def build_fcn_net(self, inp, use_dice=False): with self.graph.as_default(): self.saver = tf.train.Saver(max_to_keep=1) with tf.name_scope("Out"): bn1 = tf.layers.batch_normalization(inputs=inp, name='bn1') dnn1 = tf.layers.dense(bn1, 200, activation=None, name='f1') if use_dice: dnn1 = dice(dnn1, name='dice_1') else: dnn1 = prelu(dnn1, 'prelu1') dnn2 = tf.layers.dense(dnn1, 80, activation=None, name='f2') if use_dice: dnn2 = dice(dnn2, name='dice_2') else: dnn2 = prelu(dnn2, 'prelu2') dnn3 = tf.layers.dense(dnn2, 2, activation=None, name='f3') self.y_hat = tf.nn.softmax(dnn3) + 0.00000001 with tf.name_scope('Metrics'): # Cross-entropy loss and optimizer initialization # 'core_type_ph': [1, 1, 0,..], ctr_loss = - tf.reduce_mean(tf.log(self.y_hat) * self.target_ph) self.loss = ctr_loss # tf.summary.scalar('loss', self.loss) self.optimizer = tf.train.AdamOptimizer(learning_rate=self.lr_ph).minimize(self.loss) # self.optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.lr_ph).minimize(self.loss) # Accuracy metric self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.round(self.y_hat), self.target_ph), tf.float32)) # tf.summary.scalar('accuracy', self.accuracy) self.merged = tf.summary.merge_all()
def __init__(self, dim, samples , init_w= tf.random_uniform_initializer(minval= -0.01, maxval= 0.01) , ftype= tf.float32, scope= 'sbn'): self.dim, self.ftype, self.scope = dim, ftype, scope with tf.variable_scope(scope): self.wr = tuple( tf.get_variable(name= "wr{}".format(i), shape= (dim_d, dim_a), initializer= init_w) for i, (dim_d, dim_a) in enumerate(zip(self.dim, self.dim[1:]), 1)) self.wg = tuple( tf.get_variable(name= "wg{}".format(i), shape= (dim_a, dim_d), initializer= init_w) for i, (dim_d, dim_a) in enumerate(zip(self.dim, self.dim[1:]), 1))[::-1] self.lr_ = tf.placeholder(name= 'lr_', dtype= self.ftype, shape= ()) # wake self.v_ = tf.placeholder(name= 'v_', dtype= self.ftype, shape= (None, self.dim[0])) with tf.name_scope('wake'): recogn = [self.v_] for w in self.wr: recogn.append(binary(tf.matmul(recogn[-1], w))) self.recogn = tuple(recogn) recogn = recogn[::-1] eps = self.lr_ / tf.cast(tf.shape(self.v_)[0], dtype= self.ftype) self.wake = tuple( w.assign_add(tf.matmul(sk, (sj - pj), transpose_a= True) * eps).op for w, sk, sj, pj in zip( self.wg, recogn, recogn[1:] , (tf.sigmoid(tf.matmul(s, w)) for w, s in zip(self.wg, recogn)))) # sleep with tf.name_scope('a'): self.a = tf.round(tf.random_uniform(shape= (samples, self.dim[-1]))) with tf.name_scope('sleep'): recons = [self.a] for w in self.wg: recons.append(binary(tf.matmul(recons[-1], w))) self.recons = tuple(recons) recons = recons[::-1] eps = self.lr_ / tf.cast(tf.shape(self.a)[0], dtype= self.ftype) self.sleep = tuple( w.assign_add(tf.matmul(sj, (sk - qk), transpose_a= True) * eps).op for w, sj, sk, qk in zip( self.wr, recons, recons[1:] , (tf.sigmoid(tf.matmul(s, w)) for w, s in zip(self.wr, recons)))) # the waking world is the amnesia of dream. self.v = self.recons[-1] self.step = 0