def build_network():
u_spec = leabra.UnitSpec(act_thr=0.5,
act_gain=100,
act_sd=0.005,
g_bar_e=1.0,
g_bar_l=0.1,
g_bar_i=1.0,
e_rev_e=1.0,
e_rev_l=0.3,
e_rev_i=0.25,
avg_l_min=0.2,
avg_l_init=0.4,
avg_l_gain=2.5,
adapt_on=False)
# layers
input0_layer = leabra.Layer(4,
unit_spec=u_spec,
name='input0_layer')
input1_layer = leabra.Layer(4,
unit_spec=u_spec,
name='input1_layer')
output_spec = leabra.LayerSpec(lay_inhib=False)
output_layer = leabra.Layer(4,
spec=output_spec,
unit_spec=u_spec,
name='output_layer')
# connections
conn_spec0 = leabra.ConnectionSpec(proj='1to1',
lrule=None,
rnd_mean=0.5,
rnd_var=0.0,
wt_scale_abs=1.0,
wt_scale_rel=1.0)
conn_spec1 = leabra.ConnectionSpec(proj='1to1',
lrule=None,
rnd_mean=0.5,
rnd_var=0.0,
wt_scale_abs=2.0,
wt_scale_rel=1.0)
conn0 = leabra.Connection(input0_layer,
output_layer,
spec=conn_spec0)
conn1 = leabra.Connection(input1_layer,
output_layer,
spec=conn_spec1)
# network
network = leabra.Network(
layers=[input0_layer, input1_layer, output_layer],
connections=[conn0, conn1])
network.set_inputs({
'input0_layer': [0.5, 0.95, 0.0, 0.25],
'input1_layer': [0.0, 0.5, 0.95, 0.75]
})
return network
def build_leabra_network(n_input, n_output, n_hidden, hidden_sizes=None, training_flag=None, quarter_size=50):
# specifications
learning_rule = 'leabra' if training_flag is True else None
unit_spec = leabra.UnitSpec(adapt_on=True, noisy_act=True)
layer_spec = leabra.LayerSpec(lay_inhib=True)
conn_spec = leabra.ConnectionSpec(proj='full', rnd_type='uniform', rnd_mean=0.75, rnd_var=0.2, lrule=learning_rule)
# input/outputs
input_layer = leabra.Layer(n_input, spec=layer_spec, unit_spec=unit_spec, name='input_layer')
output_layer = leabra.Layer(n_output, spec=layer_spec, unit_spec=unit_spec, name='output_layer')
# creating the required numbers of hidden layers and connections
layers = [input_layer]
connections = []
if isinstance(hidden_sizes, numbers.Number):
hidden_sizes = [hidden_sizes] * n_hidden
for i in range(n_hidden):
if hidden_sizes is not None:
hidden_size = hidden_sizes[i]
else:
hidden_size = n_input
hidden_layer = leabra.Layer(hidden_size, spec=layer_spec, unit_spec=unit_spec, name='hidden_layer_{}'.format(i))
hidden_conn = leabra.Connection(layers[-1], hidden_layer, spec=conn_spec)
layers.append(hidden_layer)
connections.append(hidden_conn)
last_conn = leabra.Connection(layers[-1], output_layer, spec=conn_spec)
connections.append(last_conn)
layers.append(output_layer)
network_spec = leabra.NetworkSpec(quarter_size=quarter_size)
network = leabra.Network(layers=layers, connections=connections, spec=network_spec)
return network
def test_sig_inv():
conn_spec = leabra.ConnectionSpec()
assert conn_spec.sig_inv(-1.0) == 0.0
assert conn_spec.sig_inv(0.0) == 0.0
assert conn_spec.sig_inv(0.5) == 0.5
assert conn_spec.sig_inv(1.0) == 1.0
