ValueError: No gradients provided for any variable: ['embedding_2/embeddings:0', 'bidirectional_4/forward_lstm_4/lstm_cell_13/kernel:0', 'bidirectional_4/forward_lstm_4/lstm_cell_13/recurrent_kernel:0', 'bidirectional_4/forward_lstm_4/lstm_cell_13/bias:0', 'bidirectional_4/backward_lstm_4/lstm_cell_14/kernel:0', 'bidirectional_4/backward_lstm_4/lstm_cell_14/recurrent_kernel:0', 'bidirectional_4/backward_lstm_4/lstm_cell_14/bias:0', 'bidirectional_5/forward_lstm_5/lstm_cell_16/kernel:0', 'bidirectional_5/forward_lstm_5/lstm_cell_16/recurrent_kernel:0', 'bidirectional_5/forward_lstm_5/lstm_cell_16/bias:0', 'bidirectional_5/backward_lstm_5/lstm_cell_17/kernel:0', 'bidirectional_5/backward_lstm_5/lstm_cell_17/recurrent_kernel:0', 'bidirectional_5/backward_lstm_5/lstm_cell_17/bias:0', 'dense_2/kernel:0', 'dense_2/bias:0'].

[RavitejaBadugu]

I am training model in colab. The model I'm trying to build is siamese network for text similarity.
the loss function which I used finds internally the negative sample.
Explaination of loss function::
kindly, watch this two small videos where the lecturer explains about the loss function I used::
https://www.coursera.org/lecture/sequence-models-in-nlp/computing-the-cost-i-T4Ylj
https://www.coursera.org/lecture/sequence-models-in-nlp/computing-the-cost-ii-qXOjN
the dataset can be found here::
https://www.kaggle.com/c/quora-question-pairs/data
my colab notebook::
https://colab.research.google.com/drive/1NCUxSS9fiuLpPd2hOSxKv5QKH2IUuhX2?usp=sharing

sub_model=tf.keras.models.Sequential([Embedding(vocab_size,300,input_length=79), 
Bidirectional(LSTM(79,return_sequences=True)),
Bidirectional(LSTM(79,return_sequences=True)),
tf.keras.layers.GlobalAveragePooling1D(),
tf.keras.layers.Dense(units=158)])
ins1=Input((79,),name='input1')
ins2=Input((79,),name='input2')
x1=sub_model(ins1)
x2=sub_model(ins2)
norm1=tf.keras.layers.Layer(lambda x: tf.math.l2_normalize(x,axis=1))(x1)
norm2=tf.keras.layers.Layer(lambda x: tf.math.l2_normalize(x,axis=1))(x2)

model=Model([ins1,ins2],[norm1,norm2])

the loss function used is

def get_tripletloss(y_pred1,y_pred2):
    y_pred=tf.matmul(y_pred1,y_pred2,transpose_b=True) ##getting y_pred of (batch,batch)
    batch=y_pred.get_shape().as_list()[0] ##getting batch_size
    alpha_matrix=tf.cast(tf.reshape(tf.repeat(0.2,repeats=batch),shape=(batch,1)),dtype=tf.float32) #making alpha matrix of 0.2's
    diag_part=tf.cast(tf.reshape(tf.linalg.diag_part(y_pred),shape=(batch,1)),dtype=tf.float32) ##taking diag_part
    diagonal_matrix=tf.cast(tf.linalg.diag(tf.linalg.diag_part(y_pred)),dtype=tf.float32) ## making as diagonal_matrix
    sim_an=tf.reshape(tf.reduce_mean(tf.cast(y_pred,dtype=tf.float32)-diagonal_matrix,axis=1),shape=(batch,1)) ## getting only off-diagonal
    sim_an=tf.cast(sim_an,dtype=tf.float32)-diag_part+alpha_matrix ## getting sim_an-sim_ap+alpha
    sim_an=tf.maximum(sim_an,tf.cast(tf.zeros((batch,1)),dtype=tf.float32)) # getting max(loss,0)
    loss1=tf.keras.backend.mean(sim_an) ##final_loss1
    ##########
    y_pred=tf.where(y_pred<diag_part,y_pred,tf.cast(0.000001,dtype=tf.float32))##made to small number where off-diagonal elements are getter than diagonal and also diagonal elements.
    sim_an2=tf.reshape(tf.reduce_max(y_pred,axis=1),shape=(batch,1)) ##getting max value(closet_negative) 
    loss2=tf.cast(sim_an2,dtype=tf.float32)-diag_part+alpha_matrix ##getting sim_an-sim_ap+alpha
    loss2=tf.maximum(loss2,tf.cast(tf.zeros((batch,1)),dtype=tf.float32)) ## max(loss2,0)
    loss2=tf.keras.backend.mean(loss2)
    loss=loss1+loss2
    return loss

