Tensorflow疑问： 在利用for循环构造一个循环网络时，只能对本循环内的变量求导？
我想对RNN中每个时间步的loss计算关于RNNkernel和bias的梯度，于是更改了static_rnn的源码。
问题代码是这样的：
kernel = tf.Variable(initial_kernel,name='meta_rnn_kernel',trainable=False);bias = tf.Variable(initial_bias,name='meta_rnn_bias',trainable=False) for time, input_ in enumerate(inputs): if time > 0: varscope.reuse_variables() # pylint: disable=cell-var-from-loop call_cell = lambda: cell.meta_call(input_, state,kernel,bias) output = zero_output # pylint: enable=cell-var-from-loop if sequence_length is not None: if time == 0: (output, state) = _rnn_step( time=time, sequence_length=sequence_length, min_sequence_length=min_sequence_length, max_sequence_length=max_sequence_length, zero_output=zero_output, state=state, call_cell=call_cell, state_size=cell.state_size) else: _update_kernel_bias(meta_learner,initial_kernel,initial_bias,kernel,bias,y[:,time],yn[:,time],\ output,embedding_params,x_nonzero[:,time],batch_size,timestep,max_items_basket,max_items_all) (output, state) = _rnn_step( time=time, sequence_length=sequence_length, min_sequence_length=min_sequence_length, max_sequence_length=max_sequence_length, zero_output=zero_output, state=state, call_cell=call_cell, state_size=cell.state_size)
其中_update_k_b函数为：
def _update_kernel_bias(meta_learner,initial_kernel,initial_bias,kernel,bias,y,yn,rnn_output,embedding_params,x_nonzero,batch_size,timestep,max_items_basket,max_items_all): '''y是上个时刻的输出[batch_size,max_items_basket] rnn_output[batch_size,latent_dimension] x_nonzero[batch_size] 注意这些输入都要是已经从time上切片了的''' # def is_in_for_tensor(element,set): # while i in range(int(set.shape[0])): # isIn = tf.constant(False,tf.bool) # isIn = isIn | tf.equal(element,set[i]) # return isIn # def generate_negatvie_samples_for_update(y,batch_size,timestep,max_items_basket,max_items_all): # '''训练时随机动态生成负样本，以用来计算BPRLoss。 # y [batch_size,max_items_basket] # 返回：ndarray[batch_size,max_items_basket]''' # ynlist=[] # for i in range(batch_size): # temp =[] # for k in range(max_items_basket): # negative_itemid = y[i][0] # while # while negative_itemid in y[i]: # negative_itemid = random.randrange(0,max_items_all) # temp+=[negative_itemid] # ynlist.append(temp) # return np.array(ynlist) loss=tf.constant(0.0,dtype=tf.float32) ye = tf.nn.embedding_lookup(embedding_params,y) #[batch_size,maxitems,ld] yne = tf.nn.embedding_lookup(embedding_params,yn) for i in range(batch_size): pos_scores = tf.reshape(tf.matmul(ye[i],tf.transpose(rnn_output[i:i+1])),(-1,)) #[maxitems] neg_scores = tf.reshape(tf.matmul(yne[i],tf.transpose(rnn_output[i:i+1])),(-1,)) #[maxitems] scores = pos_scores[:x_nonzero[i]]-neg_scores[:x_nonzero[i]] scores = -tf.log(tf.sigmoid(scores)) #[nonzero_nums] scores = tf.reduce_sum(scores) #全空的话，0；不全空的话，就正常 loss+= scores KERNEL_SHAPE = kernel.shape BIAS_SHAPE = bias.shape loss_gradients = tf.gradients(loss,[kernel,bias]) loss_gradients_k = tf.reshape(loss_gradients[0],shape=(-1,)) loss_gradients_b = tf.reshape(loss_gradients[1],shape=(-1,)) loss_gradients = tf.concat([loss_gradients_k,loss_gradients_b],axis=0) _kernel = tf.reshape(kernel,shape=(-1,)) KERNEL_SIZE = _kernel.shape[0] print(KERNEL_SIZE) _bias = tf.reshape(bias,shape=(-1,)) _initial_kernel = tf.reshape(initial_kernel,shape=(-1,)) _initial_bias = tf.reshape(initial_bias,shape=(-1,)) meta_learner_input = tf.concat([loss_gradients,_kernel,_bias,_initial_bias,_initial_kernel],axis=0) meta_learner_input = tf.reshape(meta_learner_input,shape=(1,-1)) meta_learner_output = meta_learner(meta_learner_input) new_kernel = meta_learner_output[:,:KERNEL_SIZE] new_kernel = tf.reshape(new_kernel,shape=KERNEL_SHAPE) print("newkernel",new_kernel) new_bias = meta_learner_output[:,KERNEL_SIZE:] new_bias = tf.reshape(new_bias,shape=BIAS_SHAPE) kernel.assign(new_kernel) bias.assign(new_bias)
