def __init__(self, scope=None, **kwargs):
super().__init__(scope=scope, **kwargs)
self.training_wheels = build_scheduled_value(self.training_wheels,
"training_wheels")
self.obj_concrete_temp = build_scheduled_value(self.obj_concrete_temp,
"obj_concrete_temp")
self.obj_temp = build_scheduled_value(self.obj_temp, "obj_temp")
def __init__(self, env, updater, scope=None, **kwargs):
super().__init__(env, updater, scope=scope)
self.reconstruction_weight = build_scheduled_value(
self.reconstruction_weight, "reconstruction_weight")
self.kl_weight = build_scheduled_value(self.kl_weight, "kl_weight")
if not self.noisy and self.train_kl:
raise Exception(
"If `noisy` is False, `train_kl` must also be False.")
def __init__(self, pixels_per_cell, scope=None, **kwargs):
super(GridObjectLayer, self).__init__(scope=scope, **kwargs)
self.pixels_per_cell = pixels_per_cell
self.B = len(self.anchor_boxes)
if isinstance(self.count_prior_dist, str):
self.count_prior_dist = eval(self.count_prior_dist)
self.count_prior_log_odds = build_scheduled_value(
self.count_prior_log_odds, "count_prior_log_odds")
self.obj_concrete_temp = build_scheduled_value(self.obj_concrete_temp,
"obj_concrete_temp")
self.obj_temp = build_scheduled_value(self.obj_temp, "obj_temp")
self.yx_prior_mean = build_scheduled_value(self.yx_prior_mean,
"yx_prior_mean")
self.yx_prior_std = build_scheduled_value(self.yx_prior_std,
"yx_prior_std")
self.hw_prior_mean = build_scheduled_value(self.hw_prior_mean,
"hw_prior_mean")
self.hw_prior_std = build_scheduled_value(self.hw_prior_std,
"hw_prior_std")
self.training_wheels = build_scheduled_value(self.training_wheels,
"training_wheels")
self.anchor_boxes = np.array(self.anchor_boxes)
def __init__(self, pixels_per_cell, grid_offset=None, scope=None, **kwargs):
super(GridObjectLayer, self).__init__(scope=scope, **kwargs)
self.pixels_per_cell = pixels_per_cell
self.grid_offset = np.zeros(2) if grid_offset is None else grid_offset
self.B = self.n_objects_per_cell
self.yx_prior_mean = build_scheduled_value(self.yx_prior_mean, "yx_prior_mean")
self.yx_prior_std = build_scheduled_value(self.yx_prior_std, "yx_prior_std")
self.hw_prior_mean = build_scheduled_value(self.hw_prior_mean, "hw_prior_mean")
self.hw_prior_std = build_scheduled_value(self.hw_prior_std, "hw_prior_std")
self.anchor_box = np.array(self.anchor_box)
def __init__(self, env, updater, scope=None, **kwargs):
super().__init__(env, updater, scope=scope)
self.lmbda = build_scheduled_value(self.lmbda, "lmbda")
self.output_size_per_object = np.prod(
self.object_shape) * (self.image_depth + 1) + 1 + 4 + self.n_layers
def build_update(self, context):
self.delta = build_scheduled_value(self.delta_schedule, "delta")
tvars = self.trainable_variables(for_opt=True)
self.gradient = tf.gradients(context.objective, tvars)
mask = context.get_signal('mask')
kl = context.get_signal('kl', self.policy)
mean_kl = masked_mean(kl, mask)
self.fv_product = HessianVectorProduct(mean_kl, tvars)
self.grad_norm_pure = tf.placeholder(tf.float32,
shape=(),
name="_grad_norm_pure")
self.grad_norm_natural = tf.placeholder(tf.float32,
shape=(),
name="_grad_norm_natural")
self.step_norm = tf.placeholder(tf.float32,
shape=(),
name="_step_norm")
context.add_recorded_values(grad_norm_pure=self.grad_norm_pure,
grad_norm_natural=self.grad_norm_natural,
step_norm=self.step_norm,
train_only=True)
def build_core_signals(self, context):
self.maybe_build_mode()
self.train_exploration = build_scheduled_value(
self.train_exploration_schedule)
if self.val_exploration_schedule is None:
self.exploration = self.val_exploration = self.train_exploration
else:
self.val_exploration = build_scheduled_value(
self.val_exploration_schedule)
self.exploration = tf.cond(
tf.logical_or(
tf.equal(self.mode, "train"),
tf.equal(self.mode, "off_policy"),
), lambda: self.train_exploration,
lambda: self.val_exploration)
label = "{}-exploration".format(self.display_name)
context.add_recorded_value(label, self.exploration)
def __init__(self, env, updater, scope=None, **kwargs):
self.updater = updater
self.obs_shape = env.obs_shape
self.image_height, self.image_width, self.image_depth = self.obs_shape
self.attr_prior_mean = build_scheduled_value(self.attr_prior_mean, "attr_prior_mean")
self.attr_prior_std = build_scheduled_value(self.attr_prior_std, "attr_prior_std")
self.reconstruction_weight = build_scheduled_value(
self.reconstruction_weight, "reconstruction_weight")
self.kl_weight = build_scheduled_value(self.kl_weight, "kl_weight")
self.math_weight = build_scheduled_value(self.math_weight, "math_weight")
if not self.noisy and self.train_kl:
raise Exception("If `noisy` is False, `train_kl` must also be False.")
