def __init__(self, gan, config, losses=[], var_lists=[], metrics=None):
    BaseTrainer.__init__(self, gan, config)
    self.losses = losses
    self.var_lists = var_lists
    self.metrics = metrics or [None for i in self.losses]

def biases(self):
    """
        Biases of the GAN component.
    """
    return self.ops.biases

def build_optimizer(self, config, prefix, trainer_config, learning_rate, var_list, loss):
    with tf.variable_scope(prefix):
        defn = {k[2:]: v for k, v in config.items() if k[2:] in inspect.getargspec(trainer_config).args and k.startswith(prefix)}
        optimizer = trainer_config(learning_rate, **defn)
        if(config.clipped_gradients):
            apply_gradients = self.capped_optimizer(optimizer, config.clipped_gradients, loss, var_list)
        else:
            apply_gradients = optimizer.minimize(loss, var_list=var_list)
    return apply_gradients

def capped_optimizer(optimizer, cap, loss, var_list):
    gvs = optimizer.compute_gradients(loss, var_list=var_list)
    def create_cap(grad,var):
        if(grad == None) :
            print("Warning: No gradient for variable ",var.name)
            return None
        return (tf.clip_by_value(grad, -cap, cap), var)
    capped_gvs = [create_cap(grad,var) for grad, var in gvs]
    capped_gvs = [x for x in capped_gvs if x != None]
    return optimizer.apply_gradients(capped_gvs)

def create(self):
    self.create_called = True
    return self._create()

def create_ops(self, config):
    """
    Create the ops object as `self.ops`.  Also looks up config
    """
    if self.gan is None:
        return
    if self.gan.ops_backend is None:
        return
    self.ops = self.gan.ops_backend(config=self.config, device=self.gan.device)
    self.config = self.gan.ops.lookup(config)

def fully_connected_from_list(self, nets):
    results = []
    ops = self.ops
    for net, net2 in nets:
        net = ops.concat([net, net2], axis=3)
        shape = ops.shape(net)
        bs = shape[0]
        net = ops.reshape(net, [bs, -1])
        features = ops.shape(net)[1]
        net = ops.linear(net, features)
        #net = self.layer_regularizer(net)
        net = ops.lookup('lrelu')(net)
        #net = ops.linear(net, features)
        net = ops.reshape(net, shape)
        results.append(net)
    return results

def layer_regularizer(self, net):
    symbol = self.config.layer_regularizer
    op = self.gan.ops.lookup(symbol)
    if op:
        net = op(self, net)
    return net

def output_string(self, metrics):
    output = "\%2d: " 
    for name in sorted(metrics.keys()):
        output += " " + name
        output += " %.2f"
    return output

def output_variables(self, metrics):
    gan = self.gan
    sess = gan.session
    return [metrics[k] for k in sorted(metrics.keys())]

def permute(self, nets, k):
    return list(itertools.permutations(nets, k))

def relation_layer(self, net):
    ops = self.ops
    #hack
    shape = ops.shape(net)
    input_size = shape[1]*shape[2]*shape[3]
    netlist = self.split_by_width_height(net)
    permutations = self.permute(netlist, 2)
    permutations = self.fully_connected_from_list(permutations)
    net = ops.concat(permutations, axis=3)
    #hack
    bs = ops.shape(net)[0]
    net = ops.reshape(net, [bs, -1])
    net = ops.linear(net, input_size)
    net = ops.reshape(net, shape)
    return net

def required(self):
    return "d_trainer g_trainer d_learn_rate g_learn_rate".split()

def reuse(self, net):
    self.ops.reuse()
    net = self.build(net)
    self.ops.stop_reuse()
    return net

def split_batch(self, net, count=2):
    """ 
    Discriminators return stacked results (on axis 0).  
    
    This splits the results.  Returns [d_real, d_fake]
    """
    ops = self.ops or self.gan.ops
    s = ops.shape(net)
    bs = s[0]
    nets = []
    net = ops.reshape(net, [bs, -1])
    start = [0 for x in ops.shape(net)]
    for i in range(count):
        size = [bs//count] + [x for x in ops.shape(net)[1:]]
        nets.append(ops.slice(net, start, size))
        start[0] += bs//count
    return nets

def split_by_width_height(self, net):
    elems = []
    ops = self.gan.ops
    shape = ops.shape(net)
    bs = shape[0]
    height = shape[1]
    width = shape[2]
    for i in range(width):
        for j in range(height):
            elems.append(ops.slice(net, [0, i, j, 0], [bs, 1, 1, -1]))
    return elems

def step(self, feed_dict={}):
    if not self.create_called:
        self.create()
    step = self._step(feed_dict)
    self.current_step += 1
    return step

def validate(self):
    """
    Validates a GANComponent.  Return an array of error messages. Empty array `[]` means success.
    """
    errors = []
    required = self.required()
    for argument in required:
        if(self.config.__getattr__(argument) == None):
            errors.append("`"+argument+"` required")
    if(self.gan is None):
        errors.append("GANComponent constructed without GAN")
    return errors

def variables(self):
    """
        All variables associated with this component.
    """
    return self.ops.variables()

def weights(self):
    """
        The weights of the GAN component.
    """
    return self.ops.weights