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 test_synapse_derivative(self):
tf.reset_default_graph()
tf.disable_v2_behavior()
n_in, n_out = 10, 5
inp, out, w = random_dense(n_in, n_out)
env = blur_env.tf_env
pre, post, synapse, genome, network_spec = get_blur_state(
env, inp, out, w)
post = np.concatenate(2 * [np.zeros_like(out)],
axis=-1).astype(blur_env.NP_FLOATING_TYPE)
pre_fwd, post_fwd = blur.dense_neuron_update(
pre,
post,
synapse,
inp_act=None,
out_act=sigmoid_with_grad,
neuron_genome=genome.neuron,
update_type=synapse_util.UpdateType.FORWARD,
global_spec=network_spec,
env=env)
hebbian_update = blur.get_hebbian_update(pre_fwd, post_fwd,
genome.synapse.transform,
network_spec, env)
hebbian_update_submatrix = synapse_util.synapse_submatrix(
hebbian_update, n_in, synapse_util.UpdateType.FORWARD)
inp = tf.constant(inp.astype(blur_env.NP_FLOATING_TYPE))
inp_with_bias = tf.concat([inp[Ellipsis, 0], [[[[1]]]]], axis=-1)
ww = tf.constant(w.astype(blur_env.NP_FLOATING_TYPE))
out = tf.nn.sigmoid(inp_with_bias @ ww)
grad_w = tf.gradients(out, ww)
with tf.Session() as s:
s.run(tf.initialize_all_variables())
hebb_update, grad_w_val = s.run(
[hebbian_update_submatrix[Ellipsis, 0], grad_w[0]])
self.assertAllClose(hebb_update, grad_w_val)
def test_get_hebbian_update(self):
tf.reset_default_graph()
tf.disable_v2_behavior()
n_in, n_out = 10, 5
inp, out, w = random_dense(n_in, n_out)
env = blur_env.tf_env
pre, post, _, genome, network_spec = get_blur_state(env, inp, out, w)
hebbian_update = blur.get_hebbian_update(pre, post,
genome.synapse.transform,
network_spec, env)
hebbian_update_submatrix = synapse_util.synapse_submatrix(
hebbian_update, n_in, synapse_util.UpdateType.FORWARD)
inp = tf.constant(inp.astype(blur_env.NP_FLOATING_TYPE))
out = tf.constant(out.astype(blur_env.NP_FLOATING_TYPE))
inp_transpose = tf.transpose(inp[Ellipsis, 0], (0, 1, 3, 2))
exp_result = env.concat_row(inp_transpose) @ out[Ellipsis, 0]
with tf.Session() as s:
s.run(tf.initialize_all_variables())
self.assertAllClose(hebbian_update_submatrix[Ellipsis, 0], exp_result)
def single_synaptic_update(synapse,
input_layer,
in_channels,
update_type,
transform_gn,
synapse_transform_fn=None,
*,
env):
"""Computes one-way (Forward or Backward) synaptic update."""
include_bias = False
if update_type == synapse_util.UpdateType.FORWARD:
# Input channels are +1 to include "1" channel to simulate bias.
input_layer = env.concat_row(input_layer)
include_bias = True
subsynapse = synapse_util.synapse_submatrix(synapse,
in_channels=in_channels,
update_type=update_type,
include_bias=include_bias)
if synapse_transform_fn is not None:
subsynapse = synapse_transform_fn(subsynapse)
update = env.einsum(FC_UPDATE_PATTERN, input_layer, transform_gn,
subsynapse)
return update