tensorflow 沒有為任何可變誤差提供梯度

Question

我正在嘗試reuse選項構建一個簡單的神經網絡，但出現一個奇怪的錯誤。 我不明白問題出在哪里。 也許我沒有正確使用mse 。

import tensorflow as tf                                                                                                                                                                                            


n_inputs = 8                                                                                                                                                                                                       

x_ = tf.placeholder(tf.float32, [None, n_inputs])                                                                                                                                                                  
l1 = tf.layers.dense(x_, 100, activation=tf.nn.relu, use_bias=True, name='l1', reuse=None)                                                                                                                         
l2 = tf.layers.dense(l1, 100, activation=tf.nn.relu, use_bias=True, name='l2', reuse=None)                                                                                                                         
l3 = tf.layers.dense(l2, 20, activation=tf.nn.relu, use_bias=True, name='l3', reuse=None)                                                                                                                          

y_ = tf.placeholder(tf.float32, [None, n_inputs])                                                                                                                                                                  
w1  = tf.layers.dense(y_, 100, activation=tf.nn.relu, use_bias=True, name='l1', reuse=True)                                                                                                                        
w2 = tf.layers.dense(w1, 100, activation=tf.nn.relu, use_bias=True, name='l2', reuse=True)                                                                                                                         
w3 = tf.layers.dense(w2,  20, activation=tf.nn.relu, use_bias=True, name='l3', reuse=True)                                                                                                                         

z_ = tf.placeholder(tf.float32, [None, n_inputs])                                                                                                                                                                  
u1  = tf.layers.dense(z_, 100, activation=tf.nn.relu, use_bias=True, name='l1', reuse=True)                                                                                                                        
u2 = tf.layers.dense(u1,  100, activation=tf.nn.relu, use_bias=True, name='l2', reuse=True)                                                                                                                        
u3 = tf.layers.dense(u2,  20, activation=tf.nn.relu, use_bias=True, name='l3', reuse=True)                                                                                                                         

mse1, _ = tf.metrics.mean_squared_error(l3, w3)                                                                                                                                                                    
mse2, _ = tf.metrics.mean_squared_error(l3,u3)                                                                                                                                                                     
cost = tf.subtract(mse1, mse2)                                                                                                                                                                                     

opts = tf.train.AdamOptimizer().minimize(cost)                                                                                                                                                                     
sess = tf.InteractiveSession()                                      

錯誤：

ValueError                                Traceback (most recent call last)                                                                                                                                        
<ipython-input-4-0e3679c2a898> in <module>()                                                                                                                                                                       
----> 1 __pyfile = open('''/tmp/py3823Cbm''');exec(compile(__pyfile.read(), '''/home/lpuggini/mlp/scratch/Kerberos/flow_ui.py''', 'exec'));__pyfile.close()                                                        

/home/lpuggini/mlp/scratch/Kerberos/flow_ui.py in <module>()                                                                                                                                                       
     33 cost = tf.subtract(mse1, mse2)                                                                                                                                                                             
     34                                                                                                                                                                                                            
---> 35 opts = tf.train.AdamOptimizer().minimize(cost)                                                                                                                                                             
     36 sess = tf.InteractiveSession()                                                                                                                                                                             
     37                                                                                                                                                                                                            

/home/lpuggini/MyApps/scientific_python_2_7/lib/python2.7/site-packages/tensorflow/python/training/optimizer.pyc in minimize(self, loss, global_step, var_list, gate_gradients, aggregation_method, colocate_gradi\
ents_with_ops, name, grad_loss)                                                                                                                                                                                    
    320           "No gradients provided for any variable, check your graph for ops"                                                                                                                               
    321           " that do not support gradients, between variables %s and loss %s." %                                                                                                                            
--> 322           ([str(v) for _, v in grads_and_vars], loss))                                                                                                                                                     
    323                                                                                                                                                                                                            
    324     return self.apply_gradients(grads_and_vars, global_step=global_step,                                                                                                                                   

ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients, between variables ["<tf.Variable 'l1/kernel:0' shape=(8, 100) dtype=float32_ref>", "<tf.Variable 'l1/b\
ias:0' shape=(100,) dtype=float32_ref>", "<tf.Variable 'l2/kernel:0' shape=(100, 100) dtype=float32_ref>", "<tf.Variable 'l2/bias:0' shape=(100,) dtype=float32_ref>", "<tf.Variable 'l3/kernel:0' shape=(100, 20)\
 dtype=float32_ref>", "<tf.Variable 'l3/bias:0' shape=(20,) dtype=float32_ref>"] and loss Tensor("Sub:0", shape=(), dtype=float32).                 

Answer 1

metrics不是losses 。 指標會隨着時間的推移記錄某些統計數據。 通過它們進行區分是沒有意義的。 除了對指標的核心TF文檔，這里是一個很好的寫了。

你想要的是https://www.tensorflow.org/api_docs/python/tf/losses 。 更具體地說， https://www.tensorflow.org/api_docs/python/tf/losses/mean_squared_error