The Initial version of CLP in CPU backend (lava-nc#707)

* CLP initial commit: PrototypeLIF, NoveltyDetector, Readout procs/tests

* small linting fix

* Novelty detector upgraded to target next neuron; codacy errors fixed

* integration test; small fixes

* removed duplicate code in prototypeLIF process; linting fixes

* linting fixes

* Linting and codacy fixes

* remove duplicate test; some more codacy fixes

* PrototypeLIF spikes when it recieves a 3rd factor input

* a test for PrototypeLIF output spike after 3rd factor input

* Allocation & prototype id tracking is abstracted away from
NoveltyDetector

* Allocator process; Readout proc sends allocation trigger if error

* introduce learning rate Var in PrototypeLIF

* updated integration tests; full system test included

* Linting fixes

* Another small lintint fix

* PrototypeLIF hard reset capability to enable faster temporal  WTA

* allocation mechanism changed; proc interfaces changes; dense conns
added; lr var removed

* small linting fix

* small codacy fix

* prints removed, spelling mistakes fixed

* ignoring one check in an integration test

* Revert "small linting fix"

This reverts commit bde4fa9.

* Fix linting in test_models.py

* Test fix in utils.py

* Fix test of bug fix in utils.py

* Fix utils.py

* Implemented individual threadsafe random call

Signed-off-by: bamsumit <[email protected]>

---------

Signed-off-by: bamsumit <[email protected]>
Co-authored-by: PhilippPlank <[email protected]>
Co-authored-by: Marcus G K Williams <[email protected]>
Co-authored-by: bamsumit <[email protected]>

src/lava/proc/clp/novelty_detector/models.py

@@ -0,0 +1,71 @@
+# Copyright (C) 2023 Intel Corporation
+# SPDX-License-Identifier: BSD-3-Clause
+# See: https://spdx.org/licenses/
+
+import numpy as np
+
+from lava.magma.core.sync.protocols.loihi_protocol import LoihiProtocol
+from lava.magma.core.model.py.ports import PyInPort, PyOutPort
+from lava.magma.core.model.py.type import LavaPyType
+from lava.magma.core.resources import CPU
+from lava.magma.core.decorator import implements, requires, tag
+from lava.magma.core.model.py.model import PyLoihiProcessModel
+
+from lava.proc.clp.novelty_detector.process import NoveltyDetector
+
+
+@implements(proc=NoveltyDetector, protocol=LoihiProtocol)
+@requires(CPU)
+@tag("fixed_pt", 'bit_accurate_loihi')
+class PyNoveltyDetectorModel(PyLoihiProcessModel):
+    """Python implementation of the NoveltyDetector process
+
+    """
+    input_aval_in: PyInPort = LavaPyType(PyInPort.VEC_DENSE, np.int32)
+    output_aval_in: PyInPort = LavaPyType(PyInPort.VEC_DENSE, np.int32)
+
+    novelty_detected_out: PyOutPort = LavaPyType(PyOutPort.VEC_DENSE, np.int32)
+    t_wait: np.int32 = LavaPyType(np.ndarray, np.int32, precision=32)
+
+    def __init__(self, proc_params):
+        super().__init__(proc_params)
+        self.waiting = False  # A variable to know if we are waiting for output
+        self.t_passed = 0  # The time passed since the injection of the input
+        self.novelty_detected = False
+
+    def run_spk(self) -> None:
+
+        # If input is available, we start to clock for waiting the output.
+        a_in = self.input_aval_in.recv()
+        if a_in != 0:
+            self.waiting = True
+            self.t_passed = 0
+
+        # If output available, that means the input is a known pattern,
+        # so we turn off waiting and reset
+        a_in = self.output_aval_in.recv()
+        if a_in != 0:
+            self.waiting = False
+            self.t_passed = 0
+
+        # If not, then we check whether the time limit has been passed for
+        # waiting. If so, we assume this is a novel pattern
+        elif self.t_passed > self.t_wait:
+            self.novelty_detected = True
+            self.waiting = False
+            self.t_passed = 0
+
+        # If we are still waiting, increment the time counter
+        if self.waiting:
+            self.t_passed += 1
+
+        # If we have detected novelty, send this signal downstream, and set
+        # the flag back to the False
+        if self.novelty_detected:
+            self.novelty_detected_out.send(np.array([1]))
+            self.novelty_detected = False
+            self.waiting = False
+
+        else:
+            # Otherwise, just send zeros (i.e. no signal)
+            self.novelty_detected_out.send(np.array([0]))