and the training loop is

dataset=tf.data.Dataset.from_tensor_slices((seq1,seq2))
dataset=dataset.shuffle(149263)
dataset=dataset.batch(29,drop_remainder=True)
dataset=dataset.prefetch(tf.data.experimental.AUTOTUNE)

@tf.function
def get_grads(v1,v2):
    with tf.GradientTape() as tape:
        loss=get_tripletloss(v1,v2)
        grads=tape.gradient(loss,model.trainable_variables)
    return loss,grads
loss_per_batch=[]
loss_per_epoch=[]
optimizer=tf.keras.optimizers.Adam()
for j in range(2):
    training_batches=seq1.shape[0]//29
    for i in range(training_batches):
        data=next(iter(dataset))
        v1,v2=model(data)
        loss,grads=get_grads(v1,v2)
        loss_per_batch.append(loss)
        optimizer.apply_gradients(zip(grads,model.trainable_variables))
    loss_per_epoch.append(np.mean(loss_per_batch))
    loss_per_batch=[]
    print(f"finsihed {j+1} epoch got loss of {loss_per_epoch[j]}")

the error is ::

ValueError                                Traceback (most recent call last)
<ipython-input-98-24aaeac0e75b> in <module>()
      9         loss,grads=get_grads(v1,v2)
     10         loss_per_batch.append(loss)
---> 11         optimizer.apply_gradients(zip(grads,model.trainable_variables))
     12     loss_per_epoch.append(np.mean(loss_per_batch))
     13     loss_per_batch=[]

1 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py in apply_gradients(self, grads_and_vars, name, experimental_aggregate_gradients)
    511       ValueError: If none of the variables have gradients.
    512     """
--> 513     grads_and_vars = _filter_grads(grads_and_vars)
    514     var_list = [v for (_, v) in grads_and_vars]
    515 

/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py in _filter_grads(grads_and_vars)
   1269   if not filtered:
   1270     raise ValueError("No gradients provided for any variable: %s." %
-> 1271                      ([v.name for _, v in grads_and_vars],))
   1272   if vars_with_empty_grads:
   1273     logging.warning(

ValueError: No gradients provided for any variable: ['embedding_2/embeddings:0', 'bidirectional_4/forward_lstm_4/lstm_cell_13/kernel:0', 'bidirectional_4/forward_lstm_4/lstm_cell_13/recurrent_kernel:0', 'bidirectional_4/forward_lstm_4/lstm_cell_13/bias:0', 'bidirectional_4/backward_lstm_4/lstm_cell_14/kernel:0', 'bidirectional_4/backward_lstm_4/lstm_cell_14/recurrent_kernel:0', 'bidirectional_4/backward_lstm_4/lstm_cell_14/bias:0', 'bidirectional_5/forward_lstm_5/lstm_cell_16/kernel:0', 'bidirectional_5/forward_lstm_5/lstm_cell_16/recurrent_kernel:0', 'bidirectional_5/forward_lstm_5/lstm_cell_16/bias:0', 'bidirectional_5/backward_lstm_5/lstm_cell_17/kernel:0', 'bidirectional_5/backward_lstm_5/lstm_cell_17/recurrent_kernel:0', 'bidirectional_5/backward_lstm_5/lstm_cell_17/bias:0', 'dense_2/kernel:0', 'dense_2/bias:0'].

