def test_sync_in_out_synapses(self):
num_in = 3
num_out = 2
num_states = 2
env = blur_env.tf_env
in_out_synapse = tf.random.normal(shape=(num_in + 1, num_out, num_states))
out_in_synapse = tf.random.normal(shape=(num_out, num_in + 1, num_states))
synapse = synapse_util.combine_in_out_synapses(in_out_synapse,
out_in_synapse, env)
synapse_synced = synapse_util.sync_in_and_out_synapse(synapse, num_in, env)
fwd_sync_submatrix = synapse_util.synapse_submatrix(
synapse_synced,
num_in,
synapse_util.UpdateType.FORWARD,
include_bias=True)
bkw_sync_submatrix = synapse_util.synapse_submatrix(
synapse_synced,
num_in,
synapse_util.UpdateType.BACKWARD,
include_bias=True)
with tf.Session() as s:
bwd, fwd, inp = s.run([
synapse_util.transpose_synapse(bkw_sync_submatrix, env),
fwd_sync_submatrix, in_out_synapse
])
self.assertAllEqual(fwd, inp)
self.assertAllEqual(bwd, inp)
def get_blur_state(env, inp, out, ow):
pre = np.concatenate([inp, np.zeros_like(inp)],
axis=-1).astype(blur_env.NP_FLOATING_TYPE)
post = np.concatenate([np.zeros_like(out), out],
axis=-1).astype(blur_env.NP_FLOATING_TYPE)
ww = ow.astype(blur_env.NP_FLOATING_TYPE)
ww = ww[Ellipsis, None]
synapse = synapse_util.combine_in_out_synapses(
ww, synapse_util.transpose_synapse(ww, env), env=env)
synapse = synapse_util.sync_states_synapse(synapse, env, num_states=2)
genome = genome_util.convert_genome_to_tf_variables(
genome_util.create_backprop_genome(num_species=1))
network_spec = blur.NetworkSpec()
network_spec.symmetric_synapses = True
network_spec.batch_average = False
network_spec.backward_update = 'multiplicative_second_state'
return pre, post, synapse, genome, network_spec
def default_state_update(state, genome, network_spec, env):
"""Transforms system state (currently can normalize synapses)."""
synapse_norm_spec = network_spec.synapse_normalization_spec
if synapse_norm_spec is None:
return state
rescale_to = genome.synapse.rescale_to if synapse_norm_spec.rescale else None
normalize = ft.partial(synapse_util.normalize_synapses,
rescale_to=rescale_to,
env=env)
if synapse_norm_spec.normalize_synapses:
for i in range(len(state.synapses)):
synapse = state.synapses[i]
in_channels = state.layers[i].shape[-2]
forward = synapse[Ellipsis, :(in_channels + 1),
(in_channels + 1):, :]
# We include an extra dimension for axis=-2 here because it does not
# influence normalization, but the tensors would have proper shapes:
backward = synapse[Ellipsis,
(in_channels + 1):, :(in_channels + 1), :]
state.synapses[i] = synapse_util.combine_in_out_synapses(
normalize(forward), normalize(backward), env)
return state