src/lava/proc/clp/novelty_detector/process.py

@@ -0,0 +1,44 @@
+# Copyright (C) 2023 Intel Corporation
+# SPDX-License-Identifier: BSD-3-Clause
+# See: https://spdx.org/licenses/
+
+from lava.magma.core.process.ports.ports import InPort, OutPort
+from lava.magma.core.process.process import AbstractProcess
+from lava.magma.core.process.variable import Var
+
+
+class NoveltyDetector(AbstractProcess):
+    """Novelty detection process.
+    This process detect the mismatch between the input injection to the
+    system and the output generation by the systems. If the system processes
+    an input, but does not generate an output (i.e. all the Prototype
+    neurons are silent) during a given time window after the beginning of
+    the input processing, then NoveltyDetector process will generate a
+    signal. This signal means that a novel (unknown) input is detected.
+
+    Parameters
+        ----------
+        t_wait : int
+            The amount of time the process will wait after receiving
+            signal about input injection to the system before sending out
+            novelty detection signal. If in this time window the system (the
+            Prototype neurons) generates an output, then the process will be
+            reset and NO novelty detection signal will be sent out.
+
+    """
+
+    def __init__(self, *,
+                 t_wait: int
+                 ) -> None:
+        super().__init__()
+
+        # An input is being processed by the system
+        self.input_aval_in = InPort(shape=(1,))
+
+        # An output is generated by the system
+        self.output_aval_in = InPort(shape=(1,))
+
+        # OutPort for sending out the novelty detection signal
+        self.novelty_detected_out = OutPort(shape=(1,))
+
+        self.t_wait = Var(shape=(1,), init=t_wait)

