Keras Layer that implements an Attention mechanism, with a context/query vector, for temporal data. Supports Masking. Follows the work of Yang et al. [https://www.cs.cmu.edu/~diyiy/docs/naacl16.pdf] "Hierarchical Attention Networks for Document Classification"

AttentionWithContext.py

class AttentionWithContext(Layer):
    """
        Attention operation, with a context/query vector, for temporal data.
        Supports Masking.
        Follows the work of Yang et al. [https://www.cs.cmu.edu/~diyiy/docs/naacl16.pdf]
        "Hierarchical Attention Networks for Document Classification"
        by using a context vector to assist the attention
        # Input shape
            3D tensor with shape: `(samples, steps, features)`.
        # Output shape
            2D tensor with shape: `(samples, features)`.
        :param kwargs:
        Just put it on top of an RNN Layer (GRU/LSTM/SimpleRNN) with return_sequences=True.
        The dimensions are inferred based on the output shape of the RNN.
        Example:
            model.add(LSTM(64, return_sequences=True))
            model.add(AttentionWithContext())
        """

    def __init__(self, init='glorot_uniform', kernel_regularizer=None, bias_regularizer=None, kernel_constraint=None, bias_constraint=None,  **kwargs):
        self.supports_masking = True
        self.init = initializers.get(init)
        self.kernel_initializer = initializers.get('glorot_uniform')

        self.kernel_regularizer = regularizers.get(kernel_regularizer)
        self.bias_regularizer = regularizers.get(bias_regularizer)

        self.kernel_constraint = constraints.get(kernel_constraint)
        self.bias_constraint = constraints.get(bias_constraint)

        super(AttentionWithContext, self).__init__(**kwargs)

    def build(self, input_shape):
        self.kernel = self.add_weight((input_shape[-1], 1),
                                 initializer=self.kernel_initializer,
                                 name='{}_W'.format(self.name),
                                 regularizer=self.kernel_regularizer,
                                 constraint=self.kernel_constraint)
        self.b = self.add_weight((input_shape[1],),
                                 initializer='zero',
                                 name='{}_b'.format(self.name),
                                 regularizer=self.bias_regularizer,
                                 constraint=self.bias_constraint)

        self.u = self.add_weight((input_shape[1],),
                                 initializer=self.kernel_initializer,
                                 name='{}_u'.format(self.name),
                                 regularizer=self.kernel_regularizer,
                                 constraint=self.kernel_constraint)
        self.built = True

    def compute_mask(self, input, mask):
        return None

    def call(self, x, mask=None):
        # (x, 40, 300) x (300, 1)
        multData =  K.dot(x, self.kernel) # (x, 40, 1)
        multData = K.squeeze(multData, -1) # (x, 40)
        multData = multData + self.b # (x, 40) + (40,)

        multData = K.tanh(multData) # (x, 40)

        multData = multData * self.u # (x, 40) * (40, 1) => (x, 1)
        multData = K.exp(multData) # (X, 1)

        # apply mask after the exp. will be re-normalized next
        if mask is not None:
            mask = K.cast(mask, K.floatx()) #(x, 40)
            multData = mask*multData #(x, 40) * (x, 40, )

        # in some cases especially in the early stages of training the sum may be almost zero
        # and this results in NaN's. A workaround is to add a very small positive number ε to the sum.
        # a /= K.cast(K.sum(a, axis=1, keepdims=True), K.floatx())
        multData /= K.cast(K.sum(multData, axis=1, keepdims=True) + K.epsilon(), K.floatx())
        multData = K.expand_dims(multData)
        weighted_input = x * multData
        return K.sum(weighted_input, axis=1)


    def compute_output_shape(self, input_shape):
        return (input_shape[0], input_shape[-1],)

@rmdort Despite adding the Epsilon, I am still getting NaN loss at the start of training. Any reccomended debugging paths?

@rmdort, like your pedagogical comments of tensor shape changes. Those are really helpful for better understanding the math behind the context attention.

I dont know why, but, getting dimension error.
Code:

def generate_model(output_len, chars=None):
    """Generate the model"""
    print('Building model...')
    chars = chars or CHARS

    in_out_neurons = CONFIG.max_input_len  
    hidden_neurons = CONFIG.hidden_size
    
    model = Sequential()  
    model.add(recurrent.GRU(512, input_shape=( 128, 100), 
                             return_sequences=True, 
                            kernel_initializer=CONFIG.initialization, activation='linear'))
    model.add(AttentionWithContext())
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])
    return model

and the error is

________________________________
_________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
gru_1 (GRU)                  (None, 128, 512)          941568    
_________________________________________________________________
attention_with_context_1 (At (None, 512)               263168    
=================================================================
Total params: 1,204,736
Trainable params: 1,204,736
Non-trainable params: 0
_________________________________________________________________
None
Epoch 1/500
Traceback (most recent call last):
  File "test.py", line 580, in <module>
    train_speller()
  File "test.py", line 482, in train_speller
    itarative_train(model)
  File "test.py", line 467, in itarative_train
    class_weight=None, max_queue_size=10, workers=1)
  File "/ssd/anaconda3/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/ssd/anaconda3/lib/python3.6/site-packages/keras/models.py", line 1315, in fit_generator
    initial_epoch=initial_epoch)
  File "/ssd/anaconda3/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/ssd/anaconda3/lib/python3.6/site-packages/keras/engine/training.py", line 2230, in fit_generator
    class_weight=class_weight)
  File "/ssd/anaconda3/lib/python3.6/site-packages/keras/engine/training.py", line 1877, in train_on_batch
    class_weight=class_weight)
  File "/ssd/anaconda3/lib/python3.6/site-packages/keras/engine/training.py", line 1480, in _standardize_user_data
    exception_prefix='target')
  File "/ssd/anaconda3/lib/python3.6/site-packages/keras/engine/training.py", line 113, in _standardize_input_data
    'with shape ' + str(data_shape))
ValueError: Error when checking target: expected attention_with_context_1 to have 2 dimensions, but got array with shape (64, 128, 100)

Any idea?

as output share is anyway 3dim, i tried to change line 81 as

return (input_shape[0], input_shape[1],input_shape[2])

then for different error and model is not compiling

How can I add attention in functional api