然后这个跑出来报错：
Traceback (most recent call last): File "metaDREAM.py", line 163, in <module> train(lr=0.0001,batch_size=None,ewc_coefficient=1e-6,l2=1e-6,timestep=None,max_iter=10,pooling_type=None,variablescope='MetaTestA1',best_base_scope='TESTA4',bestepoch=1) File "metaDREAM.py", line 69, in train max_items_all,initial_state=init_state,time_major=True,sequence_length=xs) File "/home/likewise-open/SENSETIME/liupengcheng/Workspace/metaDream/test/myrnn_for_meta_testv2.py", line 126, in static_rnn_for_meta output,embedding_params,x_nonzero[:,time],batch_size,timestep,max_items_basket,max_items_all) File "/home/likewise-open/SENSETIME/liupengcheng/Workspace/metaDream/test/myrnn_for_meta_testv2.py", line 396, in _update_kernel_bias loss_gradients_k = tf.reshape(loss_gradients[0],shape=(-1,)) File "/home/likewise-open/SENSETIME/liupengcheng/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/gen_array_ops.py", line 3938, in reshape "Reshape", tensor=tensor, shape=shape, name=name) File "/home/likewise-open/SENSETIME/liupengcheng/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 528, in _apply_op_helper (input_name, err)) ValueError: Tried to convert 'tensor' to a tensor and failed. Error: None values not supported.
分析可得，问题不在reshape一行，而在reshape上面一行，求loss_gradients得到的是None，即反向传播时找不到loss和kernel的依赖关系。
很奇怪啊，我之前还写了对正常的static_rnn（包括dynamic_rnn）的BPR loss，都能正常求导的。为啥拆开了就不行了？
然后抱着试试看的心态，试出了下面的代码是没问题的：
for time, input_ in enumerate(inputs): if time > 0: varscope.reuse_variables() # pylint: disable=cell-var-from-loop call_cell = lambda: cell.meta_call(input_, state,kernel,bias) output = zero_output # pylint: enable=cell-var-from-loop if sequence_length is not None: if time != len(inputs): (output, state) = _rnn_step( time=time, sequence_length=sequence_length, min_sequence_length=min_sequence_length, max_sequence_length=max_sequence_length, zero_output=zero_output, state=state, call_cell=call_cell, state_size=cell.state_size) _update_kernel_bias(meta_learner,initial_kernel,initial_bias,kernel,bias,y[:,time],yn[:,time],\ output,embedding_params,x_nonzero[:,time],batch_size,timestep,max_items_basket,max_items_all) else: (output, state) = _rnn_step( time=time, sequence_length=sequence_length, min_sequence_length=min_sequence_length, max_sequence_length=max_sequence_length, zero_output=zero_output, state=state, call_cell=call_cell, state_size=cell.state_size)
也就是说，我在一个循环内求导是可以，而跑到下一个循环体里就不行了。
这没道理啊…我写循环其实相当于代码复用了，比如我有3个时间步，那么这两段代码展开来应该是一样的呀，为什么就不行呢？
写到这里我貌似明白了…因为time没有-1...修改
for time, input_ in enumerate(inputs): if time > 0: varscope.reuse_variables() # pylint: disable=cell-var-from-loop call_cell = lambda: cell.meta_call(input_, state,kernel,bias) output = zero_output # pylint: enable=cell-var-from-loop if sequence_length is not None: if time == 0: (output, state) = _rnn_step( time=time, sequence_length=sequence_length, min_sequence_length=min_sequence_length, max_sequence_length=max_sequence_length, zero_output=zero_output, state=state, call_cell=call_cell, state_size=cell.state_size) else: _update_kernel_bias(meta_learner,initial_kernel,initial_bias,kernel,bias,y[:,time-1],yn[:,time-1],\ output,embedding_params,x_nonzero[:,time-1],batch_size,timestep,max_items_basket,max_items_all) (output, state) = _rnn_step( time=time, sequence_length=sequence_length, min_sequence_length=min_sequence_length, max_sequence_length=max_sequence_length, zero_output=zero_output, state=state, call_cell=call_cell, state_size=cell.state_size)
依旧不对，感觉是没道理，感觉可能错误出在output上？循环体导致output处反向传播中断了？但是说不出原因。
希望聪明的你解答我的问题鸭。
