Python keras.backend.greater() Examples

def set_reset_mem(self, mem, spikes):
        """
        Reset membrane potential ``mem`` array where ``spikes`` array is
        nonzero.
        """

        spike_idxs = k.T.nonzero(spikes)
        if (hasattr(self, 'activation_str') and
                self.activation_str == 'softmax'):
            new = mem.copy()  # k.T.set_subtensor(mem[spike_idxs], 0.)
        elif self.config.get('cell', 'reset') == 'Reset by subtraction':
            if self.payloads:  # Experimental.
                new = k.T.set_subtensor(mem[spike_idxs], 0.)
            else:
                pos_spike_idxs = k.T.nonzero(k.greater(spikes, 0))
                neg_spike_idxs = k.T.nonzero(k.less(spikes, 0))
                new = k.T.inc_subtensor(mem[pos_spike_idxs], -self.v_thresh)
                new = k.T.inc_subtensor(new[neg_spike_idxs], self.v_thresh)
        elif self.config.get('cell', 'reset') == 'Reset by modulo':
            new = k.T.set_subtensor(mem[spike_idxs],
                                    mem[spike_idxs] % self.v_thresh)
        else:  # self.config.get('cell', 'reset') == 'Reset to zero':
            new = k.T.set_subtensor(mem[spike_idxs], 0.)
        self.add_update([(self.mem, new)]) 

def zero_one_rank_loss_for_scores_ties(y_true, s_pred):
    y_true, s_pred = tensorify(y_true), tensorify(s_pred)
    n_objects = K.cast(K.max(y_true) + 1, dtype="float32")
    mask = K.greater(y_true[:, None] - y_true[:, :, None], 0)
    mask2 = K.greater(s_pred[:, None] - s_pred[:, :, None], 0)
    mask3 = K.equal(s_pred[:, None] - s_pred[:, :, None], 0)

    # Calculate Transpositions
    transpositions = tf.logical_and(mask, mask2)
    transpositions = K.sum(K.cast(transpositions, dtype="float32"), axis=[1, 2])
    transpositions += (
        K.sum(K.cast(mask3, dtype="float32"), axis=[1, 2]) - n_objects
    ) / 4.0

    denominator = n_objects * (n_objects - 1.0) / 2.0
    result = transpositions / denominator
    return K.mean(result) 

def zero_one_rank_loss(y_true, y_pred):
    y_true, y_pred = tensorify(y_true), tensorify(y_pred)
    mask = K.greater(y_true[:, None] - y_true[:, :, None], 0)
    # Count the number of mistakes (here position difference less than 0)
    mask2 = K.less(y_pred[:, None] - y_pred[:, :, None], 0)
    mask3 = K.equal(y_pred[:, None] - y_pred[:, :, None], 0)

    # Calculate Transpositions
    transpositions = tf.logical_and(mask, mask2)
    transpositions = K.sum(K.cast(transpositions, dtype="float32"), axis=[1, 2])

    n_objects = K.max(y_true) + 1
    transpositions += (
        K.sum(K.cast(mask3, dtype="float32"), axis=[1, 2]) - n_objects
    ) / 4.0
    denominator = K.cast((n_objects * (n_objects - 1.0)) / 2.0, dtype="float32")
    result = transpositions / denominator
    return K.mean(result) 

def crossentropy_max_origin_wrap(_m):
    def crossentropy_max_origin_core(y_true, y_pred):

        # hyper param
        print(_m)

        # 1) determine the origin of the patch, as a boolean vector y_true_flag
        # (True = patch from noisy subset)
        _y_true_flag = K.greater(K.sum(y_true, axis=-1), 90)

        # 2) convert the input y_true (with flags inside) into a valid y_true one-hot-vector format
        # attenuating factor for data points that need it (those that came with a one-hot of 100)
        _mask_reduce = K.cast(_y_true_flag, 'float32') * 0.01

        # identity factor for standard one-hot vectors
        _mask_keep = K.cast(K.equal(_y_true_flag, False), 'float32')

        # combine 2 masks
        _mask = _mask_reduce + _mask_keep

        _y_true_shape = K.shape(y_true)
        _mask = K.reshape(_mask, (_y_true_shape[0], 1))

