MPC Quantized Machine Learning- Jacobi SVD

MPC Quantized Machine Learning- Jacobi SVD

sml/decomposition/BUILD.bazel

@@ -26,3 +26,8 @@ py_library(
     name = "nmf",
     srcs = ["nmf.py"],
 )
+
+py_library(
+    name = "jacobi_svd",
+    srcs = ["jacobi_svd.py"],
+)

sml/decomposition/emulations/BUILD.bazel

@@ -45,6 +45,15 @@ py_binary(
     ],
 )
 
+py_binary(
+    name = "jacobi_svd_emul",
+    srcs = ["jacobi_svd_emul.py"],
+    deps = [
+        "//sml/decomposition:jacobi_svd",
+        "//sml/utils:emulation",
+    ],
+)
+
 filegroup(
     name = "conf",
     srcs = [

sml/decomposition/emulations/jacobi_svd_emul.py

@@ -0,0 +1,91 @@
+import os
+import sys
+import jax.numpy as jnp
+import jax.random as random
+import jax.lax as lax
+import numpy as np
+from scipy.linalg import svd as scipy_svd
+
+sys.path.append(os.path.join(os.path.dirname(__file__), '../../../'))
+
+import sml.utils.emulation as emulation
+
+def generate_symmetric_matrix(n, seed=0):
+    A = random.normal(random.PRNGKey(seed), (n, n))
+    S = (A + A.T) / 2
+    return S
+
+def jacobi_rotation(A, p, q):
+    tau = (A[q, q] - A[p, p]) / (2 * A[p, q])
+    t = jnp.sign(tau) / (jnp.abs(tau) + jnp.sqrt(1 + tau**2))
+    c = 1 / jnp.sqrt(1 + t**2)
+    s = t * c
+    return c, s
+
+def apply_jacobi_rotation_A(A, c, s, p, q):
+    A_new = A.copy()
+    A = A.at[p, :].set(c * A_new[p, :] - s * A_new[q, :])
+    A = A.at[q, :].set(s * A_new[p, :] + c * A_new[q, :])
+    A_new = A.copy()
+    A = A.at[:, p].set(c * A_new[:, p] - s * A_new[:, q])
+    A = A.at[:, q].set(s * A_new[:, p] + c * A_new[:, q])
+    return A
+
+def jacobi_svd(A, tol=1e-10, max_iter=5):
+    n = A.shape[0]
+    A = jnp.array(A)
+
+    def body_fun(i, val):
+        A, max_off_diag = val
+        mask = jnp.abs(A - jnp.diagonal(A)) > tol  
+        for p in range(n):
+            for q in range(p + 1, n):
+                A = lax.cond(
+                    mask[p, q],
+                    lambda A: apply_jacobi_rotation_A(A, *jacobi_rotation(A, p, q), p, q),
+                    lambda A: A,
+                    A
+                )
+                max_off_diag = lax.cond(
+                    mask[p, q],
+                    lambda x: jnp.maximum(x, jnp.abs(A[p, q])),
+                    lambda x: x,
+                    max_off_diag
+                )       
+        return A, max_off_diag
+
+    max_off_diag = jnp.inf
+    A, _, = lax.fori_loop(0, max_iter, body_fun, (A, max_off_diag))
+
+    singular_values = jnp.abs(jnp.diag(A))
+    idx = jnp.argsort(-singular_values)
+    singular_values = singular_values[idx]
+
+    return singular_values
+
+def emul_jacobi_svd(mode=emulation.Mode.MULTIPROCESS):
+    print("Start Jacobi SVD emulation.")
+
+    def proc_transform(A):
+        singular_values = jacobi_svd(A)
+        return singular_values
+
+    try:
+        emulator = emulation.Emulator(
+            emulation.CLUSTER_ABY3_3PC, mode, bandwidth=300, latency=20
+        )
+        emulator.up()
+
+        A = generate_symmetric_matrix(10)
+        A_spu = emulator.seal(A)
+        singular_values = emulator.run(proc_transform)(A_spu)
+
+        _, singular_values_scipy, _ = scipy_svd(np.array(A), full_matrices=False)
+
+        np.testing.assert_allclose(np.sort(singular_values), np.sort(singular_values_scipy), atol=1e-3)
+
+    finally:
+        emulator.down()
+
+if __name__ == "__main__":
+    emul_jacobi_svd(emulation.Mode.MULTIPROCESS)

sml/decomposition/jacobi_svd.py

@@ -0,0 +1,82 @@
+import jax.numpy as jnp
+import jax.random as random
+import jax.lax as lax
+from jax import jit, vmap
+import numpy as np
+import time
+
+@jit
+def jacobi_rotation(A, p, q):
+    tau = (A[q, q] - A[p, p]) / (2 * A[p, q])
+    t = jnp.sign(tau) / (jnp.abs(tau) + jnp.sqrt(1 + tau**2))
+    c = 1 / jnp.sqrt(1 + t**2)
+    s = t * c
+    return c, s
+
+@jit
+def apply_jacobi_rotation_A(A, c, s, p, q):
+    A_new = A.copy()
+    A = A.at[p, :].set(c * A_new[p, :] - s * A_new[q, :])
+    A = A.at[q, :].set(s * A_new[p, :] + c * A_new[q, :])
+    A_new = A.copy()
+    A = A.at[:, p].set(c * A_new[:, p] - s * A_new[:, q])
+    A = A.at[:, q].set(s * A_new[:, p] + c * A_new[:, q])
+    return A
+
+@jit
+def jacobi_svd(A, tol=1e-10, max_iter=5):
+    n = A.shape[0]
+    A = jnp.array(A)
+
+    def body_fun(i, val):
+        A, max_off_diag, iterations = val
+        mask = jnp.abs(A - jnp.diagonal(A)) > tol  
+        for p in range(n):
+            for q in range(p + 1, n):
+                A = lax.cond(
+                    mask[p, q],
+                    lambda A: apply_jacobi_rotation_A(A, *jacobi_rotation(A, p, q), p, q),
+                    lambda A: A,
+                    A
+                )
+                max_off_diag = lax.cond(
+                    mask[p, q],
+                    lambda x: jnp.maximum(x, jnp.abs(A[p, q])),
+                    lambda x: x,
+                    max_off_diag
+                )       
+        return A, max_off_diag, iterations
+
+    max_off_diag = jnp.inf
+    iterations = 0
+    A, _, final_iterations = lax.fori_loop(0, max_iter, body_fun, (A, max_off_diag, iterations))
+
+    singular_values = jnp.abs(jnp.diag(A))
+    idx = jnp.argsort(-singular_values)
+    singular_values = singular_values[idx]
+
+    return singular_values
+
+def generate_symmetric_matrix(n, seed=0):
+    A = random.normal(random.PRNGKey(seed), (n, n))
+    S = (A + A.T) / 2
+    return S
+
+n = 10
+
+A_jax = generate_symmetric_matrix(n)
+
+start_time = time.time()
+singular_values = jacobi_svd(A_jax)
+end_time = time.time()
+
+elapsed_time = end_time - start_time
+print(f"Run Time: {elapsed_time:.6f} s")
+
+print("Singular Values Jacobi_svd:")
+print(singular_values)
+
+A_np = np.array(A_jax)
+_, Sigma, _ = np.linalg.svd(A_np)
+print("Sigma:")
+print(Sigma)