Python keras.backend.floatx() Examples

def labelembed_loss(out1, out2, tar, targets, tau = 2., alpha = 0.9, beta = 0.5, num_classes = 100):
    
    out2_prob = K.softmax(out2)
    tau2_prob = K.stop_gradient(K.softmax(out2 / tau))
    soft_tar = K.stop_gradient(K.softmax(tar))
    
    L_o1_y = K.sparse_categorical_crossentropy(output = K.softmax(out1), target = targets)
    
    pred = K.argmax(out2, axis = -1)
    mask = K.stop_gradient(K.cast(K.equal(pred, K.cast(targets, 'int64')), K.floatx()))
    L_o1_emb = -cross_entropy(out1, soft_tar)  # pylint: disable=invalid-unary-operand-type
    
    L_o2_y = K.sparse_categorical_crossentropy(output = out2_prob, target = targets)
    L_emb_o2 = -cross_entropy(tar, tau2_prob) * mask * (K.cast(K.shape(mask)[0], K.floatx())/(K.sum(mask)+1e-8))  # pylint: disable=invalid-unary-operand-type
    L_re = K.relu(K.sum(out2_prob * K.one_hot(K.cast(targets, 'int64'), num_classes), axis = -1) - alpha)
    
    return beta * L_o1_y + (1-beta) * L_o1_emb + L_o2_y + L_emb_o2 + L_re 

def call(self, x, mask=None):
        uit = dot_product(x, self.W)

        if self.bias:
            uit += self.b

        uit = K.tanh(uit)
        ait = dot_product(uit, self.u)

        a = K.exp(ait)

        # apply mask after the exp. will be re-normalized next
        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            a *= K.cast(mask, K.floatx())

        # 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())
        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        a = K.expand_dims(a)
        weighted_input = x * a
        return K.sum(weighted_input, axis=1) 

def call(self, inputs, mask=None, **kwargs):
        if isinstance(inputs, list):
            query, key, value = inputs
        else:
            query = key = value = inputs
        if isinstance(mask, list):
            mask = mask[1]
        feature_dim = K.shape(query)[-1]
        e = K.batch_dot(query, key, axes=2) / K.sqrt(K.cast(feature_dim, dtype=K.floatx()))
        e = K.exp(e - K.max(e, axis=-1, keepdims=True))
        if self.history_only:
            query_len, key_len = K.shape(query)[1], K.shape(key)[1]
            indices = K.tile(K.expand_dims(K.arange(key_len), axis=0), [query_len, 1])
            upper = K.expand_dims(K.arange(key_len), axis=-1)
            e *= K.expand_dims(K.cast(indices <= upper, K.floatx()), axis=0)
        if mask is not None:
            e *= K.cast(K.expand_dims(mask, axis=-2), K.floatx())
        a = e / (K.sum(e, axis=-1, keepdims=True) + K.epsilon())
        v = K.batch_dot(a, value)
        if self.return_attention:
            return [v, a]
        return v 

def call(self, x, mask=None):
        # computes a probability distribution over the timesteps
        # uses 'max trick' for numerical stability
        # reshape is done to avoid issue with Tensorflow
        # and 1-dimensional weights
        logits = K.dot(x, self.W)
        x_shape = K.shape(x)
        logits = K.reshape(logits, (x_shape[0], x_shape[1]))
        ai = K.exp(logits - K.max(logits, axis=-1, keepdims=True))

        # masked timesteps have zero weight
        if mask is not None:
            mask = K.cast(mask, K.floatx())
            ai = ai * mask
        att_weights = ai / (K.sum(ai, axis=1, keepdims=True) + K.epsilon())
        weighted_input = x * K.expand_dims(att_weights)
        result = K.sum(weighted_input, axis=1)
        if self.return_attention:
            return [result, att_weights]
        return result 

def call(self, x, mask=None):
        # computes a probability distribution over the timesteps
        # uses 'max trick' for numerical stability
        # reshape is done to avoid issue with Tensorflow
        # and 1-dimensional weights
        logits = K.dot(x, self.W)
        x_shape = K.shape(x)
        logits = K.reshape(logits, (x_shape[0], x_shape[1]))
        ai = K.exp(logits - K.max(logits, axis=-1, keepdims=True))

        # masked timesteps have zero weight
        if mask is not None:
            mask = K.cast(mask, K.floatx())
            ai = ai * mask
        att_weights = ai / (K.sum(ai, axis=1, keepdims=True) + K.epsilon())
        weighted_input = x * K.expand_dims(att_weights)
        result = K.sum(weighted_input, axis=1)
        if self.return_attention:
            return [result, att_weights]
        return result 

def add_boundary_energy(x, b_start=None, b_end=None, mask=None):
    '''Given the observations x, it adds the start boundary energy b_start (resp.
    end boundary energy b_end on the start (resp. end) elements and multiplies
    the mask.'''
    if mask is None:
        if b_start is not None:
            x = K.concatenate([x[:, :1, :] + b_start, x[:, 1:, :]], axis=1)
        if b_end is not None:
            x = K.concatenate([x[:, :-1, :], x[:, -1:, :] + b_end], axis=1)
    else:
        mask = K.cast(mask, K.floatx())
        mask = K.expand_dims(mask, 2)
        x *= mask
        if b_start is not None:
            mask_r = K.concatenate([K.zeros_like(mask[:, :1]), mask[:, :-1]], axis=1)
            start_mask = K.cast(K.greater(mask, mask_r), K.floatx())
            x = x + start_mask * b_start
        if b_end is not None:
            mask_l = K.concatenate([mask[:, 1:], K.zeros_like(mask[:, -1:])], axis=1)
            end_mask = K.cast(K.greater(mask, mask_l), K.floatx())
            x = x + end_mask * b_end
    return x 

def call(self,x,mask=None):
        conv_input,theta = x
        s = theta.shape
        theta = T.reshape(theta,[-1,s[2]])
        m = K.not_equal(conv_input,0.)

        #### For translation
        trans = _trans(theta)
        output = _transform_trans(trans, conv_input)
        output = output * K.cast(m,K.floatx())

        ### For rotation
        M = _fusion(theta)
        output = _transform_rot(M,output)

        return output 

def _forward(x, reduce_step, initial_states, U, mask=None):
    '''Forward recurrence of the linear chain crf.'''

    def _forward_step(energy_matrix_t, states):
        alpha_tm1 = states[-1]
        new_states = reduce_step(K.expand_dims(alpha_tm1, 2) + energy_matrix_t)
        return new_states[0], new_states

    U_shared = K.expand_dims(K.expand_dims(U, 0), 0)

    if mask is not None:
        mask = K.cast(mask, K.floatx())
        mask_U = K.expand_dims(K.expand_dims(mask[:, :-1] * mask[:, 1:], 2), 3)
        U_shared = U_shared * mask_U

    inputs = K.expand_dims(x[:, 1:, :], 2) + U_shared
    inputs = K.concatenate([inputs, K.zeros_like(inputs[:, -1:, :, :])], axis=1)

    last, values, _ = K.rnn(_forward_step, inputs, initial_states)
    return last, values 

def create_transformer(embedding_dim: int = 768, embedding_dropout: float = 0.1,
                       vocab_size: int = 30000, max_len: int = 512,
                       trainable_pos_embedding: bool = True, num_heads: int = 12, num_layers: int = 12,
                       attention_dropout: float = 0.1, use_one_embedding_dropout: bool = False,
                       d_hid: int = 768 * 4, residual_dropout: float = 0.1,
                       use_attn_mask: bool = True) -> keras.Model:
    vocab_size += TextEncoder.SPECIAL_COUNT
    tokens = Input(batch_shape=(None, max_len), name='token_input', dtype='int32')
    segment_ids = Input(batch_shape=(None, max_len), name='segment_input', dtype='int32')
    pos_ids = Input(batch_shape=(None, max_len), name='position_input', dtype='int32')
    attn_mask = Input(batch_shape=(None, 1, max_len, max_len), name='attention_mask_input',
                      dtype=K.floatx()) if use_attn_mask else None
    inputs = [tokens, segment_ids, pos_ids]
    embedding_layer = Embedding(embedding_dim, embedding_dropout, vocab_size, max_len, trainable_pos_embedding,
                                use_one_embedding_dropout)
    x = embedding_layer(inputs)
    for i in range(num_layers):
        x = EncoderLayer(embedding_dim, num_heads, d_hid, residual_dropout,
                         attention_dropout, use_attn_mask, i)(x, attn_mask)
    inputs = inputs + ([attn_mask] if use_attn_mask else [])
    return keras.Model(inputs=inputs, outputs=x, name='Transformer') 

def next(self):
        # Keeps under lock only the mechanism which advances
        # the indexing of each batch.
        with self.lock:
            index_array, current_index, current_batch_size = next(self.index_generator)
        # The transformation of images is not under thread lock
        # so it can be done in parallel
        batch_x = np.zeros(tuple([current_batch_size] + list(self.image_size)), dtype=K.floatx())
        for i, j in enumerate(index_array):
            x = scipy.misc.imread(self.x[j])
            x = scipy.misc.imresize(x, self.image_size)
            x = self.image_data_generator.random_transform(x.astype(K.floatx()))
            x = self.image_data_generator.standardize(x)
            batch_x[i] = x
        if self.save_to_dir:
            for i in range(current_batch_size):
                img = image.array_to_img(batch_x[i], self.data_format, scale=True)
                fname = '{prefix}_{index}_{hash}.{format}'.format(prefix=self.save_prefix,
                                                                  index=current_index + i,
                                                                  hash=np.random.randint(1e4),
                                                                  format=self.save_format)
                img.save(os.path.join(self.save_to_dir, fname))
        batch_y = self.y[index_array]
        return batch_x, batch_y 

def gen_cosine_amp(amp=100, period=1000, x0=0, xn=50000, step=1, k=0.0001):
	"""Generates an absolute cosine time series with the amplitude
	exponentially decreasing

	Arguments:
	    amp: amplitude of the cosine function
	    period: period of the cosine function
	    x0: initial x of the time series
	    xn: final x of the time series
	    step: step of the time series discretization
	    k: exponential rate
	"""
	cos = np.zeros(((xn - x0) * step, 1, 1), dtype=K.floatx())
	for i in range(len(cos)):
		idx = x0 + i * step
		cos[i, 0, 0] = amp * np.cos(2 * np.pi * idx / period)
		cos[i, 0, 0] = cos[i, 0, 0] * np.exp(-k * idx)
	return cos 

def call(self, x, mask=None):
		if self.mode == 'maximum_likelihood':
			# draw maximum likelihood sample from Bernoulli distribution
			#    x* = argmax_x p(x) = 1         if p(x=1) >= 0.5
			#                         0         otherwise
			return K.round(x)
		elif self.mode == 'random':
			# draw random sample from Bernoulli distribution
			#    x* = x ~ p(x) = 1              if p(x=1) > uniform(0, 1)
			#                    0              otherwise
			#return self.srng.binomial(size=x.shape, n=1, p=x, dtype=K.floatx())
			return K.random_binomial(x.shape, p=x, dtype=K.floatx())
		elif self.mode == 'mean_field':
			# draw mean-field approximation sample from Bernoulli distribution
			#    x* = E[p(x)] = E[Bern(x; p)] = p
			return x
		elif self.mode == 'nrlu':
			return nrlu(x)
		else:
			raise NotImplementedError('Unknown sample mode!') 

def build_masked_loss(loss_function, mask_value):
    """Builds a loss function that masks based on targets

    Args:
        loss_function: The loss function to mask
        mask_value: The value to mask in the targets

    Returns:
        function: a loss function that acts like loss_function with masked inputs
    """

    def masked_loss_function(y_true, y_pred):
        mask = K.cast(K.not_equal(y_true, mask_value), K.floatx())
        return loss_function(y_true * mask, y_pred * mask)

    return masked_loss_function 

def add_boundary_energy(x, b_start=None, b_end=None, mask=None):
    '''Given the observations x, it adds the start boundary energy b_start (resp.
    end boundary energy b_end on the start (resp. end) elements and multiplies
    the mask.'''
    if mask is None:
        if b_start is not None:
            x = K.concatenate([x[:, :1, :] + b_start, x[:, 1:, :]], axis=1)
        if b_end is not None:
            x = K.concatenate([x[:, :-1, :], x[:, -1:, :] + b_end], axis=1)
    else:
        mask = K.cast(mask, K.floatx())
        mask = K.expand_dims(mask, 2)
        x *= mask
        if b_start is not None:
            mask_r = K.concatenate([K.zeros_like(mask[:, :1]), mask[:, :-1]], axis=1)
            start_mask = K.cast(K.greater(mask, mask_r), K.floatx())
            x = x + start_mask * b_start
        if b_end is not None:
            mask_l = K.concatenate([mask[:, 1:], K.zeros_like(mask[:, -1:])], axis=1)
            end_mask = K.cast(K.greater(mask, mask_l), K.floatx())
            x = x + end_mask * b_end
    return x 

def conv_kernel_initializer(shape, dtype=K.floatx()):
    """Initialization for convolutional kernels.
    The main difference with tf.variance_scaling_initializer is that
    tf.variance_scaling_initializer uses a truncated normal with an uncorrected
    standard deviation, whereas here we use a normal distribution. Similarly,
    tf.contrib.layers.variance_scaling_initializer uses a truncated normal with
    a corrected standard deviation.
    Args:
        shape: shape of variable
        dtype: dtype of variable
    Returns:
        an initialization for the variable
    """
    kernel_height, kernel_width, _, out_filters = shape
    fan_out = int(kernel_height * kernel_width * out_filters)
    return tf.random_normal(
        shape, mean=0.0, stddev=np.sqrt(2.0 / fan_out), dtype=dtype) 

def call(self, x, mask=None):
        # size of x :[batch_size, sel_len, attention_dim]
        # size of u :[batch_size, attention_dim]
        # uit = tanh(xW+b)
        uit = K.tanh(K.bias_add(K.dot(x, self.W), self.b))
        ait = K.dot(uit, self.u)
        ait = K.squeeze(ait, -1)

        ait = K.exp(ait)

        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            ait *= K.cast(mask, K.floatx())
        ait /= K.cast(K.sum(ait, axis=1, keepdims=True) + K.epsilon(), K.floatx())
        ait = K.expand_dims(ait)
        weighted_input = x * ait
        output = K.sum(weighted_input, axis=1)

        return output 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = self.learning_rate * (K.sqrt(1. - K.pow(self.beta_2, t)) / (1. - K.pow(self.beta_1, t)))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            p_t = lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            self.updates.append(K.update_sub(p, p_t))
        return self.updates 

def call(self, x, mask=None):
        uit = dot_product(x, self.W)

        if self.bias:
            uit += self.b

        uit = K.tanh(uit)
        ait = dot_product(uit, self.u)

        a = K.exp(ait)

        # apply mask after the exp. will be re-normalized next
        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            a *= K.cast(mask, K.floatx())

        # 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())
        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        a = K.expand_dims(a)
        weighted_input = x * a
        return K.sum(weighted_input, axis=1) 

def call(self, x, mask=None):
        uit = dot_product(x, self.W)

        if self.bias:
            uit += self.b

        uit = K.tanh(uit)
        ait = dot_product(uit, self.u)

        a = K.exp(ait)

        # apply mask after the exp. will be re-normalized next
        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            a *= K.cast(mask, K.floatx())

        # 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())
        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        a = K.expand_dims(a)
        weighted_input = x * a
        return K.sum(weighted_input, axis=1) 

def call(self, x, mask=None):
        uit = dot_product(x, self.W)

        if self.bias:
            uit += self.b

        uit = K.tanh(uit)
        ait = dot_product(uit, self.u)

        a = K.exp(ait)

        # apply mask after the exp. will be re-normalized next
        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            a *= K.cast(mask, K.floatx())

        # 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())
        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        a = K.expand_dims(a)
        weighted_input = x * a
        return K.sum(weighted_input, axis=1) 

def call(self, x, mask=None):
        eij = dot_product(x, self.W)

        if self.bias:
            eij += self.b

        eij = K.tanh(eij)

        a = K.exp(eij)

        if mask is not None:
            a *= K.cast(mask, K.floatx())

        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        weighted_input = x * K.expand_dims(a)

        result = K.sum(weighted_input, axis=1)

        if self.return_attention:
            return [result, a]
        return result 

def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
                 epsilon=None, decay=0., amsgrad=False, accum_iters=1, **kwargs):
        if accum_iters < 1:
            raise ValueError('accum_iters must be >= 1')
        super(AdamAccumulate, self).__init__(**kwargs)
        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.lr = K.variable(lr, name='lr')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.decay = K.variable(decay, name='decay')
        if epsilon is None:
            epsilon = K.epsilon()
        self.epsilon = epsilon
        self.initial_decay = decay
        self.amsgrad = amsgrad
        self.accum_iters = K.variable(accum_iters, K.dtype(self.iterations))
        self.accum_iters_float = K.cast(self.accum_iters, K.floatx()) 

def call(self, x, mask=None):
        # computes a probability distribution over the timesteps
        # uses 'max trick' for numerical stability
        # reshape is done to avoid issue with Tensorflow
        # and 1-dimensional weights
        logits = K.dot(x, self.W)
        x_shape = K.shape(x)
        logits = K.reshape(logits, (x_shape[0], x_shape[1]))
        ai = K.exp(logits - K.max(logits, axis=-1, keepdims=True))

        # masked timesteps have zero weight
        if mask is not None:
            mask = K.cast(mask, K.floatx())
            ai = ai * mask
        att_weights = ai / (K.sum(ai, axis=1, keepdims=True) + K.epsilon())
        weighted_input = x * K.expand_dims(att_weights)
        result = K.sum(weighted_input, axis=1)
        if self.return_attention:
            return [result, att_weights]
        return result 

def _forward(x, reduce_step, initial_states, U, mask=None):
    '''Forward recurrence of the linear chain crf.'''

    def _forward_step(energy_matrix_t, states):
        alpha_tm1 = states[-1]
        new_states = reduce_step(K.expand_dims(alpha_tm1, 2) + energy_matrix_t)
        return new_states[0], new_states

    U_shared = K.expand_dims(K.expand_dims(U, 0), 0)

    if mask is not None:
        mask = K.cast(mask, K.floatx())
        mask_U = K.expand_dims(K.expand_dims(mask[:, :-1] * mask[:, 1:], 2), 3)
        U_shared = U_shared * mask_U

    inputs = K.expand_dims(x[:, 1:, :], 2) + U_shared
    inputs = K.concatenate([inputs, K.zeros_like(inputs[:, -1:, :, :])], axis=1)

    last, values, _ = K.rnn(_forward_step, inputs, initial_states)
    return last, values 

def create_transformer(embedding_dim: int = 768, embedding_dropout: float = 0.1, vocab_size: int = 30000,
                       max_len: int = 512, trainable_pos_embedding: bool = True, num_heads: int = 12,
                       num_layers: int = 12, attention_dropout: float = 0.1, use_one_embedding_dropout: bool = False,
                       d_hid: int = 768 * 4, residual_dropout: float = 0.1, use_attn_mask: bool = True,
                       embedding_layer_norm: bool = False, neg_inf: float = -1e9, layer_norm_epsilon: float = 1e-5,
                       accurate_gelu: bool = False) -> keras.Model:
    vocab_size += TextEncoder.SPECIAL_COUNT
    tokens = Input(batch_shape=(None, max_len), name='token_input', dtype='int32')
    segment_ids = Input(batch_shape=(None, max_len), name='segment_input', dtype='int32')
    pos_ids = Input(batch_shape=(None, max_len), name='position_input', dtype='int32')
    attn_mask = Input(batch_shape=(None, 1, max_len, max_len), name='attention_mask_input',
                      dtype=K.floatx()) if use_attn_mask else None
    inputs = [tokens, segment_ids, pos_ids]
    embedding_layer = Embedding(embedding_dim, embedding_dropout, vocab_size, max_len, trainable_pos_embedding,
                                use_one_embedding_dropout, embedding_layer_norm, layer_norm_epsilon)
    x = embedding_layer(inputs)
    for i in range(num_layers):
        x = EncoderLayer(embedding_dim, num_heads, d_hid, residual_dropout,
                         attention_dropout, use_attn_mask, i, neg_inf, layer_norm_epsilon, accurate_gelu)(x, attn_mask)
    if use_attn_mask:
        inputs.append(attn_mask)
    return keras.Model(inputs=inputs, outputs=[x], name='Transformer') 

def load_img(img_path, target_size):
    """
    loads the image the same way darknet does, processes it and returns it (as array).
    uses PIL, like keras.preprocessing.image module.
    This loads image in RGB format.
    """
    from PIL import Image as pil_image
    import keras.backend as K
    import numpy as np
    img = pil_image.open(img_path)
    # TODO: check format and convert to RGB
    #resize
    x = np.asarray(img, dtype=K.floatx())/255.0
    #print(x[0,0,0], x[1,0,0], x[0,1,0], x[1,1,0], img.mode)
    x = letterbox_image(x, target_size)
    return x 

def masked_accuracy(y_true, y_pred):
    a = K.sum(K.cast(K.equal(y_true, K.round(y_pred)), K.floatx()))
    c = K.sum(K.cast(K.not_equal(y_true, 0.5), K.floatx()))
    acc = (a) / c
    return acc 

def __call__(self, w):
        other_weights = K.cast(K.greater_equal(w, 0)[:-1], K.floatx())
        last_weight = K.cast(K.equal(K.reshape(w[-1, :], (1, K.shape(w)[1])), 0.), K.floatx())
        appended = K.concatenate([other_weights, last_weight], axis=0)
        w *= appended
        return w 

def _get_accuracy(args, null_token_value):
    y_pred, y_true = args

    y_true = K.cast(y_true, "int32")
    mask = 1.0 - K.cast(K.equal(y_true, null_token_value), K.floatx())

    y_pred = K.cast(K.argmax(y_pred, axis=-1), "int32")
    correct = K.cast(
        K.equal(y_pred, y_true),
        K.floatx()
    )
    correct = K.sum(correct * mask, -1) / K.sum(mask, -1)

    return K.mean(correct) 

def __call__(self, w):
        other_weights = K.cast(K.greater_equal(w, 0)[:-1], K.floatx())
        last_weight = K.cast(K.equal(K.reshape(w[-1, :], (1, K.shape(w)[1])), 0.), K.floatx())
        appended = K.concatenate([other_weights, last_weight], axis=0)
        w *= appended
        return w