        # applying mask to have a valid y_true that we can use as always
        y_true = y_true * _mask

        y_true = K.clip(y_true, K.epsilon(), 1)
        y_pred = K.clip(y_pred, K.epsilon(), 1)

        # compute loss for every data point
        _loss = -K.sum(y_true * K.log(y_pred), axis=-1)

        # threshold m
        t_m = K.max(_loss) * _m

        _mask_m = 1 - (K.cast(K.greater(_loss, t_m), 'float32') * K.cast(_y_true_flag, 'float32'))
        _loss = _loss * _mask_m

        return _loss
    return crossentropy_max_origin_core 

def binary_sigmoid_activation(self, mem):
        """Binary sigmoid activation."""

        return k.T.mul(k.greater(mem, 0), self.v_thresh) 

def binary_tanh_activation(self, mem):
        """Binary tanh activation."""

        output_spikes = k.T.mul(k.greater(mem, 0), self.v_thresh)
        output_spikes += k.T.mul(k.less(mem, 0), -self.v_thresh)

        return output_spikes 

def get_new_mem(self):
        """Add input to membrane potential."""

        # Destroy impulse if in refractory period
        masked_impulse = self.impulse if self.tau_refrac == 0 else \
            k.T.set_subtensor(
                self.impulse[k.T.nonzero(self.refrac_until > self.time)], 0.)

        # Add impulse
        if clamp_var:
            # Experimental: Clamp the membrane potential to zero until the
            # presynaptic neurons fire at their steady-state rates. This helps
            # avoid a transient response.
            new_mem = theano.ifelse.ifelse(
                k.less(k.mean(self.var), 1e-4) +
                k.greater(self.time, self.duration / 2),
                self.mem + masked_impulse, self.mem)
        elif hasattr(self, 'clamp_idx'):
            # Set clamp-duration by a specific delay from layer to layer.
            new_mem = theano.ifelse.ifelse(k.less(self.time, self.clamp_idx),
                                           self.mem, self.mem + masked_impulse)
        elif v_clip:
            # Clip membrane potential to prevent too strong accumulation.
            new_mem = k.clip(self.mem + masked_impulse, -3, 3)
        else:
            new_mem = self.mem + masked_impulse

        if self.config.getboolean('cell', 'leak'):
            # Todo: Implement more flexible version of leak!
            new_mem = k.T.inc_subtensor(
                new_mem[k.T.nonzero(k.T.gt(new_mem, 0))], -0.1 * self.dt)

        return new_mem 

def crossentropy_max_wrap(_m):
    def crossentropy_max_core(y_true, y_pred):
        """
        This function is based on the one proposed in
        Il-Young Jeong and Hyungui Lim, "AUDIO TAGGING SYSTEM FOR DCASE 2018: FOCUSING ON LABEL NOISE,
         DATA AUGMENTATION AND ITS EFFICIENT LEARNING", Tech Report, DCASE 2018
        https://github.com/finejuly/dcase2018_task2_cochlearai

        :param y_true:
        :param y_pred:
        :return:
        """

        # hyper param
        print(_m)
        y_pred = K.clip(y_pred, K.epsilon(), 1)

        # compute loss for every data point
        _loss = -K.sum(y_true * K.log(y_pred), axis=-1)

        # threshold
        t_m = K.max(_loss) * _m
        _mask_m = 1 - (K.cast(K.greater(_loss, t_m), 'float32'))
        _loss = _loss * _mask_m

        return _loss
    return crossentropy_max_core 

def get_new_thresh(self):
        """Get new threshhold."""

        thr_min = self._v_thresh / 100
        thr_max = self._v_thresh
        r_lim = 1 / self.dt
        return thr_min + (thr_max - thr_min) * self.max_spikerate / r_lim

        # return theano.ifelse.ifelse(
        #     k.equal(self.time / self.dt % settings['timestep_fraction'], 0) *
        #     k.greater(self.max_spikerate, settings['diff_to_min_rate']/1000)*
        #     k.greater(1 / self.dt - self.max_spikerate,
        #          settings['diff_to_max_rate'] / 1000),
        #     self.max_spikerate, self.v_thresh) 

def discriminator_loss(y_true, y_pred):
    loss = mean_squared_error(y_true, y_pred)
    is_large = k.greater(loss, k.constant(_disc_train_thresh)) # threshold
    is_large = k.cast(is_large, k.floatx())
    return loss * is_large # binary threshold the loss to prevent overtraining the discriminator 

def get_split_averages(input_tensor, input_mask, indices):
        # Splits input tensor into three parts based on the indices and
        # returns average of values prior to index, values at the index and
        # average of values after the index.
        # input_tensor: (batch_size, input_length, input_dim)
        # input_mask: (batch_size, input_length)
        # indices: (batch_size, 1)
        # (1, input_length)
        length_range = K.expand_dims(K.arange(K.shape(input_tensor)[1]), dim=0)
        # (batch_size, input_length)
        batched_range = K.repeat_elements(length_range, K.shape(input_tensor)[0], 0)
        tiled_indices = K.repeat_elements(indices, K.shape(input_tensor)[1], 1)  # (batch_size, input_length)
        greater_mask = K.greater(batched_range, tiled_indices)  # (batch_size, input_length)
        lesser_mask = K.lesser(batched_range, tiled_indices)  # (batch_size, input_length)
        equal_mask = K.equal(batched_range, tiled_indices)  # (batch_size, input_length)

        # We also need to mask these masks using the input mask.
        # (batch_size, input_length)
        if input_mask is not None:
            greater_mask = switch(input_mask, greater_mask, K.zeros_like(greater_mask))
            lesser_mask = switch(input_mask, lesser_mask, K.zeros_like(lesser_mask))

        post_sum = K.sum(switch(K.expand_dims(greater_mask), input_tensor, K.zeros_like(input_tensor)), axis=1)  # (batch_size, input_dim)
        pre_sum = K.sum(switch(K.expand_dims(lesser_mask), input_tensor, K.zeros_like(input_tensor)), axis=1)  # (batch_size, input_dim)
        values_at_indices = K.sum(switch(K.expand_dims(equal_mask), input_tensor, K.zeros_like(input_tensor)), axis=1)  # (batch_size, input_dim)

        post_normalizer = K.expand_dims(K.sum(greater_mask, axis=1) + K.epsilon(), dim=1)  # (batch_size, 1)
        pre_normalizer = K.expand_dims(K.sum(lesser_mask, axis=1) + K.epsilon(), dim=1)  # (batch_size, 1)

        return K.cast(pre_sum / pre_normalizer, 'float32'), values_at_indices, K.cast(post_sum / post_normalizer, 'float32') 

def acc_class1(y_true, y_pred):
    """ Function to estimate accuracy over the class 1 prediction. This estimation is global (i.e. abstaining samples are not removed)
    
    Parameters
    ----------
    y_true : keras tensor
        True values to predict
    y_pred : keras tensor
        Prediction made by the model. It is assumed that this keras tensor includes extra columns to store the abstaining classes.
    """

    # Find samples in ground truth belonging to class 1
    ytrueint = K.argmax(y_true, axis=-1)

    # Compute total number of ground truth samples in class 1
    total_true1 = K.sum(ytrueint)

    # Find samples in prediction belonging to class 1
    ypredint = K.argmax(y_pred[:,:2], axis=-1)

    # Find correctly predicted class 1 samples
    true1_pred = K.sum(ytrueint*ypredint)

    # Compute accuracy in class 1
    acc = true1_pred / total_true1

    # Since there are so few samples in class 1
    # it is possible that ground truth does not
    # have any sample in class 1, leading to a divide
    # by zero and not valid accuracy
    # Therefore, for the accuracy to be valid
    # total_true1 should be greater than zero
    # otherwise, return 0.

    condition = K.greater(total_true1, 0)

    return K.switch(condition, acc, K.zeros_like(acc, dtype=acc.dtype)) 

def abs_acc_class1(y_true, y_pred):
    """ Function to estimate accuracy over the class 1 prediction after removing the samples where the model is abstaining
    
    Parameters
    ----------
    y_true : keras tensor
        True values to predict
    y_pred : keras tensor
        Prediction made by the model. It is assumed that this keras tensor includes extra columns to store the abstaining classes.
    """

    # Find locations of true 1 prediction
    ytrueint = K.argmax(y_true, axis=-1)

    # Find locations that are predicted (not abstained)
    mask_pred = K.cast(K.not_equal(K.argmax(y_pred, axis=-1), nb_classes), 'int64')

    # Compute total number of ground truth samples in class 1 filtering abstaining predictions
    total_true1 = K.sum(ytrueint * mask_pred)

    # matching in original class 1 after removing abstention
    true1_pred = K.sum(mask_pred * ytrueint * K.cast(K.equal(K.argmax(y_true, axis=-1), K.argmax(y_pred, axis=-1)), 'int64'))

    # Compute accuracy in class 1
    acc = true1_pred / total_true1

    # Since there are so few samples in class 1
    # it is possible that ground truth does not
    # have any sample in class 1, leading to a divide
    # by zero and not valid accuracy
    # Therefore, for the accuracy to be valid
    # total_true1 should be greater than zero
    # otherwise, return 0.

    condition = K.greater(total_true1, 0)

    return K.switch(condition, acc, K.zeros_like(acc, dtype=acc.dtype)) 

def crossentropy_outlier_wrap(_l):
    def crossentropy_outlier_core(y_true, y_pred):

        # hyper param
        print(_l)
        y_pred = K.clip(y_pred, K.epsilon(), 1)

        # compute loss for every data point
        _loss = -K.sum(y_true * K.log(y_pred), axis=-1)

        def _get_real_median(_v):
            """
            given a tensor with shape (batch_size,), compute and return the median

            :param v:
            :return:
            """
            _val = tf.nn.top_k(_v, 33).values
            return 0.5 * (_val[-1] + _val[-2])

        _mean_loss, _var_loss = tf.nn.moments(_loss, axes=[0])
        _median_loss = _get_real_median(_loss)
        _std_loss = tf.sqrt(_var_loss)

        # threshold
        t_l = _median_loss + _l*_std_loss
        _mask_l = 1 - (K.cast(K.greater(_loss, t_l), 'float32'))
        _loss = _loss * _mask_l

        return _loss
    return crossentropy_outlier_core



#########################################################################
# from here on we distinguish data points in the batch, based on its origin
# we only apply robustness measures to the data points coming from the noisy subset
# Therefore, the next functions are used only when training with the entire train set
######################################################################### 

def lq_loss_origin_wrap(_q):
    def lq_loss_origin_core(y_true, y_pred):

        # hyper param
        print(_q)

        # 1) determine the origin of the patch, as a boolean vector in y_true_flag
        # (True = patch from noisy subset)
        _y_true_flag = K.greater(K.sum(y_true, axis=-1), 90)

        # 2) convert the input y_true (with flags inside) into a valid y_true one-hot-vector format
        # attenuating factor for data points that need it (those that came with a one-hot of 100)
        _mask_reduce = K.cast(_y_true_flag, 'float32') * 0.01

        # identity factor for standard one-hot vectors
        _mask_keep = K.cast(K.equal(_y_true_flag, False), 'float32')

        # combine 2 masks
        _mask = _mask_reduce + _mask_keep

        _y_true_shape = K.shape(y_true)
        _mask = K.reshape(_mask, (_y_true_shape[0], 1))

        # applying mask to have a valid y_true that we can use as always
        y_true = y_true * _mask

        y_true = K.clip(y_true, K.epsilon(), 1)
        y_pred = K.clip(y_pred, K.epsilon(), 1)

        # compute two types of losses, for all the data points
        # (1) compute CCE loss for every data point
        _loss_CCE = -K.sum(y_true * K.log(y_pred), axis=-1)

        # (2) compute lq_loss for every data point
        _tmp = y_pred * y_true
        _loss_tmp = K.max(_tmp, axis=-1)
        # compute the Lq loss between the one-hot encoded label and the predictions
        _loss_q = (1 - (_loss_tmp + 10 ** (-8)) ** _q) / _q

