法为形状为（（？，1）''的张量'Placeholder_1：0'提供形状（100，）的值

3 个答案: