Graded (lava-nc#734)

* prod neuron * trying to get prod neuron to work... * trying to get prod neuron cpu to work... * prod neuron process cpu backend working with unit test * remove init file from prod_neuron * gradedvec process and test * working on norm vec * fixed prod neuron license headers * invsqrt model and tests, reconfigured to process and models * normvecdelay and tests, timing weirdness with normvecdelay * test for second channel of norm vec * renamed to prodneuron. * fixing some linting errors * cleanup * adding some docstring, fixing unused imports * Fix conv python model to send() before recv() (lava-nc#751) Co-authored-by: Gavin Parpart <[email protected]> * Adds support for Monitor a Port to observe if it is blocked (lava-nc#755) * Adds support for Monitor a Port to observe if it is blocked * Fix lint issues * Redesigned Watchdog to use Multiprocessing Manager; Invoke only 2 Event Monitors and use 2 queues for watching events; Configs are piped in via compiler now * Incorporate Codacy Suggestions * Fix lint comments * Fix failing unit tests to add the watchdog builder * Code review comments * Set version to dev0 in pyproject.toml * Update README.md Updated version in install instructions. * Update README.md (lava-nc#758) Updated the installation branch to the most recent version. Co-authored-by: PhilippPlank <[email protected]> * Fix DelayDense buffer issue (lava-nc#767) * update refport unittest to always wait when it writes to port for consistent behavior Signed-off-by: bamsumit <[email protected]> * Removed pyproject changes Signed-off-by: bamsumit <[email protected]> * Fix to convolution tests. Fixed imcompatible mnist_pretrained for old python versions. Signed-off-by: bamsumit <[email protected]> * Missing moudle parent fix Signed-off-by: bamsumit <[email protected]> * Added ConvVarModel * Added iterable callback function Signed-off-by: bamsumit <[email protected]> * Fix codacy issues in callback_fx.py * Fix linting in callback_fx.py * Fix codacy sig issue in callback_fx.py * Bugfix to pass the args by keyword * Delay Dense PyModel fix Signed-off-by: bamsumit <[email protected]> * Fixed unittests Signed-off-by: bamsumit <[email protected]> * Fixed sparse delay Signed-off-by: bamsumit <[email protected]> --------- Signed-off-by: bamsumit <[email protected]> Co-authored-by: Joyesh Mishra <[email protected]> Co-authored-by: Marcus G K Williams <[email protected]> * Allow np.array as input weights for Sparse (lava-nc#772) * ndarray as input weights for Sparse * docs * codacy * remove implementation details from docstring and from tests * move tests to corresponding classes * put weight casting into extra method * Removed unused import --------- Co-authored-by: Mathis Richter <[email protected]> * Bump tornado from 6.3.2 to 6.3.3 (lava-nc#778) Bumps [tornado](https://github.com/tornadoweb/tornado) from 6.3.2 to 6.3.3. - [Changelog](https://github.com/tornadoweb/tornado/blob/master/docs/releases.rst) - [Commits](tornadoweb/tornado@v6.3.2...v6.3.3) --- updated-dependencies: - dependency-name: tornado dependency-type: indirect ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Bump cryptography from 41.0.2 to 41.0.3 (lava-nc#779) Bumps [cryptography](https://github.com/pyca/cryptography) from 41.0.2 to 41.0.3. - [Changelog](https://github.com/pyca/cryptography/blob/main/CHANGELOG.rst) - [Commits](pyca/cryptography@41.0.2...41.0.3) --- updated-dependencies: - dependency-name: cryptography dependency-type: indirect ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Mathis Richter <[email protected]> * small docstring, typing and other formatting changes * small docstring, typing and other formatting changes * doc strings for graded vec * removed prodneuron tests --------- Signed-off-by: bamsumit <[email protected]> Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: Gavin Parpart <[email protected]> Co-authored-by: Gavin Parpart <[email protected]> Co-authored-by: Joyesh Mishra <[email protected]> Co-authored-by: Marcus G K Williams <[email protected]> Co-authored-by: PhilippPlank <[email protected]> Co-authored-by: Alexander Henkes <[email protected]> Co-authored-by: bamsumit <[email protected]> Co-authored-by: Svea Marie Meyer <[email protected]> Co-authored-by: Mathis Richter <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
monkin77 · Oct 6, 2023 · 8a9861f · 8a9861f
1 parent 9cbe5a0
commit 8a9861f
Show file tree

Hide file tree

Showing 3 changed files with 596 additions and 0 deletions.
diff --git a/src/lava/proc/graded/models.py b/src/lava/proc/graded/models.py
@@ -0,0 +1,179 @@
+# 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.graded.process import GradedVec, NormVecDelay, InvSqrt
+
+
+class AbstractGradedVecModel(PyLoihiProcessModel):
+    """Implementation of GradedVec"""
+
+    a_in = None
+    s_out = None
+    v = None
+    vth = None
+    exp = None
+
+    def run_spk(self) -> None:
+        """The run function that performs the actual computation during
+        execution orchestrated by a PyLoihiProcessModel using the
+        LoihiProtocol.
+        """
+        a_in_data = self.a_in.recv()
+        self.v += a_in_data
+
+        is_spike = np.abs(self.v) > self.vth
+        sp_out = self.v * is_spike
+
+        self.v[:] = 0
+
+        self.s_out.send(sp_out)
+
+
+@implements(proc=GradedVec, protocol=LoihiProtocol)
+@requires(CPU)
+@tag('fixed_pt')
+class PyGradedVecModelFixed(AbstractGradedVecModel):
+    """Fixed point implementation of GradedVec"""
+    a_in = LavaPyType(PyInPort.VEC_DENSE, np.int32, precision=24)
+    s_out = LavaPyType(PyOutPort.VEC_DENSE, np.int32, precision=24)
+    vth: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+    v: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+    exp: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+
+
+@implements(proc=NormVecDelay, protocol=LoihiProtocol)
+@requires(CPU)
+@tag('fixed_pt')
+class NormVecDelayModel(PyLoihiProcessModel):
+    """Implementation of NormVecDelay. This process is typically part of
+    a network for normalization.
+    """
+    a_in1 = LavaPyType(PyInPort.VEC_DENSE, np.int32, precision=24)
+    a_in2 = LavaPyType(PyInPort.VEC_DENSE, np.int32, precision=24)
+    s_out = LavaPyType(PyOutPort.VEC_DENSE, np.int32, precision=24)
+    s2_out = LavaPyType(PyOutPort.VEC_DENSE, np.int32, precision=24)
+
+    vth: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+    exp: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+    v: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+    v2: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+
+    def run_spk(self) -> None:
+        """The run function that performs the actual computation during
+        execution orchestrated by a PyLoihiProcessModel using the
+        LoihiProtocol.
+        """
+        a_in_data1 = self.a_in1.recv()
+        a_in_data2 = self.a_in2.recv()
+
+        vsq = a_in_data1 ** 2
+        self.s2_out.send(vsq)
+
+        self.v2 = self.v
+        self.v = a_in_data1
+
+        output = self.v2 * a_in_data2
+
+        is_spike = np.abs(output) > self.vth
+        sp_out = output * is_spike
+
+        self.s_out.send(sp_out)
+
+
+@implements(proc=InvSqrt, protocol=LoihiProtocol)
+@requires(CPU)
+@tag('float')
+class InvSqrtModelFloat(PyLoihiProcessModel):
+    """Implementation of InvSqrt in floating point"""
+    a_in = LavaPyType(PyInPort.VEC_DENSE, float)
+    s_out = LavaPyType(PyOutPort.VEC_DENSE, float)
+
+    fp_base: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+
+    def run_spk(self) -> None:
+        """The run function that performs the actual computation during
+        execution orchestrated by a PyLoihiProcessModel using the
+        LoihiProtocol.
+        """
+        a_in_data = self.a_in.recv()
+        sp_out = 1 / (a_in_data ** 0.5)
+
+        self.s_out.send(sp_out)
+
+
+def make_fpinv_table(fp_base):
+    """
+    Creates the table for fp inverse algorithm.
+    """
+    n_bits = 24
+    B = 2**fp_base
+
+    Y_est = np.zeros((n_bits), dtype='int')
+    n_adj = 1.238982962
+
+    for m in range(n_bits):  # span the 24 bits, negate the decimal base
+        Y_est[n_bits - m - 1] = 2 * int(B / (2**((m - fp_base) / 2) * n_adj))
+
+    return Y_est
+
+
+def clz(val):
+    """
+    Count lead zeros.
+    """
+    return (24 - (int(np.log2(val)) + 1))
+
+
+def inv_sqrt(s_fp, n_iters=5, b_fraction=12):
+    """
+    Runs the fixed point inverse square root algorithm
+    """
+    Y_est = make_fpinv_table(b_fraction)
+    m = clz(s_fp)
+    b_i = int(s_fp)
+    Y_i = Y_est[m]
+    y_i = Y_i // 2
+
+    for _ in range(n_iters):
+        b_i = np.right_shift(np.right_shift(b_i * Y_i,
+                                            b_fraction + 1) * Y_i,
+                             b_fraction + 1)
+        Y_i = np.left_shift(3, b_fraction) - b_i
+        y_i = np.right_shift(y_i * Y_i, b_fraction + 1)
+
+    return y_i
+
+
+@implements(proc=InvSqrt, protocol=LoihiProtocol)
+@requires(CPU)
+@tag('fixed_pt')
+class InvSqrtModelFP(PyLoihiProcessModel):
+    """Implementation of InvSqrt in fixed point"""
+
+    a_in = LavaPyType(PyInPort.VEC_DENSE, np.int32, precision=24)
+    s_out = LavaPyType(PyOutPort.VEC_DENSE, np.int32, precision=24)
+    fp_base: np.ndarray = LavaPyType(np.ndarray, np.int32, precision=24)
+
+    def run_spk(self) -> None:
+        """The run function that performs the actual computation during
+        execution orchestrated by a PyLoihiProcessModel using the
+        LoihiProtocol.
