Merge pull request #2 from UM-ARM-Lab/refactor/torch-compile

Refactor to improve performance/torch compile compatibility

Author: LemonPi

benchmarks/bench_compile.py

@@ -0,0 +1,217 @@
+#!/usr/bin/env python
+"""Benchmark script for arm_pytorch_utilities torch.compile optimization.
+
+Measures eager-mode and torch.compile performance for refactor-target functions.
+Outputs a printed table and JSON file for before/after comparison.
+
+Usage:
+    python benchmarks/bench_compile.py --device cpu
+    python benchmarks/bench_compile.py --device cuda
+"""
+import argparse
+import json
+import time
+from datetime import datetime
+from pathlib import Path
+
+import torch
+
+from arm_pytorch_utilities import math_utils, linalg, tensor_utils, preprocess, softknn
+
+
+def bench(fn, *args, warmup=5, repeats=20, device='cpu'):
+    """Time a function with warmup and repeats. Returns median time in ms."""
+    for _ in range(warmup):
+        fn(*args)
+    if device == 'cuda':
+        torch.cuda.synchronize()
+
+    times = []
+    for _ in range(repeats):
+        if device == 'cuda':
+            torch.cuda.synchronize()
+        t0 = time.perf_counter()
+        fn(*args)
+        if device == 'cuda':
+            torch.cuda.synchronize()
+        t1 = time.perf_counter()
+        times.append((t1 - t0) * 1000)
+
+    times.sort()
+    return times[len(times) // 2]
+
+
+def try_compile_bench(fn, *args, device='cpu', warmup=5, repeats=20):
+    """Try to compile and benchmark a function. Returns (time_ms, success) or (None, False)."""
+    try:
+        compiled_fn = torch.compile(fn, fullgraph=True)
+        # Warmup includes compilation
+        for _ in range(warmup):
+            compiled_fn(*args)
+        if device == 'cuda':
+            torch.cuda.synchronize()
+
+        times = []
+        for _ in range(repeats):
+            if device == 'cuda':
+                torch.cuda.synchronize()
+            t0 = time.perf_counter()
+            compiled_fn(*args)
+            if device == 'cuda':
+                torch.cuda.synchronize()
+            t1 = time.perf_counter()
+            times.append((t1 - t0) * 1000)
+
+        times.sort()
+        return times[len(times) // 2], True
+    except Exception as e:
+        return None, False, str(e)
+
+
+def make_psd(n, device):
+    R = torch.randn(n, n, device=device)
+    return R.t() @ R + torch.eye(n, device=device) * 0.1
+
+
+def run_benchmarks(device_str):
+    device = torch.device(device_str)
+    results = {}
+
+    benchmarks = {}
+
+    # --- Setup inputs ---
+    # replace_nan_and_inf
+    x_nan = torch.randn(10000, 100, device=device)
+    mask = torch.rand_like(x_nan) < 0.1
+    x_nan[mask] = float('nan')
+    benchmarks['replace_nan_and_inf'] = (math_utils.replace_nan_and_inf, (x_nan.clone(), 0), True)
+
+    # angular_diff_batch
+    a_ang = torch.randn(100000, device=device)
+    b_ang = torch.randn(100000, device=device)
+    benchmarks['angular_diff_batch'] = (math_utils.angular_diff_batch, (a_ang, b_ang), False)
+
+    # angle_between_stable
+    u_abs = torch.randn(200, 50, device=device)
+    v_abs = torch.randn(150, 50, device=device)
+    benchmarks['angle_between_stable'] = (math_utils.angle_between_stable, (u_abs, v_abs), False)
+
+    # cos_sim_pairwise
+    x1_cos = torch.randn(500, 50, device=device)
+    x2_cos = torch.randn(300, 50, device=device)
+    benchmarks['cos_sim_pairwise'] = (math_utils.cos_sim_pairwise, (x1_cos, x2_cos), False)
+
+    # batch_batch_product
+    X_bbp = torch.randn(10000, 20, device=device)
+    A_bbp = torch.randn(10000, 20, 20, device=device)
+    benchmarks['batch_batch_product'] = (linalg.batch_batch_product, (X_bbp, A_bbp), False)
+
+    # batch_quadratic_product
+    X_bqp = torch.randn(10000, 20, device=device)
+    A_bqp = make_psd(20, device)
