def test_layer_hard_clamping_respects_clamp_max() -> None:
layer = lr.Layer(name="in", size=3, spec=sp.LayerSpec(clamp_max=0.5))
layer.hard_clamp(act_ext=[0, 1])
layer.handle(ev.HardClamp(layer_name="lr1", acts=[0, 1]))
expected = [0, 0.5, 0]
for i in range(len(expected)):
assert math.isclose(layer.units.act[i], expected[i], abs_tol=1e-6)
def test_kwta_suppresses_all_but_k_units() -> None:
n = net.Net()
n.new_layer(name="lr1", size=1)
lr2_spec = sp.LayerSpec(
inhibition_type="kwta",
k=2,
log_on_cycle=("unit_act", ),
unit_spec=sp.UnitSpec(adapt_dt=0, spike_gain=0))
n.new_layer(name="lr2", size=3, spec=lr2_spec)
pr1_spec = sp.ProjnSpec(dist=rn.Scalar(0.3))
n.new_projn("proj1", "lr1", "lr2", pr1_spec)
pr2_spec = sp.ProjnSpec(dist=rn.Scalar(0.5), post_mask=[0, 1, 1])
n.new_projn("proj2", "lr1", "lr2", pr2_spec)
n.force_layer("lr1", [1])
for i in range(100):
n.cycle()
logs = n.logs("cycle", "lr2")
acts = logs[logs.time == 99]["act"]
assert (acts > 0.8).sum() == 2
def test_net_can_pause_and_resume_logging() -> None:
n = net.Net()
n.new_layer(
"layer1",
2,
spec=specs.LayerSpec(log_on_cycle=(
"unit_act",
"avg_act",
)))
for i in range(2):
n.cycle()
n.pause_logging()
for i in range(2):
n.cycle()
n.resume_logging()
for i in range(2):
n.cycle()
parts_time = torch.Tensor(n.logs("cycle", "layer1").parts["time"])
whole_time = torch.Tensor(n.logs("cycle", "layer1").whole["time"])
assert list(parts_time.size()) == [8]
assert list(whole_time.size()) == [4]
assert (parts_time == torch.Tensor([0, 0, 1, 1, 4, 4, 5, 5])).all()
assert (whole_time == torch.Tensor([0, 1, 4, 5])).all()
def test_net_trial_log_pausing_and_resuming() -> None:
n = net.Net()
n.new_layer(
"layer1",
2,
spec=specs.LayerSpec(log_on_trial=(
"unit_act",
"avg_act",
)))
n.phase_cycle(phase=events.MinusPhase, num_cycles=5)
n.phase_cycle(phase=events.PlusPhase, num_cycles=5)
n.end_trial()
n.pause_logging("trial")
n.phase_cycle(phase=events.MinusPhase, num_cycles=5)
n.phase_cycle(phase=events.PlusPhase, num_cycles=5)
n.end_trial()
n.resume_logging("trial")
n.phase_cycle(phase=events.MinusPhase, num_cycles=5)
n.phase_cycle(phase=events.PlusPhase, num_cycles=5)
n.end_trial()
parts_time = torch.Tensor(n.logs("trial", "layer1").parts["time"])
whole_time = torch.Tensor(n.logs("trial", "layer1").whole["time"])
assert list(parts_time.size()) == [4]
assert list(whole_time.size()) == [2]
assert (parts_time == torch.Tensor([0, 0, 2, 2])).all()
assert (whole_time == torch.Tensor([0, 2])).all()
def __init__(self, name: str, size: int,
spec: specs.LayerSpec = None) -> None:
self._name = name
self.size = size
if spec is None:
self._spec = specs.LayerSpec()
else:
self._spec = spec
self.units = unit.UnitGroup(size=size, spec=self.spec.unit_spec)
# Feedback inhibition
self.fbi = 0.0
# Global inhibition
self.gc_i = 0.0
# Is the layer activation clamped?
self.clamped = False
# Is this a hidden layer? (i.e. has never been clamped)
self.hidden = True
# Set k units for inhibition
self.k = max(1, int(round(self.size * self.spec.kwta_pct)))
# Desired clamping values
self.act_ext = torch.Tensor(self.size).zero_()
# Last plus phase activation
self.acts_p = torch.Tensor(self.size).zero_()
# Last minus phase activation
self.acts_m = torch.Tensor(self.size).zero_()
# Cosine similarity between acts_p and acts_m
self.cos_diff = 0.0
# Cosine similiarity between acts_p and acts_m, integrated over trials
self.cos_diff_avg = 0.0
# The following two buffers are filled every time self.add_input() is
# called, and reset at the end of self.activation_cycle()
# Net input (excitation) input buffer. For every cycle, we
# store the layer inputs here. Once we have all the inputs, we
# normalize by wt_scale_rel_sum and send to the unit group.
self.input_buffer = torch.Tensor(self.size).zero_()
# Sum of the wt_scale_rel parameters for each projection terminating in
# this layer. We use this to normalize the inputs before propagating to
# unit group
self.wt_scale_rel_sum = 0.0
# When adding any loggable attribute or property to these lists, update
# specs.LayerSpec._valid_log_on_cycle (we represent in two places to
# avoid a circular dependency)
whole_attrs: List[str] = ["avg_act", "avg_net", "cos_diff_avg", "fbi"]
parts_attrs: List[str] = [
"unit_net", "unit_net_raw", "unit_gc_i", "unit_act", "unit_i_net",
"unit_i_net_r", "unit_v_m", "unit_v_m_eq", "unit_adapt",
"unit_spike"
]
super().__init__(whole_attrs=whole_attrs, parts_attrs=parts_attrs)
def test_attrs_to_log_gets_the_attrs_to_log_by_frequency() -> None:
spec = sp.LayerSpec(log_on_cycle=("unit_act", ),
log_on_trial=("unit_v_m", ),
log_on_epoch=("unit_spike", ),
log_on_batch=("unit_i_net", ))
assert spec.attrs_to_log(ev.CycleFreq) == ("unit_act", )
assert spec.attrs_to_log(ev.TrialFreq) == ("unit_v_m", )
assert spec.attrs_to_log(ev.EpochFreq) == ("unit_spike", )
assert spec.attrs_to_log(ev.BatchFreq) == ("unit_i_net", )
def test_you_can_retrieve_the_logs_for_a_layer() -> None:
n = net.Net()
n.new_layer(name="layer1",
size=3,
spec=specs.LayerSpec(log_on_cycle=("avg_act", ),
log_on_trial=("avg_act", ),
log_on_epoch=("avg_act", ),
log_on_batch=("avg_act", )))
n.plus_phase_cycle(1)
n.end_epoch()
n.end_batch()
for freq in ("cycle", "trial", "epoch", "batch"):
assert "avg_act" in n.logs(freq, "layer1").whole.columns
def __init__(self, name: str, size: int,
spec: specs.LayerSpec = None) -> None:
self.size = size
if spec is None:
self.spec = specs.LayerSpec()
else:
self.spec = spec
self.units = [unit.Unit(self.spec.unit_spec) for _ in range(size)]
self.fbi = 0.0
super().__init__(name)
def test_projn_can_uninhibit_flush() -> None:
pre = lr.Layer("lr1", size=1, spec=sp.LayerSpec(clamp_max=1))
post = lr.Layer("lr2", size=1)
projn = pr.Projn("proj", pre, post)
pre.clamp(act_ext=[1])
projn.handle(ev.InhibitProjns("proj"))
projn.flush()
projn.handle(ev.UninhibitProjns("proj"))
projn.flush()
assert post.input_buffer == 0.5
def test_layer_soft_clamping_equivalent_to_input() -> None:
layer1 = lr.Layer(name="test", size=3)
