Python keras.backend.zeros() Examples

def call(self, inputs, **kwargs):
        # (batch_size, 1, input_num_capsule, input_dim_capsule)
        expand_inputs = K.expand_dims(inputs, axis=1)
        # (batch_size, num_capsule, input_num_capsule, input_dim_capsule)
        expand_inputs = K.tile(expand_inputs, (1, self.num_capsule, 1, 1))
        # (batch_size, num_capsule, input_num_capsule, dim_capsule)
        u_hat = K.map_fn(lambda x: K.batch_dot(x, self.W, axes=[2, 3]), expand_inputs)

        if self.num_routing <= 0:
            self.num_routing = 3
        # (batch_size, num_capsule, input_num_capsule)
        b = K.zeros((K.shape(u_hat)[0], self.num_capsule, self.input_num_capsule))
        for i in xrange(self.num_routing):
            # (batch_size, num_capsule, input_num_capsule)
            c = softmax(b, axis=1)
            # (batch_size, num_capsule, dim_capsule)
            s = K.batch_dot(c, u_hat, axes=[2, 2])
            squashed_s = squash(s)
            if i < self.num_routing - 1:
                # (batch_size, num_capsule, input_num_capsule)
                b += K.batch_dot(squashed_s, u_hat, axes=[2, 3])
        return squashed_s 

def reset_states(self):
		assert self.stateful, 'Layer must be stateful.'
		input_shape = self.input_spec[0].shape

		if not input_shape[0]:
			raise Exception('If a RNN is stateful, a complete ' +
							'input_shape must be provided (including batch size).')

		if hasattr(self, 'states'):
			K.set_value(self.states[0],
			            np.zeros((input_shape[0], self.hidden_recurrent_dim)))
			K.set_value(self.states[1],
			            np.zeros((input_shape[0], self.input_dim)))
			K.set_value(self.states[2],
			            np.zeros((input_shape[0], self.hidden_dim)))
		else:
			self.states = [K.zeros((input_shape[0], self.hidden_recurrent_dim)),
							K.zeros((input_shape[0], self.input_dim)),
							K.zeros((input_shape[0], self.hidden_dim))] 

def build(self, input_shape):
		self.input_spec = [InputSpec(shape=input_shape)]
		self.input_dim = input_shape[2]

		self.W = self.init((self.output_dim, 4 * self.input_dim),
		                   name='{}_W'.format(self.name))
		self.U = self.inner_init((self.input_dim, 4 * self.input_dim),
		                         name='{}_U'.format(self.name))
		self.b = K.variable(np.hstack((np.zeros(self.input_dim),
		                               K.get_value(self.forget_bias_init((self.input_dim,))),
		                               np.zeros(self.input_dim),
		                               np.zeros(self.input_dim))),
		                    name='{}_b'.format(self.name))

		self.A = self.init((self.input_dim, self.output_dim),
		                    name='{}_A'.format(self.name))
		self.ba = K.zeros((self.output_dim,), name='{}_ba'.format(self.name))


		self.trainable_weights = [self.W, self.U, self.b, self.A, self.ba]

		if self.initial_weights is not None:
			self.set_weights(self.initial_weights)
			del self.initial_weights 

def reset_states(self):
		assert self.stateful, 'Layer must be stateful.'
		input_shape = self.input_spec[0].shape
		if not input_shape[0]:
			raise ValueError('If a RNN is stateful, it needs to know '
			                 'its batch size. Specify the batch size '
			                 'of your input tensors: \n'
			                 '- If using a Sequential model, '
			                 'specify the batch size by passing '
			                 'a `batch_input_shape` '
			                 'argument to your first layer.\n'
			                 '- If using the functional API, specify '
			                 'the time dimension by passing a '
			                 '`batch_shape` argument to your Input layer.')
		if hasattr(self, 'states'):
			K.set_value(self.states[0],
			            np.zeros((input_shape[0], self.input_dim)))
			K.set_value(self.states[1],
			            np.zeros((input_shape[0], self.output_dim)))
		else:
			self.states = [K.zeros((input_shape[0], self.input_dim)),
							K.zeros((input_shape[0], self.output_dim))] 