assert conn_spec.sig_inv(2.0) == 1.0
def build_network(inhib, fixed_lrn_factor=None):
if fixed_lrn_factor is not None:
class UnitSpecFixedLrnFactor(leabra.UnitSpec):
def avg_l_lrn(self, unit):
return fixed_lrn_factor
unitspec_class = UnitSpecFixedLrnFactor
else:
unitspec_class = leabra.UnitSpec
u_spec = unitspec_class(act_thr=0.5,
act_gain=100,
act_sd=0.005,
g_bar_e=1.0,
g_bar_i=1.0,
g_bar_l=0.1,
e_rev_e=1.0,
e_rev_i=0.25,
e_rev_l=0.3,
avg_l_min=0.2,
avg_l_init=0.4,
avg_l_gain=2.5,
adapt_on=False)
input_layer = leabra.Layer(1,
unit_spec=u_spec,
genre=leabra.INPUT,
name='input_layer')
for unit in input_layer.units:
unit.log_names = log_names
unit.logs = {name: [] for name in unit.log_names}
output_spec = leabra.LayerSpec(lay_inhib=inhib,
g_i=1.8,
ff=1.0,
fb=1.0,
fb_dt=1 / 1.4,
ff0=0.1) #FIXME fb_tau
output_layer = leabra.Layer(1,
spec=output_spec,
unit_spec=u_spec,
genre=leabra.OUTPUT,
name='output_layer')
for unit in output_layer.units:
unit.log_names = log_names
unit.logs = {name: [] for name in unit.log_names}
conspec = leabra.ConnectionSpec(proj='full',
lrule='leabra',
lrate=0.04,
m_lrn=1.0,
rnd_mean=0.5,
rnd_var=0.0)
conn = leabra.Connection(input_layer, output_layer, spec=conspec)
network = leabra.Network(layers=[input_layer, output_layer],
connections=[conn])
network.set_inputs({'input_layer': [0.95]})
network.set_outputs({'output_layer': [0.95]})
return network
def build_network(n_input, n_output, n_hidden):
# specifications
unit_spec = leabra.UnitSpec(adapt_on=True, noisy_act=True)
inpout_layer_spec = leabra.LayerSpec(lay_inhib=True, g_i=2.0, ff=1, fb=0.5)
hidden_layer_spec = leabra.LayerSpec(lay_inhib=True, g_i=1.8, ff=1, fb=1)
conn_spec = leabra.ConnectionSpec(proj='full',
lrule='leabra',
lrate=0.04,
rnd_type='uniform',
rnd_mean=0.50,
rnd_var=0.25)
# input/outputs
input_layer = leabra.Layer(n_input,
spec=inpout_layer_spec,
unit_spec=unit_spec,
genre=leabra.INPUT,
name='input_layer')
output_layer = leabra.Layer(n_output,
spec=inpout_layer_spec,
unit_spec=unit_spec,
genre=leabra.OUTPUT,
name='output_layer')
# creating the required numbers of hidden layers and connections
layers = [input_layer]
connections = []
for i in range(n_hidden):
hidden_layer = leabra.Layer(n_input,
spec=hidden_layer_spec,
unit_spec=unit_spec,
genre=leabra.HIDDEN,
name='hidden_layer_{}'.format(i))
hidden_conn = leabra.Connection(layers[-1],
hidden_layer,
spec=conn_spec)
layers.append(hidden_layer)
connections.append(hidden_conn)
last_conn = leabra.Connection(layers[-1], output_layer, spec=conn_spec)
connections.append(last_conn)
layers.append(output_layer)
network_spec = leabra.NetworkSpec(quarter_size=25)
network = leabra.Network(layers=layers, connections=connections)
return network
def test_emergent_layer(self):
"""Test quantitative equivalence with emergent on a basic layer inhibition project."""