super(VariationalAutoencoder, self).__init__(scope=scope, **kwargs)
def __init__(self, pixels_per_cell, scope=None, **kwargs):
super(GridObjectLayer, self).__init__(scope=scope, **kwargs)
self.pixels_per_cell = pixels_per_cell
self.B = self.n_objects_per_cell
if isinstance(self.count_prior_dist, str):
self.count_prior_dist = eval(self.count_prior_dist)
self.count_prior_log_odds = build_scheduled_value(self.count_prior_log_odds, "count_prior_log_odds")
self.yx_prior_mean = build_scheduled_value(self.yx_prior_mean, "yx_prior_mean")
self.yx_prior_std = build_scheduled_value(self.yx_prior_std, "yx_prior_std")
self.hw_prior_mean = build_scheduled_value(self.hw_prior_mean, "hw_prior_mean")
self.hw_prior_std = build_scheduled_value(self.hw_prior_std, "hw_prior_std")
self.anchor_box = np.array(self.anchor_box)
def __init__(self, env, updater, scope=None, **kwargs):
ap_iou_values = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
self.eval_funcs = {
"AP_at_point_{}".format(int(10 * v)): BaselineAP(v)
for v in ap_iou_values
}
self.eval_funcs["AP"] = BaselineAP(ap_iou_values)
self.eval_funcs["MOT"] = BaselineMOTMetrics()
self.cc_threshold = build_scheduled_value(self.cc_threshold,
'cc_threshold')
try:
self.cosine_threshold = build_scheduled_value(
self.cosine_threshold, 'cosine_threshold')
except Exception:
self.cosine_threshold = None
super().__init__(env, updater, scope=scope, **kwargs)
def __init__(self, env, updater, scope=None, **kwargs):
super(AIR_Network, self).__init__(env, updater, scope=scope, **kwargs)
ap_iou_values = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
self.eval_funcs = {
"AP_at_point_{}".format(int(10 * v)): AIR_AP(v)
for v in ap_iou_values
}
self.eval_funcs["AP"] = AIR_AP(ap_iou_values)
self.training_wheels = build_scheduled_value(self.training_wheels,
"training_wheels")
self.scale_prior_mean = build_scheduled_value(self.scale_prior_mean,
"scale_prior_mean")
self.scale_prior_std = build_scheduled_value(self.scale_prior_std,
"scale_prior_std")
self.shift_prior_mean = build_scheduled_value(self.shift_prior_mean,
"shift_prior_mean")
self.shift_prior_std = build_scheduled_value(self.shift_prior_std,
"shift_prior_std")
self.z_pres_prior_log_odds = build_scheduled_value(
self.z_pres_prior_log_odds, "z_pres_prior_log_odds")
def build_core_signals(self, context):
self.weight = build_scheduled_value(self.weight_schedule,
"{}-weight".format(self.name))
def __init__(self, **kwargs):
if isinstance(self.count_prior_dist, str):
self.count_prior_dist = eval(self.count_prior_dist)
self.count_prior_log_odds = build_scheduled_value(
self.count_prior_log_odds, "count_prior_log_odds")
super().__init__(**kwargs)
def __init__(self, **kwargs):
self.exp_rate = build_scheduled_value(self.exp_rate, "exp_rate")
super().__init__(**kwargs)
def __init__(self, **kwargs):
self.prior_log_odds = build_scheduled_value(self.prior_log_odds,
"prior_log_odds")
super().__init__(**kwargs)
def __init__(self, **kwargs):
self.obj_concrete_temp = build_scheduled_value(self.obj_concrete_temp,
"obj_concrete_temp")
super().__init__(**kwargs)
def __init__(self, env, scope=None, **kwargs):
self.lr_schedule = build_scheduled_value(self.lr_schedule, "lr")
self.l2_schedule = build_scheduled_value(self.l2_schedule, "l2_weight")
super().__init__(env, scope=scope, **kwargs)
def generate_signal(self, signal_key, context):
if signal_key == "beta":
return build_scheduled_value(self.beta_schedule,
'{}-beta'.format(self.name))
else:
raise Exception("NotImplemented")