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 build(self, input_shape):
        # Create mean and count
        # These are weights because just maintaining variables don't get saved with the model, and we'd like
        # to have these numbers saved when we save the model.
        # But we need to make sure that the weights are untrainable.
        self.mean = self.add_weight(name='mean', 
                                      shape=input_shape[1:],
                                      initializer='zeros',
                                      trainable=False)
        self.count = self.add_weight(name='count', 
                                      shape=[1],
                                      initializer='zeros',
                                      trainable=False)

        # self.mean = K.zeros(input_shape[1:], name='mean')
        # self.count = K.variable(0.0, name='count')
        super(MeanStream, self).build(input_shape)  # Be sure to call this somewhere! 

def cmc(model):
    
    def cmc_curve(model, camera1, camera2, rank_max=50):
        num = camera1.shape[0]    
        rank = []
        score = []    
        camera_batch1 = np.zeros(camera1.shape)
        for i in range(num):
            for j in range(num):
                camera_batch1[j] = camera1[i]
            similarity_batch = model.predict_on_batch([camera_batch1, camera2])
            sim_trans = similarity_batch.transpose()
            similarity_rate_sorted = np.argsort(sim_trans[0])
            for k in range(num):
                if similarity_rate_sorted[k] == i:
                    rank.append(k+1)
                    break
        rank_val = 0
        for i in range(rank_max):
            rank_val = rank_val + len([j for j in rank if i == j-1])        
            score.append(rank_val / float(num))
        return np.array(score)  
        
    a,b = get_data_for_cmc()
    return cmc_curve(model,a,b) 

def call(self, x, mask=None):

            s = K.shape(x)
            b = s[0]
            r = s[1]
            c = s[2]
            ch = s[3]

            half_n = self.n // 2 # half the local region

            input_sqr = K.square(x) # square the input

            extra_channels = K.zeros((b, r, c, ch + 2 * half_n))
            input_sqr = K.concatenate([extra_channels[:, :, :, :half_n],input_sqr, extra_channels[:, :, :, half_n + ch:]], axis = 3)

            scale = self.k # offset for the scale
            norm_alpha = self.alpha / self.n # normalized alpha
            for i in range(self.n):
                scale += norm_alpha * input_sqr[:, :, :, i:i+ch]
            scale = scale ** self.beta
            x = x / scale
            
            return x 

def vectorizeData(xContext, xQuestion, xAnswerBeing, xAnswerEnd, word_index, context_maxlen, question_maxlen):
    '''Vectorize the words to their respective index and pad context to max context length and question to max question length.
       Answers vectors are padded to the max context length as well.
    '''
    X = []
    Xq = []
    YBegin = []
    YEnd = []
    for i in xrange(len(xContext)):
        x = [word_index[w] for w in xContext[i]]
        xq = [word_index[w] for w in xQuestion[i]]
        # map the first and last words of answer span to one-hot representations
        y_Begin =  np.zeros(len(xContext[i]))
        y_Begin[xAnswerBeing[i]] = 1
        y_End = np.zeros(len(xContext[i]))
        y_End[xAnswerEnd[i]] = 1
        X.append(x)
        Xq.append(xq)
        YBegin.append(y_Begin)
        YEnd.append(y_End)
    return pad_sequences(X, maxlen=context_maxlen, padding='post'), pad_sequences(Xq, maxlen=question_maxlen, padding='post'), pad_sequences(YBegin, maxlen=context_maxlen, padding='post'), pad_sequences(YEnd, maxlen=context_maxlen, padding='post')

# Note: Need to download and unzip Glove pre-train model files into same file as this script 

def build(self, input_shape):
        # input shape is (batch_size, num_words, num_senses, num_hyps)
        self.num_senses = input_shape[-2]
        self.num_hyps = input_shape[-1] - 1  # -1 because the last value is a word index
        # embedding of size 1.
        if self.set_sense_priors:
            self.sense_priors = self._get_initial_sense_priors((self.word_index_size, 1), name='{}_sense_priors'.format(self.name))
        else:
            # OntoLSTM makes sense proabilities uniform if the passed sense parameters are zero.
            self.sense_priors = K.zeros((self.word_index_size, 1))  # uniform sense probs
        # Keeping aside the initial weights to not let Embedding set them. It wouldn't know what sense priors are.
        if self.initial_weights is not None:
            self.onto_aware_embedding_weights = self.initial_weights
            self.initial_weights = None
        # The following method will set self.trainable_weights
        super(OntoAwareEmbedding, self).build(input_shape)  # input_shape will not be used by Embedding's build.
        if not self.tune_embedding:
            # Move embedding to non_trainable_weights
            self._non_trainable_weights.append(self._trainable_weights.pop())