        # decide which loss to take for each datapoint
        _mask_noisy = K.cast(_y_true_flag, 'float32')                   # only allows patches from noisy set
        _mask_clean = K.cast(K.equal(_y_true_flag, False), 'float32')   # only allows patches from clean set

        # points coming from clean set contribute with CCE loss
        # points coming from noisy set contribute with lq_loss
        _loss_final = _loss_CCE * _mask_clean + _loss_q * _mask_noisy

        return _loss_final
    return lq_loss_origin_core 

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

        lr = self.lr
        if self.inital_decay > 0:
            lr *= (1. / (1. + self.decay * self.iterations))

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

        shapes = [K.get_variable_shape(p) for p in params]
        ms = [K.zeros(shape) for shape in shapes]
        vs = [K.zeros(shape) for shape in shapes]
        f = K.variable(0)
        d = K.variable(1)
        self.weights = [self.iterations] + ms + vs + [f, d]

        cond = K.greater(t, K.variable(1))
        small_delta_t = K.switch(K.greater(loss, f), self.small_k + 1, 1. / (self.big_K + 1))
        big_delta_t = K.switch(K.greater(loss, f), self.big_K + 1, 1. / (self.small_k + 1))

        c_t = K.minimum(K.maximum(small_delta_t, loss / (f + self.epsilon)), big_delta_t)
        f_t = c_t * f
        r_t = K.abs(f_t - f) / (K.minimum(f_t, f))
        d_t = self.beta_3 * d + (1 - self.beta_3) * r_t

        f_t = K.switch(cond, f_t, loss)
        d_t = K.switch(cond, d_t, K.variable(1.))

        self.updates.append(K.update(f, f_t))
        self.updates.append(K.update(d, d_t))

        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 = p - lr_t * m_t / (d_t * K.sqrt(v_t) + self.epsilon)

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))

            new_p = p_t
            self.updates.append(K.update(p, new_p))
        return self.updates 

def crossentropy_reed_origin_wrap(_beta):
    def crossentropy_reed_origin_core(y_true, y_pred):
        # hyper param
        print(_beta)

        # 1) determine the origin of the patch, as a boolean vector in y_true_flag
        # (True = patch from noisy subset)
        _y_true_flag = K.greater(K.sum(y_true, axis=-1), 90)

        # 2) convert the input y_true (with flags inside) into a valid y_true one-hot-vector format
        # attenuating factor for data points that need it (those that came with a one-hot of 100)
        _mask_reduce = K.cast(_y_true_flag, 'float32') * 0.01

        # identity factor for standard one-hot vectors
        _mask_keep = K.cast(K.equal(_y_true_flag, False), 'float32')

        # combine 2 masks
        _mask = _mask_reduce + _mask_keep

        _y_true_shape = K.shape(y_true)
        _mask = K.reshape(_mask, (_y_true_shape[0], 1))

        # applying mask to have a valid y_true that we can use as always
        y_true = y_true * _mask

        y_true = K.clip(y_true, K.epsilon(), 1)
        y_pred = K.clip(y_pred, K.epsilon(), 1)

        # (1) dynamically update the targets based on the current state of the model: bootstrapped target tensor
        # use predicted class proba directly to generate regression targets
        y_true_bootstrapped = _beta * y_true + (1 - _beta) * y_pred

        # at this point we have 2 versions of y_true
        # decide which target label to use for each datapoint
        _mask_noisy = K.cast(_y_true_flag, 'float32')                   # only allows patches from noisy set
        _mask_clean = K.cast(K.equal(_y_true_flag, False), 'float32')   # only allows patches from clean set
        _mask_noisy = K.reshape(_mask_noisy, (_y_true_shape[0], 1))
        _mask_clean = K.reshape(_mask_clean, (_y_true_shape[0], 1))

        # points coming from clean set use the standard true one-hot vector. dim is (batch_size, 1)
        # points coming from noisy set use the Reed bootstrapped target tensor
        y_true_final = y_true * _mask_clean + y_true_bootstrapped * _mask_noisy

        # (2) compute loss as always
        _loss = -K.sum(y_true_final * K.log(y_pred), axis=-1)

        return _loss
    return crossentropy_reed_origin_core