[Saduf2019]

@RavitejaBadugu
The colab note book shared, is empty, please update with complete code and error.
Please refer to these issues with same error: #42038 link,

[RavitejaBadugu]

@Saduf2019
that's not a gist. just copy the link and paste it in a new tab. It's not empty.

[RavitejaBadugu]

How can we create a loss function which doesn't take y_true. But, My loss function which I showed in the issue calculates the loss without y_true. (for information check the links)In documentation of tf, I found that when we are writing a custom loss function, then
we need to define as
def loss(y_true,t_pred, any smoothing_parameter):

So, the function is expecting us to give y_true, the logic in my loss function is correct. My doubt is how to define without y_true. May be it is the one causing the problem. I think! but don't know. Please! clarify this doubt @gowthamkpr

[RavitejaBadugu]

I got the solution. I did a small mistake in my code.

[KrishPro]

@RavitejaBadugu

I got the solution. I did a small mistake in my code.

Can you tell me that mistake because i'm also getting same mistake

[yhetman]

@RavitejaBadugu
Also have same mistake.. i would appreciate if you tell how you solved the mistake

[RavitejaBadugu]

In custom training loop within tape we need to mention model(data) if we mention it before I got the error.

loss_per_epoch=[]
optimizer=tf.keras.optimizers.Adam()
for j in range(2):
    training_batches=x_train.shape[0]//12
    for i in tqdm(range(training_batches)):
        data=next(iter(dataset))
        outputs=model(data)## it is mentioned outside the tape
        with tf.GradientTape() as tape:
          losses=loss(outputs.get('out1'),outputs.get('out2'),0.3,-2.0)
        grads=tape.gradient(losses,model.trainable_variables)
        loss_per_batch.append(loss)
        optimizer.apply_gradients(zip(grads,model.trainable_variables))
    loss_per_epoch.append(np.mean(loss_per_batch))
    loss_per_batch=[]
    print(f"finsihed {j+1} epoch got loss of {loss_per_epoch[j]}")

getting error as

ValueError: No gradients provided for any variable: ['tf_roberta_model/roberta/encoder/layer_._0/attention/self/query/kernel:0', 'tf_roberta_model/roberta/encoder/layer_._0/attention/self/query/bias:0', 'tf_roberta_model/roberta/encoder/layer_._0/attention/self/key/kernel:0', 'tf_roberta_model/roberta/encoder/layer_._0/attention/self/key/bias:0', 'tf_roberta_model/roberta/encoder/layer_._0/attention/self/value/kernel:0', 'tf_roberta_model/roberta/encoder/layer_._0/attention/self/value/bias:0', 'tf_roberta_model/roberta/encoder/layer_._0/attention/o....

but when I mention the model within the tape the error is gone

loss_per_epoch=[]
optimizer=tf.keras.optimizers.Adam()
for j in range(2):
    training_batches=x_train.shape[0]//12
    for i in tqdm(range(training_batches)):
        data=next(iter(dataset))
        with tf.GradientTape() as tape:
          outputs=model(data)###################mentioned here
          losses=loss(outputs.get('out1'),outputs.get('out2'),0.3,-2.0)
        grads=tape.gradient(losses,model.trainable_variables)
        loss_per_batch.append(loss)
        optimizer.apply_gradients(zip(grads,model.trainable_variables))
    loss_per_epoch.append(np.mean(loss_per_batch))
    loss_per_batch=[]
    print(f"finsihed {j+1} epoch got loss of {loss_per_epoch[j]}")