        if self.set_sense_priors:
            self._trainable_weights.append(self.sense_priors)

        if self.onto_aware_embedding_weights is not None:
            self.set_weights(self.onto_aware_embedding_weights) 

def reset_states(self):
    assert self.stateful, 'Layer must be stateful.'
    input_shape = self.input_spec[0].shape
    if not input_shape[0]:
      raise Exception('If a RNN is stateful, a complete ' +
                      'input_shape must be provided (including batch size).')
    if hasattr(self, 'states'):
      K.set_value(self.states[0],
                  np.zeros((input_shape[0], self.output_dim)))
    else:
      self.states = [K.zeros((input_shape[0], self.output_dim))] 

def build(self, input_shape):
    super(LSTM_LN, self).build(input_shape)
    self.gs, self.bs = [], []
    for i in xrange(3):
      f = 1 if i == 2 else 4
      self.gs += [ K.ones((f*self.output_dim,), name='{}_g%i'.format(self.name, i)) ]
      self.bs += [ K.zeros((f*self.output_dim,), name='{}_b%d'.format(self.name, i)) ]
    self.trainable_weights += self.gs + self.bs 

def build(self):
        stack_size = self.input_shape[2]
        dtensor5 = T.TensorType('float32', (False,)*5)
        self.input = dtensor5()
        self.W_shape = (self.nb_filter, stack_size, self.nb_row, self.nb_col)
        self.W = self.init(self.W_shape)
        self.b = shared_zeros((self.nb_filter,))

        self.params = [self.W, self.b]

        self.regularizers = []

        if self.W_regularizer:
            self.W_regularizer.set_param(self.W)
            self.regularizers.append(self.W_regularizer)

        if self.b_regularizer:
            self.b_regularizer.set_param(self.b)
            self.regularizers.append(self.b_regularizer)

        if self.activity_regularizer:
            self.activity_regularizer.set_layer(self)
            self.regularizers.append(self.activity_regularizer)

        if self.initial_weights is not None:
            self.set_weights(self.initial_weights)
            del self.initial_weights 

def build(self, input_shape):
		self.input_spec = [InputSpec(shape=input_shape)]
		input_dim = input_shape[2]
		self.input_dim = input_dim
		
		if self.stateful:
			self.reset_states()
		else:
			self.states = [None, None]
			self.states_dim = [self.input_dim, self.output_dim]


		self.weight_size = self.output_dim * 4
		self.W = self.add_weight((input_dim, self.weight_size),
                                 initializer=self.init,
                                 name='{}_W'.format(self.name),
                                 regularizer=self.W_regularizer)
		self.U = self.add_weight((input_dim, self.weight_size),
                                 initializer=self.inner_init,
                                 name='{}_U'.format(self.name),
                                 regularizer=self.U_regularizer)

		def b_reg(shape, name=None):
			return K.variable(np.hstack((np.zeros(self.output_dim),
										K.get_value(self.forget_bias_init((self.output_dim,))),
										np.zeros(self.output_dim),
										np.zeros(self.output_dim))),
										name='{}_b'.format(self.name))
		self.b = self.add_weight((self.weight_size,),
                                     initializer=b_reg,
                                     name='{}_b'.format(self.name),
                                     regularizer=self.b_regularizer)


		if self.initial_weights is not None:
			self.set_weights(self.initial_weights)
			del self.initial_weights

		self.built = True 

def reset_states(self):
        assert self.stateful, 'Layer must be stateful.'
        input_shape = self.input_spec[0].shape
        if not input_shape[0]:
            raise Exception('If a RNN is stateful, a complete ' +
                            'input_shape must be provided (including batch size).')
        if hasattr(self, 'states'):
            K.set_value(self.states[0],
                        np.zeros((input_shape[0], self.output_dim)))
            K.set_value(self.states[1],
                        np.zeros((input_shape[0], self.output_dim)))
        else:
            self.states = [K.zeros((input_shape[0], self.output_dim)),
                           K.zeros((input_shape[0], self.output_dim))] 

def initial(number_of_memory_locations, memory_vector_size):
    return K.zeros((number_of_memory_locations, memory_vector_size)) 

def build(self, input_shape):
		'''Builds internal weights and paramer attribute'''
		# NOTE: NEED TO TILE AND EVALUATE SO THAT PARAMS CAN BE VARIABLES
		# OTHERWISE K.GET_VALUE() DOES NOT WORK

		# Define template weights for inner FxF
		W_inner = self.init_inner((self.inner_dim, self.inner_dim))
		b_inner = K.zeros((1, self.inner_dim))
		# Initialize weights tensor 
		self.W_inner = K.variable(T.tile(W_inner, (self.depth + 1, 1, 1)).eval() + \
			initializations.uniform((self.depth + 1, self.inner_dim, self.inner_dim)).eval())
		self.W_inner.name = 'T:W_inner'
		self.b_inner = K.variable(T.tile(b_inner, (self.depth + 1, 1, 1)).eval()  + \
			initializations.uniform((self.depth + 1, 1, self.inner_dim)).eval())
		self.b_inner.name = 'T:b_inner'
		# # Concatenate third dimension (depth) so different layers can have 
		# # different weights. Now, self.W_inner[#,:,:] corresponds to the 
		# # weight matrix for layer/depth #.

		# Define template weights for output FxL
		W_output = self.init_output((self.inner_dim, self.output_dim), scale = self.scale_output)
		b_output = K.zeros((1, self.output_dim))
		# Initialize weights tensor
		self.W_output = K.variable(T.tile(W_output, (self.depth + 1, 1, 1)).eval())
		self.W_output.name = 'T:W_output'
		self.b_output = K.variable(T.tile(b_output, (self.depth + 1, 1, 1)).eval())
		self.b_output.name = 'T:b_output'
		# # Concatenate third dimension (depth) so different layers can have 
		# # different weights. Now, self.W_output[#,:,:] corresponds to the 
		# # weight matrix for layer/depth #.

		# Pack params
		self.trainable_weights = [self.W_inner, 
					   self.b_inner,
					   self.W_output,
					   self.b_output]
		self.params = [self.W_inner, 
					   self.b_inner,
					   self.W_output,
					   self.b_output] 

def build(self, input_shape):
		'''Builds internal weights and paramer attribute'''
		# NOTE: NEED TO TILE AND EVALUATE SO THAT PARAMS CAN BE VARIABLES
		# OTHERWISE K.GET_VALUE() DOES NOT WORK

		# Define template weights for inner FxF
		W_inner = self.init_inner((self.inner_dim, self.inner_dim))
		b_inner = K.zeros((1, self.inner_dim))
		# Initialize weights tensor 
		self.W_inner = K.variable(T.tile(W_inner, (self.depth + 1, 1, 1)).eval() + \
			initializations.uniform((self.depth + 1, self.inner_dim, self.inner_dim)).eval())
		self.W_inner.name = 'T:W_inner'
		self.b_inner = K.variable(T.tile(b_inner, (self.depth + 1, 1, 1)).eval()  + \
			initializations.uniform((self.depth + 1, 1, self.inner_dim)).eval())
		self.b_inner.name = 'T:b_inner'
		# # Concatenate third dimension (depth) so different layers can have 
		# # different weights. Now, self.W_inner[#,:,:] corresponds to the 
		# # weight matrix for layer/depth #.

		# Define template weights for output FxL
		W_output = self.init_output((self.inner_dim, self.output_dim), scale = self.scale_output)
		b_output = K.zeros((1, self.output_dim))
		# Initialize weights tensor
		self.W_output = K.variable(T.tile(W_output, (self.depth + 1, 1, 1)).eval())
		self.W_output.name = 'T:W_output'
		self.b_output = K.variable(T.tile(b_output, (self.depth + 1, 1, 1)).eval())
		self.b_output.name = 'T:b_output'
		# # Concatenate third dimension (depth) so different layers can have 
		# # different weights. Now, self.W_output[#,:,:] corresponds to the 
		# # weight matrix for layer/depth #.