layer2 = lr.Layer(name="alt", size=3, spec=sp.LayerSpec(clamp_max=1))
layer2.clamp([0, 1, 0], hard=False)
for i in range(50):
layer1.add_input(torch.Tensor([0, 1, 0]))
layer1.activation_cycle()
layer2.activation_cycle()
print(layer1.units.act)
print(layer2.units.act)
for i in range(3):
assert math.isclose(layer1.units.act[i],
layer2.units.act[i],
abs_tol=1e-6)
def test_network_can_be_retrieved_and_continue_logging() -> None:
n = net.Net()
n.new_layer("layer1",
2,
spec=specs.LayerSpec(log_on_cycle=(
"unit_act",
"avg_act",
)))
for i in range(2):
n.cycle()
n.pause_logging()
location = "tests/mynet.pkl"
n.save(location)
m = net.Net()
m.load(filename=location)
before_parts_n = n.logs("cycle", "layer1").parts
before_parts_m = m.logs("cycle", "layer1").parts
before_whole_n = n.logs("cycle", "layer1").whole
before_whole_m = m.logs("cycle", "layer1").whole
assert np.all((before_parts_n == before_parts_m).values)
assert np.all((before_whole_n == before_whole_m).values)
for i in range(2):
m.cycle()
m.resume_logging()
for i in range(2):
m.cycle()
after_parts_time = torch.Tensor(m.logs("cycle", "layer1").parts["time"])
after_whole_time = torch.Tensor(m.logs("cycle", "layer1").whole["time"])
assert list(after_parts_time.size()) == [8]
assert list(after_whole_time.size()) == [4]
assert (after_parts_time == torch.Tensor([0, 0, 1, 1, 4, 4, 5, 5])).all()
assert (after_whole_time == torch.Tensor([0, 1, 4, 5])).all()
def test_net_batch_log_pausing_and_resuming() -> None:
n = net.Net()
n.new_layer("layer1",
2,
spec=specs.LayerSpec(log_on_batch=(
"unit_act",
"avg_act",
)))
n.end_batch()
n.pause_logging("batch")
n.end_batch()
n.resume_logging("batch")
n.end_batch()
parts_time = torch.Tensor(n.logs("batch", "layer1").parts["time"])
whole_time = torch.Tensor(n.logs("batch", "layer1").whole["time"])
assert list(parts_time.size()) == [4]
assert list(whole_time.size()) == [2]
assert (parts_time == torch.Tensor([0, 0, 2, 2])).all()
assert (whole_time == torch.Tensor([0, 2])).all()
def test_it_should_check_for_invalid_log_on_batch_attrs() -> None:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(log_on_batch=("whales", )).validate()
def test_every_valid_log_on_cycle_attribute_can_be_logged() -> None:
valid_attrs = sp.LayerSpec()._valid_attrs_to_log
lr1 = lr.Layer("lr1", 3)
for attr in valid_attrs:
lr1.validate_attr(attr)
def test_it_should_validate_the_unit_spec() -> None:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(unit_spec=sp.UnitSpec(vm_dt=-1)).validate()
def test_it_should_validate_global_inhibition_multiplier(f) -> None:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(gi=f).validate()
def test_it_should_validate_feedback_integration_constant(f) -> None:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(fb_dt=f).validate()
def test_it_should_validate_feedback_inhibition_multiplier(f) -> None:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(fb=f).validate()
def test_it_should_validate_feedforward_inhibition_multiplier(f):
with pytest.raises(sp.ValidationError):
sp.LayerSpec(ff=f).validate()
def test_it_should_check_for_invalid_clamp_max(f) -> None:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(clamp_max=f).validate()
def test_layer_init_uses_the_spec_you_pass_it() -> None:
spec = sp.LayerSpec()
layer = lr.Layer(name="in", spec=spec, size=1)