src/lava/proc/clp/nsm/models.py

@@ -0,0 +1,149 @@
+# Copyright (C) 2023 Intel Corporation
+# SPDX-License-Identifier: BSD-3-Clause
+# See: https://spdx.org/licenses/
+
+import numpy as np
+
+from lava.magma.core.sync.protocols.loihi_protocol import LoihiProtocol
+from lava.magma.core.model.py.ports import PyInPort, PyOutPort
+from lava.magma.core.model.py.type import LavaPyType
+from lava.magma.core.resources import CPU
+from lava.magma.core.decorator import implements, requires, tag
+from lava.magma.core.model.py.model import PyLoihiProcessModel
+
+from lava.proc.clp.nsm.process import Readout
+from lava.proc.clp.nsm.process import Allocator
+
+
+@implements(proc=Readout, protocol=LoihiProtocol)
+@requires(CPU)
+@tag("fixed_pt")
+class PyReadoutModel(PyLoihiProcessModel):
+    """Python implementation of the Readout process.
+    This process will run in super host and will be the main interface
+    process with the user.
+    """
+    inference_in: PyInPort = LavaPyType(PyInPort.VEC_DENSE, np.int32,
+                                        precision=24)
+    label_in: PyInPort = LavaPyType(PyInPort.VEC_DENSE, np.int32)
+
+    user_output: PyOutPort = LavaPyType(PyOutPort.VEC_DENSE, np.int32)
+    trigger_alloc: PyOutPort = LavaPyType(PyOutPort.VEC_DENSE, np.int32)
+    feedback: PyOutPort = LavaPyType(PyOutPort.VEC_DENSE, np.int32)
+    proto_labels: np.ndarray = LavaPyType(np.ndarray, np.int32)
+    last_winner_id: np.int32 = LavaPyType(np.ndarray, np.int32)
+
+    def run_spk(self) -> None:
+        # Read the output of the prototype neurons
+        output_vec = self.inference_in.recv()
+        # Read the user-provided label
+        user_label = self.label_in.recv()[0]
+        # Feedback about the correctness of prediction. +1 if correct,
+        # -1 if incorrect, 0 if no label is provided by the user at this point.
+        infer_check = 0
+
+        # If there is an active prototype neuron, this will temporarily store
+        # the label of that neuron
+        inferred_label = 0
+
+        # Flag for allocation trigger
+        allocation_trigger = False
+
+        # If any prototype neuron is active, then we go here. We assume there
+        # is only one neuron active in the prototype population
+        if output_vec.any():
+
+            # Find the id of the winner neuron and store it
+            winner_proto_id = np.nonzero(output_vec)[0][0]
+            self.last_winner_id = winner_proto_id
+
+            # Get the label of this neuron from the labels' list
+            inferred_label = self.proto_labels[winner_proto_id]
+
+            # If this label is zero, that means this prototype is not labeled.
+            if inferred_label == 0:
+                # So, we give a pseudo label to the unknown winner.
+                # These are negative temporary labels that is based on the id
+                # of the prototype and generated as follows.
+                self.proto_labels[winner_proto_id] = -1 * (winner_proto_id + 1)
+
+                # So now this pseudo-label is our inferred label.
+                inferred_label = self.proto_labels[winner_proto_id]
+
+        # Next we check if a user-provided label is available.
+        if user_label != 0:
+
+            # If so we need to access the most recent winner's label,
+            # assuming the temporal causality between the prediction by the
+            # system and the providence of the label;l by the user
+            last_inferred_label = self.proto_labels[self.last_winner_id]
+
+            # If the most recently predicted label (i.e. the one for the
+            # current input which is also the user-provided label refer to)
+            # is an actual label (not a pseudo one), then we check the
+            # correctness of the predicted label against user-provided one.
+
+            if last_inferred_label > 0:  # "Known Known class"
+                if last_inferred_label == user_label:
+                    infer_check = 1
+                else:
+                    # If the error occurs, trigger allocation by sending an
+                    # allocation signal
+                    infer_check = -1
+                    allocation_trigger = True
+
+            # If this prototype has a pseudo-label, then we label it with
+            # the user-provided label and do not send any feedback (because
+            # we did not have an actual prediction)
+
+            elif last_inferred_label < 0:  # "Known Unknown class"
+                self.proto_labels[self.last_winner_id] = user_label
+                inferred_label = user_label
+
+        # Send out the readout predicted label (if any) and the feedback
+        # about the correctness of this prediction after user providing the
+        # actual label
+        self.user_output.send(np.array([inferred_label]))
+        self.feedback.send(np.array([infer_check]))
+        if allocation_trigger:
+            self.trigger_alloc.send(np.array([1]))
+        else:
+            self.trigger_alloc.send(np.array([0]))
+
+
+@implements(proc=Allocator, protocol=LoihiProtocol)
+@requires(CPU)
+@tag("fixed_pt")
+class PyAllocatorModel(PyLoihiProcessModel):
+    """Python implementation of the Allocator process.
+    """
+
+    trigger_in: PyInPort = LavaPyType(PyInPort.VEC_DENSE, np.int32)
+    allocate_out: PyOutPort = LavaPyType(PyOutPort.VEC_DENSE, np.int32)
+    next_alloc_id: np.int32 = LavaPyType(np.ndarray, np.int32)
+    n_protos: np.int32 = LavaPyType(np.ndarray, np.int32)
+
+    def __init__(self, proc_params):
+        super().__init__(proc_params)
+
+    def run_spk(self) -> None:
+        # Allocation signal, initialized to a vector of zeros
+        alloc_signal = np.zeros(shape=self.allocate_out.shape, dtype=np.int32)
+
+        # Check the input, if a trigger for allocation is received then we
+        # send allocation signal to the next neuron
+        allocating = self.trigger_in.recv()[0]
+        if allocating:
+            # Choose the specific element of the OutPort to send allocate
+            # signal. This is a single graded spike that has the payload of
+            # the id of the next neuron to be allocated. Note that these id's
+            # are starting from id=1, as the graded value of zero means no
+            # signal. Hence, the initial value of next_alloc_id is one and
+            # after each allocation it is incremented by one
+            alloc_signal[0] = self.next_alloc_id
+
+            # Increment this counter to point to the next neuron
+            self.next_alloc_id += 1
+
+        # Otherwise, just send zeros (i.e. no signal)
+        self.allocate_out.send(alloc_signal)