		# Pack params
		self.trainable_weights = [self.W_inner, 
					   self.b_inner,
					   self.W_output,
					   self.b_output]
		self.params = [self.W_inner, 
					   self.b_inner,
					   self.W_output,
					   self.b_output] 

def __init__(self, filters,
                 kernel_size,
                 strides=(1, 1, 1),
                 padding='valid',
                 data_format=None,
                 activation=None,
                 use_bias=True,
                 kernel_initializer='glorot_uniform',
                 bias_initializer='zeros',
                 kernel_regularizer=None,
                 bias_regularizer=None,
                 activity_regularizer=None,
                 kernel_constraint=None,
                 bias_constraint=None,
                 **kwargs):
        
        super(LocallyConnected3D, self).__init__(**kwargs)
        self.filters = filters
        self.kernel_size = conv_utils.normalize_tuple(
            kernel_size, 3, 'kernel_size')
        self.strides = conv_utils.normalize_tuple(strides, 3, 'strides')
        self.padding = conv_utils.normalize_padding(padding)
        if self.padding != 'valid':
            raise ValueError('Invalid border mode for LocallyConnected3D '
                             '(only "valid" is supported): ' + padding)
        self.data_format = conv_utils.normalize_data_format(data_format)
        self.activation = activations.get(activation)
        self.use_bias = use_bias
        self.kernel_initializer = initializers.get(kernel_initializer)
        self.bias_initializer = initializers.get(bias_initializer)
        self.kernel_regularizer = regularizers.get(kernel_regularizer)
        self.bias_regularizer = regularizers.get(bias_regularizer)
        self.activity_regularizer = regularizers.get(activity_regularizer)
        self.kernel_constraint = constraints.get(kernel_constraint)
        self.bias_constraint = constraints.get(bias_constraint)
        self.input_spec = InputSpec(ndim=5) 

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

    lr = self.lr
    if self.initial_decay > 0:
      lr *= (1. / (1. + self.decay * K.cast(self.iterations, K.dtype(self.decay))))

    t = K.cast(self.iterations, K.floatx()) + 1
    lr_t = lr * (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]
    vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params] 
    self.weights = [self.iterations] + ms + vs + vhats

    for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
      m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
      v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
      vhat_t = K.maximum(vhat, v_t)
      p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)

      self.updates.append(K.update(m, m_t))
      self.updates.append(K.update(v, v_t))
      self.updates.append(K.update(vhat, vhat_t))
      new_p = p_t

      # Apply constraints.
      if getattr(p, 'constraint', None) is not None:
        new_p = p.constraint(new_p)

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

def residual_drop(x, input_shape, output_shape, strides=(1, 1)):
    global add_tables

    nb_filter = output_shape[0]
    conv = Convolution2D(nb_filter, 3, 3, subsample=strides,
                         border_mode="same", W_regularizer=l2(weight_decay))(x)
    conv = BatchNormalization(axis=1)(conv)
    conv = Activation("relu")(conv)
    conv = Convolution2D(nb_filter, 3, 3,
                         border_mode="same", W_regularizer=l2(weight_decay))(conv)
    conv = BatchNormalization(axis=1)(conv)

    if strides[0] >= 2:
        x = AveragePooling2D(strides)(x)

    if (output_shape[0] - input_shape[0]) > 0:
        pad_shape = (1,
                     output_shape[0] - input_shape[0],
                     output_shape[1],
                     output_shape[2])
        padding = K.zeros(pad_shape)
        padding = K.repeat_elements(padding, K.shape(x)[0], axis=0)
        x = Lambda(lambda y: K.concatenate([y, padding], axis=1),
                   output_shape=output_shape)(x)

    _death_rate = K.variable(death_rate)
    scale = K.ones_like(conv) - _death_rate
    conv = Lambda(lambda c: K.in_test_phase(scale * c, c),
                  output_shape=output_shape)(conv)

    out = merge([conv, x], mode="sum")
    out = Activation("relu")(out)

    gate = K.variable(1, dtype="uint8")
    add_tables += [{"death_rate": _death_rate, "gate": gate}]
    return Lambda(lambda tensors: K.switch(gate, tensors[0], tensors[1]),
                  output_shape=output_shape)([out, x]) 

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

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        beta_1_t = K.pow(self.beta_1, t)
        beta_2_t = K.pow(self.beta_2, t)
        rho = 2 / (1 - self.beta_2) - 1
        rho_t = rho - 2 * t * beta_2_t / (1 - beta_2_t)
        r_t = K.sqrt(
            K.relu(rho_t - 4) * K.relu(rho_t - 2) * rho / ((rho - 4) * (rho - 2) * rho_t)
        )
        flag = K.cast(rho_t > 4, K.floatx())