+        """
+        a_in_data = self.a_in.recv()
+
+        if np.any(a_in_data) == 0:
+            sp_out = 0 * a_in_data
+        else:
+            sp_out = np.array([inv_sqrt(a_in_data, 5)])
+
+        self.s_out.send(sp_out)
diff --git a/src/lava/proc/graded/process.py b/src/lava/proc/graded/process.py
@@ -0,0 +1,144 @@
+# Copyright (C) 2023 Intel Corporation
+# SPDX-License-Identifier: BSD-3-Clause
+# See: https://spdx.org/licenses/
+
+import numpy as np
+import typing as ty
+
+from lava.magma.core.process.process import AbstractProcess
+from lava.magma.core.process.variable import Var
+from lava.magma.core.process.ports.ports import InPort, OutPort
+
+
+def loihi2round(vv):
+    """
+    Round values in numpy array the way loihi 2
+    performs rounding/truncation.
+    """
+    return np.fix(vv + (vv > 0) - 0.5).astype('int')
+
+
+class GradedVec(AbstractProcess):
+    """GradedVec
+    Graded spike vector layer. Transmits accumulated input as
+    graded spike with no dynamics.
+
+    v[t] = a_in
+    s_out = v[t] * (v[t] > vth)
+
+    Parameters
+    ----------
+    shape: tuple(int)
+        number and topology of neurons
+    vth: int
+        threshold for spiking
+    exp: int
+        fixed point base
+    """
+
+    def __init__(
+            self,
+            shape: ty.Tuple[int, ...],
+            vth: ty.Optional[int] = 1,
+            exp: ty.Optional[int] = 0) -> None:
+
+        super().__init__(shape=shape)
+
+        self.a_in = InPort(shape=shape)
+        self.s_out = OutPort(shape=shape)
+
+        self.v = Var(shape=shape, init=0)
+        self.vth = Var(shape=(1,), init=vth)
+        self.exp = Var(shape=(1,), init=exp)
+
+    @property
+    def shape(self) -> ty.Tuple[int, ...]:
+        """Return shape of the Process."""
+        return self.proc_params['shape']
+
+
+class NormVecDelay(AbstractProcess):
+    """NormVec
+    Normalizable graded spike vector. Used in conjunction with
+    InvSqrt process to create normalization layer.
+
+    When configured with InvSqrt, the process will output a normalized
+    vector with graded spike values. The output is delayed by 2 timesteps.
+
+    NormVecDelay has two input and two output channels. The process
+    outputs the square input values on the second channel to the InvSqrt
+    neuron. The process waits two timesteps to receive the inverse
+    square root value returned by the InvSqrt process. The value received
+    on the second input channel is multiplied by the primary input
+    value, and the result is output on the primary output channel.
+
+    v[t] = a_in1
+    s2_out = v[t] ** 2
+    s_out = v[t-2] * a_in2
+
+    Parameters
+    ----------
+    shape: tuple(int)
+        number and topology of neurons
+    vth: int
+        threshold for spiking
+    exp: int
+        fixed point base
+    """
+
+    def __init__(
+            self,
+            shape: ty.Tuple[int, ...],
+            vth: ty.Optional[int] = 1,
+            exp: ty.Optional[int] = 0) -> None:
+
+        super().__init__(shape=shape)
+
+        self.a_in1 = InPort(shape=shape)
+        self.a_in2 = InPort(shape=shape)
+
+        self.s_out = OutPort(shape=shape)
+        self.s2_out = OutPort(shape=shape)
+
+        self.vth = Var(shape=(1,), init=vth)
+        self.exp = Var(shape=(1,), init=exp)
+
+        self.v = Var(shape=shape, init=np.zeros(shape, 'int32'))
+        self.v2 = Var(shape=shape, init=np.zeros(shape, 'int32'))
+
+    @property
+    def shape(self) -> ty.Tuple[int, ...]:
+        """Return shape of the Process."""
+        return self.proc_params['shape']
+
+
+class InvSqrt(AbstractProcess):
+    """InvSqrt
+    Neuron model for computing inverse square root with 24-bit
+    fixed point values. Designed to be used in conjunction with
+    NormVecDelay.
+
+    v[t] = a_in
+    s_out = 1 / sqrt(v[t])
+
+    Parameters
+    ----------
+    fp_base : int
+        Base of the fixed-point representation
+    """
+
+    def __init__(
+            self,
+            shape: ty.Tuple[int, ...],
+            fp_base: ty.Optional[int] = 12) -> None:
+        super().__init__(shape=shape)
+
+        # base of the decimal point
+        self.fp_base = Var(shape=(1,), init=fp_base)
+        self.a_in = InPort(shape=shape)
+        self.s_out = OutPort(shape=shape)
+
+    @property
+    def shape(self) -> ty.Tuple[int, ...]:
+        """Return shape of the Process."""
+        return self.proc_params['shape']