src/lava/proc/clp/nsm/process.py

@@ -0,0 +1,87 @@
+# Copyright (C) 2023 Intel Corporation
+# SPDX-License-Identifier: BSD-3-Clause
+# See: https://spdx.org/licenses/
+
+import typing as ty
+import numpy as np
+
+from lava.magma.core.process.ports.ports import InPort, OutPort
+from lava.magma.core.process.process import AbstractProcess
+from lava.magma.core.process.variable import Var
+
+
+class Readout(AbstractProcess):
+    """ Readout process of the CLP system. It receives the output spikes from
+     PrototypeLIF neurons, look up the label of the winner prototype and
+     send it out to the user as the inferred label.
+     Additionally, if the winner neuron does not have a label this process
+     assigns a pseudo-label (a negative-valued label) for the time-being.
+     When a user-provided label is available, this process refer to the most
+     recent predicted label. If that is a pseudo-label, then it assigns the
+     user-provided label to this neuron. On the other hand if that is a
+     normal label (i.e. a positive number) then the process will check the
+     correctness of the predicted label and provide feedback through another
+     channel
+
+    Parameters
+        ----------
+        n_protos : int
+            Number of Prototype LIF neurons that this process need to read from.
+        proto_labels : numpy.ndarray, optional
+            Initial labels of the Prototype LIF neurons. If not provided,
+            by default this array will be initialized with zeros, meaning
+            they are not labelled.
+    """
+
+    def __init__(self, *,
+                 n_protos: int,
+                 proto_labels: ty.Optional[np.ndarray] = None) -> None:
+        # If not provided by the user initialize it to the zeros
+        if proto_labels is None:
+            proto_labels = np.zeros(shape=(n_protos,), dtype=int)
+
+        super().__init__(proto_labels=proto_labels, n_protos=n_protos)
+
+        self.inference_in = InPort(shape=(n_protos,))  # To read output spikes
+        self.label_in = InPort(shape=(1,))  # User-provided labels goes in here
+        self.user_output = OutPort(shape=(1,))  # Output for predicted labels
+
+        # Feedback to user about correctness of the prediction
+        self.feedback = OutPort(shape=(1,))
+        self.trigger_alloc = OutPort(shape=(1,))
+
+        # The array for the labels of the prototype neurons
+        self.proto_labels = Var(shape=(n_protos,), init=proto_labels)
+
+        # The id of the most recent winner prototype
+        self.last_winner_id = Var(shape=(1,), init=0)
+
+
+class Allocator(AbstractProcess):
+    """ Allocator process of CLP system. When triggered by other processes
+    it will send a one-hot-encoded allocation signal to the prototype
+    population, specifically targeting next neuron to be allocated. It holds
+    the reference to the id of the next neuron to be allocated.
+
+    Parameters
+        ----------
+        n_protos : int
+            The number of prototypes that this Allocator process can
+            target. Each time a allocation trigger input is received the
+            next unallocated prototype will be targeted by the output of the
+            Allocator process.
+    """
+
+    def __init__(self, *,
+                 n_protos: int) -> None:
+
+        super().__init__()
+
+        # Input for triggering allocation
+        self.trigger_in = InPort(shape=(1,))
+        # One-hot-encoded output for allocating specific prototype
+        self.allocate_out = OutPort(shape=(1,))
+
+        # The id of the next prototype to be allocated
+        self.next_alloc_id = Var(shape=(1,), init=1)
+        self.n_protos = Var(shape=(1,), init=n_protos)