        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)
            mhat_t = m_t / (1 - beta_1_t)
            vhat_t = K.sqrt(v_t / (1 - beta_2_t))
            p_t = p - lr * mhat_t * (flag * r_t / (vhat_t + self.epsilon) + (1 - flag))

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

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

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

def next(self):
        with self.lock:
            index_array, current_index, current_batch_size = next(self.index_generator)
        batch_x1 = np.zeros(tuple([current_batch_size * 2] + [128,64,3]))
        batch_x2 = np.zeros(tuple([current_batch_size * 2] + [128,64,3]))
        batch_y  = np.zeros([current_batch_size * 2, 2])
        for i, j in enumerate(index_array):
            x1 = np.array(Image.open(self.folder_dir + self.f[self.train_or_validation][j,0])) / 255.
            x2 = np.array(Image.open(self.folder_dir + self.f[self.train_or_validation][j,1])) / 255.  
            if np.random.rand() > self.flag:
                x1 = self.image_data_generator.random_transform(x1.astype('float32'))
            if np.random.rand() > self.flag:
                x2 = self.image_data_generator.random_transform(x2.astype('float32'))
            batch_x1[2*i] = x1
            batch_x2[2*i] = x2
            batch_y[2*i][1] = 1
            while True:
                index_1,index_2 = np.random.choice(self.path_list,2)
                if index_1[6] != index_2[6] and index_1[0:4] != index_2[0:4]:
                    break
            x1 = np.array(Image.open(self.folder_dir + index_1)) / 255.
            x2 = np.array(Image.open(self.folder_dir + index_2)) / 255.
            batch_x1[2*i+1] = x1
            batch_x2[2*i+1] = x2
            batch_y[2*i+1][0] = 1
            
        return [batch_x1,batch_x2], batch_y 

def agumentation(self, X, rounds=1, seed=None):
        
        if seed is not None:
            np.random.seed(seed)

        X = np.copy(X)
        aX = np.zeros(tuple([rounds * X.shape[0]] + list(X.shape)[1:]))
        for r in range(rounds):
            for i in range(X.shape[0]):
                aX[i + r * X.shape[0]] = self.random_transform(X[i])
        X = aX
        return X 

def get_initial_state(self, inputs):
        if isinstance(inputs, list):
            assert len(inputs) == 2  # inputs == [encoder_outputs, y_true]
            encoder_outputs = inputs[0]
        else:
            encoder_outputs = inputs

        memory_shape = K.shape(encoder_outputs)

        # apply the matrix on the first time step to get the initial s0.
        s0 = activations.tanh(K.dot(encoder_outputs[:, 0], self.W_s))

        y0 = K.zeros((memory_shape[0],), dtype='int64') + self.start_token
        t0 = K.zeros((memory_shape[0],), dtype='int64')

        initial_states = [y0, s0, t0]
        if self.is_monotonic:
            # initial attention has form: [1, 0, 0, ..., 0] for each sample in batch
            alpha0 = K.ones((memory_shape[0], 1))
            alpha0 = K.switch(K.greater(memory_shape[1], 1),
                              lambda: K.concatenate([alpha0, K.zeros((memory_shape[0], memory_shape[1] - 1))], axis=-1),
                              alpha0)
            # like energy, attention is stored in shape (samples, time, 1)
            alpha0 = K.expand_dims(alpha0, -1)
            initial_states.append(alpha0)

        return initial_states 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        weights = self.get_weights()
        self.updates = [K.update_add(self.iterations, 1)]
        scaled_lr = self.lr
        w_norm = K.sqrt(K.sum([K.sum(K.square(weight))
                               for weight in weights]))
        g_norm = K.sqrt(K.sum([K.sum(K.square(grad))
                               for grad in grads]))
        scaled_lr = K.switch(K.greater(w_norm * g_norm, K.zeros([1])),
                             K.expand_dims((self.eeta * w_norm /
                                            (g_norm + self.weight_decay * w_norm +
                                             self.epsilon)) * self.lr),
                             K.ones([1]) * self.lr)
        if K.backend() == 'theano':
            scaled_lr = scaled_lr[0]  # otherwise theano raise broadcasting error
        # momentum
        moments = [K.zeros(K.int_shape(param), dtype=K.dtype(param))
                   for param in params]
        self.weights = [self.iterations] + moments
        for param, grad, moment in zip(params, grads, moments):
            v0 = (moment * self.momentum)
            v1 = scaled_lr * grad  # velocity
            veloc = v0 - v1
            self.updates.append(K.update(moment, veloc))

            if self.nesterov:
                new_param = param + (veloc * self.momentum) - v1
            else:
                new_param = param + veloc