emergent_data = data.parse_unit('layer_fffb.dat')
unit_spec = leabra.UnitSpec(adapt_on=True,
noisy_act=True,
g_bar_e=0.3,
g_bar_l=0.3,
g_bar_i=1.0,
act_thr=0.5,
act_gain=40,
act_sd=0.01)
layer_spec = leabra.LayerSpec(g_i=0.4, ff=1.0, fb=0.5)
connection_spec = leabra.ConnectionSpec(proj='1to1',
rnd_mean=1.0,
rnd_var=0.0)
src_layer = leabra.Layer(10, spec=layer_spec, unit_spec=unit_spec)
dst_layer = leabra.Layer(10, spec=layer_spec, unit_spec=unit_spec)
connection0 = leabra.Connection(src_layer,
dst_layer,
spec=connection_spec)
connection0.wt_scale_rel_eff = 1.0 # because we don't use the Network that initialize this
# value here.
input_pattern = 5 * [1.0, 0.0]
for i in range(200):
if ((i >= 10) and (i < 160)):
src_layer.force_activity(input_pattern)
else:
src_layer.force_activity(10 * [0.0])
src_layer.cycle('minus')
connection0.cycle()
dst_layer.cycle('minus')
self.assertTrue(
quantitative_match(dst_layer.units[0].logs,
emergent_data,
rtol=2e-05,
atol=0))
def test_emergent_layer(self):
"""Test quantitative equivalence with emergent on a basic layer inhibition project."""
emergent_data = data.parse_unit('layer_1.txt')
unit_spec0 = leabra.UnitSpec(adapt_on=True, noisy_act=True)
layer_spec0 = leabra.LayerSpec(g_i=0.4, ff=1.0, fb=0.5)
connection_spec0 = leabra.ConnectionSpec(proj='1to1',
rnd_mean=1.0,
rnd_var=0.0)
src_layer = leabra.Layer(10, spec=layer_spec0, unit_spec=unit_spec0)
dst_layer = leabra.Layer(10, spec=layer_spec0, unit_spec=unit_spec0)
connection0 = leabra.Connection(src_layer,
dst_layer,
spec=connection_spec0)
input_pattern = 5 * [1.0, 0.0]
for i in range(200):
if ((i >= 10) and (i < 160)):
src_layer.force_activity(input_pattern)
else:
src_layer.force_activity(10 * [0.0])
src_layer.cycle()
connection0.cycle()
dst_layer.cycle()
check = True
for name in dst_layer.units[0].logs.keys():
for t, (py, em) in enumerate(
zip(dst_layer.units[0].logs[name], emergent_data[name])):
if not np.allclose(py, em, rtol=1e-05, atol=1e-07):
print('{}:{} [py] {:.10f} != {:.10f} [emergent]'.format(
name, t, py, em))
check = False
self.assertTrue(check)
def test_simple_usage(self):
"""Test the basic Network API"""
input_layer = leabra.Layer(4, name='input_layer')
output_spec = leabra.LayerSpec(g_i=1.5,
ff=1,
fb=0.5,
fb_dt=1 / 1.4,
ff0=0.1)
output_layer = leabra.Layer(2, spec=output_spec, name='output_layer')
conspec = leabra.ConnectionSpec(proj="full", lrule='leabra')
conn = leabra.Connection(input_layer, output_layer, spec=conspec)
network = leabra.Network(layers=[input_layer, output_layer],
connections=[conn])
network.set_inputs({'input_layer': [1.0, 1.0, 0.0, 0.0]})
network.set_outputs({'output_layer': [1.0, 0.0]})
for _ in range(20):
network.trial()
self.assertTrue(True)
def build_network(n):
log_names = ('net', 'I_net', 'v_m', 'act', 'v_m_eq', 'adapt',
'avg_ss', 'avg_s', 'avg_s_eff', 'avg_m', 'avg_l')
u_spec = leabra.UnitSpec(act_thr=0.5,
act_gain=100,
act_sd=0.005,
g_bar_e=1.0,
g_bar_l=0.1,
g_bar_i=1.0,
e_rev_e=1.0,
e_rev_l=0.3,
e_rev_i=0.25,
avg_l_min=0.2,
avg_l_init=0.4,
avg_l_gain=2.5,
adapt_on=False)
# layers
layer_spec = leabra.LayerSpec(lay_inhib=False)