+    benchmarks['batch_quadratic_product'] = (linalg.batch_quadratic_product, (X_bqp, A_bqp), False)
+
+    # batch_outer_product
+    u_bop = torch.randn(10000, 20, device=device)
+    v_bop = torch.randn(10000, 20, device=device)
+    benchmarks['batch_outer_product'] = (linalg.batch_outer_product, (u_bop, v_bop), False)
+
+    # squeeze_n
+    x_sq = torch.randn(1, 1, 1, 1000, 50, device=device)
+    benchmarks['squeeze_n'] = (tensor_utils.squeeze_n, (x_sq, 3), False)
+
+    # MinMaxScaler.transform
+    x_mm = torch.randn(10000, 50, device=device)
+    scaler = preprocess.MinMaxScaler()
+    scaler.fit(x_mm)
+    benchmarks['MinMaxScaler.transform'] = (scaler.transform, (x_mm,), False)
+
+    # SoftKNN.forward
+    x_knn = torch.randn(200, 10, device=device)
+    knn = softknn.SoftKNN(min_k=20)
+    benchmarks['SoftKNN.forward'] = (knn, (x_knn,), False)
+
+    # sqrtm (CPU only due to .numpy())
+    if device_str == 'cpu':
+        A_sqrtm = make_psd(50, device)
+        benchmarks['sqrtm'] = (linalg.sqrtm, (A_sqrtm,), False)
+
+    # --- Run benchmarks ---
+    print(f"\n{'Function':<30} {'Eager (ms)':>12} {'Compile (ms)':>14} {'Speedup':>10} {'Compile OK':>12}")
+    print("-" * 80)
+
+    for name, (fn, args, needs_clone) in benchmarks.items():
+        # Eager benchmark
+        if needs_clone:
+            # For in-place functions, clone first arg each call
+            template = args[0]
+            rest_args = args[1:]
+
+            def cloning_fn(*a, _fn=fn, _tpl=template, _rest=rest_args):
+                return _fn(_tpl.clone(), *_rest)
+
+            eager_ms = bench(cloning_fn, warmup=5, repeats=20, device=device_str)
+        else:
+            try:
+                eager_ms = bench(fn, *args, device=device_str)
+            except Exception as e:
+                print(f"{name:<30} {'ERROR':>12} {'N/A':>14} {'N/A':>10} {'N/A':>12}  ({e})")
+                results[name] = {'eager_ms': None, 'compile_ms': None, 'compile_ok': False,
+                                 'compile_error': None, 'eager_error': str(e)}
+                continue
+
+        # Compile benchmark
+        compile_result = try_compile_bench(fn, *args, device=device_str)
+
+        if len(compile_result) == 2:
+            compile_ms, compile_ok = compile_result
+            compile_err = None
+        else:
+            compile_ms, compile_ok, compile_err = compile_result
+
+        speedup = f"{eager_ms / compile_ms:.2f}x" if compile_ms else "N/A"
+        compile_str = f"{compile_ms:.3f}" if compile_ms else "FAIL"
+
+        print(f"{name:<30} {eager_ms:>12.3f} {compile_str:>14} {speedup:>10} {'yes' if compile_ok else 'no':>12}")
+
+        results[name] = {
+            'eager_ms': eager_ms,
+            'compile_ms': compile_ms,
+            'compile_ok': compile_ok,
+            'compile_error': compile_err,
+        }
+
+    return results
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Benchmark arm_pytorch_utilities functions')
+    parser.add_argument('--device', choices=['cpu', 'cuda'], default='cpu')
+    args = parser.parse_args()
+
+    if args.device == 'cuda' and not torch.cuda.is_available():
+        print("CUDA not available, falling back to CPU")
+        args.device = 'cpu'
+
+    print(f"Running benchmarks on {args.device}")
+    print(f"PyTorch version: {torch.__version__}")
+
+    results = run_benchmarks(args.device)
+
+    # Save JSON
+    output_dir = Path(__file__).parent
+    timestamp = datetime.now().strftime('%Y%m%d_%H%M%S')
+    output_file = output_dir / f"results_{args.device}_{timestamp}.json"
+    output = {
+        'device': args.device,
+        'torch_version': torch.__version__,
+        'timestamp': timestamp,
+        'results': results,
+    }
+    with open(output_file, 'w') as f:
+        json.dump(output, f, indent=2)
+    print(f"\nResults saved to {output_file}")
+
+
+if __name__ == '__main__':
+    main()

