ValueError: tf.function-decorated function tried to create variables on non-first call while using C-CodePudding

I am trying to create a triplet loss function to calculate the similarity between two sentences as follows:

def TripletLoss(y_true,y_pred, margin=0.25,batch_size = 64):
    v1, v2 = y_pred[:,:128],y_pred[:,-128:]
    scores = K.dot(v1, K.transpose(v2))
    positive = tf.linalg.diag_part(scores)
    negative_without_positive = scores - 2 * K.eye(batch_size)

    closest_negative = tf.reduce_max(negative_without_positive, axis=1)

    negative_zero_on_duplicate = scores * (1.0 - K.eye(batch_size))
    
    mean_negative = K.sum(negative_zero_on_duplicate, axis=1) / (batch_size-1)
    
    triplet_loss1 = K.maximum(0.0, margin - positive   closest_negative)
    
    triplet_loss2 = K.maximum(0.0, margin - positive   mean_negative)
    
    triplet_loss = K.mean(triplet_loss1   triplet_loss2)

    return triplet_loss

My Model is as follows:

input1 = keras.Input(shape=(train_data1.shape[1],))
input2 = keras.Input(shape=(train_data1.shape[1],))

encoding1 = base_model(input1)
encoding2 = base_model(input2)

merged = layers.Concatenate()([encoding1, encoding2])

model = models.Model(inputs = [input1, input2], outputs = merged)

where the base model is:

def calculate_mean(x, axis=1):
    return K.mean(x, axis=axis)

def normalize(x):
        return x / K.sqrt(K.sum(x * x, axis=-1, keepdims=True))

base_model = models.Sequential()
base_model.add(layers.Embedding(input_dim=len(vocab) 2, output_dim=128))
base_model.add(layers.LSTM(128, return_sequences=True))
base_model.add(layers.Lambda(calculate_mean, name='mean'))
base_model.add(layers.Lambda(normalize, name='normalize'))

Now when I use that loss function to compile the model with

model.compile(
    optimizer = Adam(0.001),
    loss = TripletLoss
)

It does not give any error. But when I train it using the fit method it gives me errors as:

ValueError: tf.function-decorated function tried to create variables on non-first call.

if I use other losses it works perfectly. I don't know what is wrong with the loss function here.

CodePudding user response：

Maybe try the following with tf.eye:

import tensorflow as tf

from tensorflow.keras import backend as K

def TripletLoss(margin=0.25):
    def triplet(y_true,y_pred):
      batch_size = tf.cast(tf.shape(y_true)[0], dtype=tf.float32)
      v1, v2 = y_pred[:,:128],y_pred[:,-128:]
      scores = K.dot(v1, K.transpose(v2))
      positive = tf.linalg.diag_part(scores)
      negative_without_positive = scores - 2 * tf.eye(batch_size)

      closest_negative = tf.reduce_max(negative_without_positive, axis=1)

      negative_zero_on_duplicate = scores * (1.0 - tf.eye(batch_size))
      
      mean_negative = K.sum(negative_zero_on_duplicate, axis=1) / (batch_size-1)
      
      triplet_loss1 = K.maximum(0.0, margin - positive   closest_negative)
      
      triplet_loss2 = K.maximum(0.0, margin - positive   mean_negative)
      
      triplet_loss = K.mean(triplet_loss1   triplet_loss2)

      return triplet_loss
    return triplet

triplet_loss = TripletLoss()

def calculate_mean(x, axis=1):
    return K.mean(x, axis=axis)

def normalize(x):
        return x / K.sqrt(K.sum(x * x, axis=-1, keepdims=True))

base_model = tf.keras.Sequential()
base_model.add(tf.keras.layers.Embedding(input_dim=50, output_dim=128))
base_model.add(tf.keras.layers.LSTM(128, return_sequences=True))
base_model.add(tf.keras.layers.Lambda(calculate_mean, name='mean'))
base_model.add(tf.keras.layers.Lambda(normalize, name='normalize'))

input1 = tf.keras.layers.Input(shape=(50,))
input2 = tf.keras.layers.Input(shape=(50,))

encoding1 = base_model(input1)
encoding2 = base_model(input2)

merged = tf.keras.layers.Concatenate()([encoding1, encoding2])

model = tf.keras.Model(inputs = [input1, input2], outputs = merged)
model.compile(
    optimizer = tf.keras.optimizers.Adam(0.001),
    loss = triplet_loss
)

x = tf.random.uniform((500, 50), maxval=50, dtype=tf.int32)
y = tf.random.uniform((500, 256))
model.fit([x, x], y, epochs=2, batch_size=64)

Epoch 1/2
8/8 [==============================] - 6s 237ms/step - loss: 0.0037
Epoch 2/2
8/8 [==============================] - 2s 233ms/step - loss: 5.4691e-04
<keras.callbacks.History at 0x7fd249072d50>