input_layer = leabra.Layer(n,
spec=layer_spec,
unit_spec=u_spec,
genre=leabra.INPUT,
name='input_layer')
hidden_layer = leabra.Layer(n,
spec=layer_spec,
unit_spec=u_spec,
genre=leabra.HIDDEN,
name='hidden_layer')
output_layer = leabra.Layer(n,
spec=layer_spec,
unit_spec=u_spec,
genre=leabra.OUTPUT,
name='output_layer')
for layer in [input_layer, hidden_layer, output_layer]:
for unit in layer.units:
unit.log_names = log_names
unit.logs = {name: [] for name in unit.log_names}
# connections
weights = read_weights(
os.path.join(os.path.dirname(__file__),
'emergent_projects/leabra_std{}.wts'.format(n)))
inphid_conn_spec = leabra.ConnectionSpec(proj='full',
lrule='leabra',
lrate=0.04,
rnd_mean=0.5,
rnd_var=0.0,
wt_scale_abs=1.0,
wt_scale_rel=1.0)
hidout_conn_spec = leabra.ConnectionSpec(proj='full',
lrule='leabra',
lrate=0.04,
rnd_mean=0.5,
rnd_var=0.0,
wt_scale_abs=1.0,
wt_scale_rel=1.0)
inphid_conn = leabra.Connection(input_layer,
hidden_layer,
spec=inphid_conn_spec)
inphid_conn.weights = weights[('Input', 'Hidden')]
hidout_conn = leabra.Connection(hidden_layer,
output_layer,
spec=hidout_conn_spec)
hidout_conn.weights = weights[('Hidden', 'Output')]
# network
network = leabra.Network(
layers=[input_layer, hidden_layer, output_layer],
connections=[inphid_conn, hidout_conn])
n_sqrt = int(round(np.sqrt(n)))
network.set_inputs(
{'input_layer': [0.95] * n_sqrt + [0.0] * (n - n_sqrt)})
network.set_outputs(
{'output_layer':
[0.0] * (n - n_sqrt) + [0.95] * n_sqrt}) # FIXME 0.95 -> 1.0
return network
def test_simple_pattern_learning(self):
"""Quantitative test on the pair of neurons scenario"""
check = True
for inhib in [False, True]:
if inhib:
emergent_data = data.parse_weights('neuron_pair_inhib.txt')
else:
emergent_data = data.parse_weights('neuron_pair.txt')
u_spec = leabra.UnitSpec(act_thr=0.5,
act_gain=100,
act_sd=0.01,
g_bar_e=1.0,
g_bar_i=1.0,
g_bar_l=0.1,
e_rev_e=1.0,
e_rev_i=0.25,
e_rev_l=0.3,
avg_l_min=0.2,
avg_l_init=0.155,
avg_l_max=1.5,
adapt_on=False)
input_layer = leabra.Layer(1, unit_spec=u_spec, name='input_layer')
g_i = 1.5 if inhib else 0.0
output_spec = leabra.LayerSpec(g_i=g_i,
ff=1.0,
fb=0.5,
fb_dt=1 / 1.4,
ff0=0.1)
output_layer = leabra.Layer(1,
spec=output_spec,
unit_spec=u_spec,
name='output_layer')
for u in output_layer.units:
u.avg_l_lrn = 1.0
conspec = leabra.ConnectionSpec(proj='full',
lrule='leabra',
lrate=0.04,
m_lrn=0.0,
rnd_mean=0.5,
rnd_var=0.0)
conn = leabra.Connection(input_layer, output_layer, spec=conspec)
network = leabra.Network(layers=[input_layer, output_layer],
connections=[conn])
network.set_inputs({'input_layer': [0.95]})
network.set_outputs({'output_layer': [0.95]})
logs = {'wt': [], 'sse': []}
for t in range(50):
logs['wt'].append(conn.links[0].wt)
sse = network.trial()
logs['sse'].append(sse)
for name in ['wt', 'sse']:
for t, (py,
em) in enumerate(zip(logs[name], emergent_data[name])):
if not np.allclose(py, em, rtol=0, atol=1e-05):
print(
'{}:{:2d} [py] {:.10f} != {:.10f} [emergent] ({}inhib) diff={:g}'
.format(name, t, py, em, '' if inhib else 'no ',
py - em))
check = False
self.assertTrue(check)