assert layer.spec is spec
def test_layer_should_be_able_to_update_its_units_kwta_avg_inhibition(
) -> None:
layer_spec = sp.LayerSpec(inhibition_type="kwta_avg")
layer = lr.Layer(name="in", size=3, spec=layer_spec)
layer.update_inhibition()
def test_a_new_layer_validates_its_spec() -> None:
n = net.Net()
with pytest.raises(specs.ValidationError):
n.new_layer("layer1", 3, spec=specs.LayerSpec(integ=-1))
def test_you_can_retrieve_the_logs_for_a_layer():
n = net.Net()
n.new_layer("layer1", 3, spec=specs.LayerSpec(log_on_cycle=("avg_act", )))
n.cycle()
assert "avg_act" in n.logs("cycle", "layer1").columns
def test_observing_unlogged_attr_raises_error_if_obj_not_observable() -> None:
n = net.Net()
n.new_layer("layer1", 1, spec=specs.LayerSpec(log_on_cycle=("avg_act", )))
with pytest.raises(ValueError):
n.observe("layer1_cycle_logger", "cos_diff_avg")
def test_layer_spec_validates_inhibition_type(f) -> None:
if f not in ["kwta", "kwta_avg", "fffb", "none"]:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(inhibition_type=f).validate()
else:
sp.LayerSpec(inhibition_type=f).validate()
def test_layer_spec_validates_kwta_pt(f) -> None:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(kwta_pt=f).validate()
def test_layer_spec_validates_integ(f) -> None:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(integ=f).validate()
def test_net_multiple_freq_log_pausing_and_resuming() -> None:
n = net.Net()
loggables = ("unit_act", "avg_act")
n.new_layer(
"layer1",
2,
spec=specs.LayerSpec(
log_on_cycle=loggables,
log_on_epoch=loggables,
log_on_trial=loggables,
log_on_batch=loggables))
n.cycle()
n.end_trial()
n.end_epoch()
n.end_batch()
n.pause_logging("cycle", "trial", "epoch", "batch")
n.cycle()
n.end_trial()
n.end_epoch()
n.end_batch()
n.resume_logging("cycle", "trial", "epoch", "batch")
n.cycle()
n.end_trial()
n.end_epoch()
n.end_batch()
parts_time = torch.Tensor(n.logs("cycle", "layer1").parts["time"])
whole_time = torch.Tensor(n.logs("cycle", "layer1").whole["time"])
assert list(parts_time.size()) == [4]
assert list(whole_time.size()) == [2]
assert (parts_time == torch.Tensor([0, 0, 2, 2])).all()
assert (whole_time == torch.Tensor([0, 2])).all()
parts_time = torch.Tensor(n.logs("trial", "layer1").parts["time"])
whole_time = torch.Tensor(n.logs("trial", "layer1").whole["time"])
assert list(parts_time.size()) == [4]
assert list(whole_time.size()) == [2]
assert (parts_time == torch.Tensor([0, 0, 2, 2])).all()
assert (whole_time == torch.Tensor([0, 2])).all()
parts_time = torch.Tensor(n.logs("epoch", "layer1").parts["time"])
whole_time = torch.Tensor(n.logs("epoch", "layer1").whole["time"])
assert list(parts_time.size()) == [4]
assert list(whole_time.size()) == [2]
assert (parts_time == torch.Tensor([0, 0, 2, 2])).all()
assert (whole_time == torch.Tensor([0, 2])).all()
parts_time = torch.Tensor(n.logs("batch", "layer1").parts["time"])
whole_time = torch.Tensor(n.logs("batch", "layer1").whole["time"])
assert list(parts_time.size()) == [4]
assert list(whole_time.size()) == [2]
assert (parts_time == torch.Tensor([0, 0, 2, 2])).all()
assert (whole_time == torch.Tensor([0, 2])).all()
def test_layer_spec_validates_k(f) -> None:
if f < 1:
with pytest.raises(sp.ValidationError):
sp.LayerSpec(inhibition_type="kwta", k=f).validate()