def pbroadcast_value(value, value_axis_names, output_axis_names):
value_axis_names = distribute_lib.canonicalize_named_axis(value_axis_names)
pbroadcast_axes = [
axis_name for axis_name in output_axis_names
if axis_name not in value_axis_names
]
return distribute_lib.pbroadcast(value, named_axis=pbroadcast_axes)
def manual_sharded_model():
# This one has manual pbroadcasts; the goal is to get sharded_model above
# to do this automatically.
x = yield root(tfd.LogNormal(0., 1., name='x'))
x = distribute_lib.pbroadcast(x, axis_name=self.axis_name)
yield sharded.Sharded(
tfd.Uniform(0., x), shard_axis_name=self.axis_name, name='y')
yield sharded.Sharded(
tfb.Scale(x)(tfd.Normal(0., 1.)),
shard_axis_name=self.axis_name,
name='z')
def target_log_prob(a, b):
return (
tfd.Normal(0., 1.).log_prob(a)
+ distribute_lib.psum(tfd.Normal(
distribute_lib.pbroadcast(a, 'foo'), 1.).log_prob(b), 'foo'))
def adjust_state(x, v, shard_axes=None):
broadcasted_dt = distribute_lib.pbroadcast(
bu.left_justified_expand_dims_like(dt, v), shard_axes)
return x + broadcasted_dt * v