            # Apply constraints.
            if getattr(param, 'constraint', None) is not None:
                new_param = param.constraint(new_param)

            self.updates.append(K.update(param, new_param))
        return self.updates 

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

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = lr * (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]
        vhats = [K.zeros(1) for _ in params]
        self.weights = [self.iterations] + ms + vs + vhats

        for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
            g2 = K.square(g)
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = v - (1. - self.beta_2) * K.sign(v - g2) * g2
            p_t = p - 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))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

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

def get_updates(self, loss, params):
        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 / (1. - K.pow(self.beta_1, t))

        shapes = [K.int_shape(p) for p in params]
        zs = [K.zeros(shape) for shape in shapes]
        vs = [K.zeros(shape) for shape in shapes]
        ds = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + zs + vs + ds

        for p, g, z, v, d in zip(params, grads, zs, vs, ds):
            v_t = self.beta_2 * v + (1. - self.beta_2) * K.square(g)
            d_t = (K.sqrt(v_t / (1. - K.pow(self.beta_2, t)))
                   + self.epsilon) / lr_t
            sigma_t = d_t - self.beta_1 * d
            z_t = self.beta_1 * z + (1. - self.beta_1) * g - sigma_t * p

            p_t = - z_t / d_t

            self.updates.append(K.update(z, z_t))
            self.updates.append(K.update(v, v_t))
            self.updates.append(K.update(d, d_t))

            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

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

def test_DSSIM_channels_last():
    prev_data = K.image_data_format()
    K.set_image_data_format('channels_last')
    for input_dim, kernel_size in zip([32, 33], [2, 3]):
        input_shape = [input_dim, input_dim, 3]
        X = np.random.random_sample(4 * input_dim * input_dim * 3)
        X = X.reshape([4] + input_shape)
        y = np.random.random_sample(4 * input_dim * input_dim * 3)
        y = y.reshape([4] + input_shape)

        model = Sequential()
        model.add(Conv2D(32, (3, 3), padding='same', input_shape=input_shape,
                         activation='relu'))
        model.add(Conv2D(3, (3, 3), padding='same', input_shape=input_shape,
                         activation='relu'))
        adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-8)
        model.compile(loss=DSSIMObjective(kernel_size=kernel_size),
                      metrics=['mse'],
                      optimizer=adam)
        model.fit(X, y, batch_size=2, epochs=1, shuffle='batch')

        # Test same
        x1 = K.constant(X, 'float32')
        x2 = K.constant(X, 'float32')
        dssim = DSSIMObjective(kernel_size=kernel_size)
        assert_allclose(0.0, K.eval(dssim(x1, x2)), atol=1e-4)

        # Test opposite
        x1 = K.zeros([4] + input_shape)
        x2 = K.ones([4] + input_shape)
        dssim = DSSIMObjective(kernel_size=kernel_size)
        assert_allclose(0.5, K.eval(dssim(x1, x2)), atol=1e-4)

    K.set_image_data_format(prev_data) 

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

        t = self.iterations + 1
        lr_t = self.lr * 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]
        gs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        self.weights = ms + vs
        
        flag = K.equal(t % self.accum_iters, 0)
        flag = K.cast(flag, dtype='float32')
        
        for p, g, m, v, gg in zip(params, grads, ms, vs, gs):

            gg_t = (1 - flag) * (gg + g)
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * (gg + flag * g) / self.accum_iters
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square((gg + flag * g) / self.accum_iters)
            p_t = p - flag * lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

            self.updates.append((m, flag * m_t + (1 - flag) * m))
            self.updates.append((v, flag * v_t + (1 - flag) * v))
            self.updates.append((gg, gg_t))
            
            # apply constraints.
            new_p = p_t
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)
            self.updates.append(K.update(p, new_p))
        return self.updates