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "arm_pytorch_utilities"
-version = "0.4.3"
+version = "0.5.0"
 description = "Utilities for working with pytorch"
 readme = "README.md" # Optional
 
@@ -68,7 +68,7 @@ dependencies = [# Optional
 # Similar to `dependencies` above, these must be valid existing
 # projects.
 [project.optional-dependencies] # Optional
-test = ["pytest"]
+test = ["pytest", "pytest-benchmark"]
 
 # List URLs that are relevant to your project
 #

src/arm_pytorch_utilities/linalg.py

@@ -42,20 +42,7 @@ def kronecker_product(t1, t2):
     Computes the Kronecker product between two tensors.
     See https://en.wikipedia.org/wiki/Kronecker_product
     """
-    t1_height, t1_width = t1.size()
-    t2_height, t2_width = t2.size()
-    out_height = t1_height * t2_height
-    out_width = t1_width * t2_width
-
-    tiled_t2 = t2.repeat(t1_height, t1_width)
-    expanded_t1 = (
-        t1.unsqueeze(2)
-            .unsqueeze(3)
-            .repeat(1, t2_height, t2_width, 1)
-            .view(out_height, out_width)
-    )
-
-    return expanded_t1 * tiled_t2
+    return torch.kron(t1, t2)
 
 
 def cov(x, rowvar=False, bias=False, ddof=None, aweights=None):
@@ -115,7 +102,7 @@ class MatrixSquareRoot(Function):
 
     @staticmethod
     def forward(ctx, input):
-        m = input.detach().numpy().astype(np.float_)
+        m = input.detach().numpy().astype(np.float64)
         sqrtm = torch.from_numpy(scipy.linalg.sqrtm(m).real).type_as(input)
         ctx.save_for_backward(sqrtm)
         return sqrtm
@@ -125,8 +112,8 @@ def backward(ctx, grad_output):
         grad_input = None
         if ctx.needs_input_grad[0]:
             sqrtm, = ctx.saved_tensors
-            sqrtm = sqrtm.data.numpy().astype(np.float_)
-            gm = grad_output.data.numpy().astype(np.float_)
+            sqrtm = sqrtm.data.numpy().astype(np.float64)
+            gm = grad_output.data.numpy().astype(np.float64)
 
             # Given a positive semi-definite matrix X,
             # since X = X^{1/2}X^{1/2}, we can compute the gradient of the
@@ -152,16 +139,16 @@ def forward(ctx, A, b):
         # A: (..., M, N)
         # b: (..., M, K)
         # https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/linalg_ops.py#L267
-        u = torch.cholesky(torch.matmul(A.transpose(-1, -2), A), upper=True)
-        ret = torch.cholesky_solve(torch.matmul(A.transpose(-1, -2), b), u, upper=True)
-        ctx.save_for_backward(u, ret, A, b)
+        L = torch.linalg.cholesky(torch.matmul(A.transpose(-1, -2), A))
+        ret = torch.cholesky_solve(torch.matmul(A.transpose(-1, -2), b), L, upper=False)
+        ctx.save_for_backward(L, ret, A, b)
         return ret
 
     @staticmethod
     def backward(ctx, grad_output):
         # https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/linalg_grad.py#L223
         chol, x, a, b = ctx.saved_tensors
-        z = torch.cholesky_solve(grad_output, chol, upper=True)
+        z = torch.cholesky_solve(grad_output, chol, upper=False)
         xzt = torch.matmul(x, z.transpose(-1, -2))
         zx_sym = xzt + xzt.transpose(-1, -2)
         grad_A = - torch.matmul(a, zx_sym) + torch.matmul(b, z.transpose(-1, -2))
@@ -197,11 +184,13 @@ def ls(X, Y, weights=None):
 def ls_cov(X, Y, weights=None, make_symmetric=True, sigreg=1e-4):
     X, Y = _apply_weights(X, Y, weights)
 
-    pinvXX = X.pinverse()
-    params = (pinvXX @ Y).t()
+    # Solve least squares via lstsq (more stable than pinverse)
+    result = torch.linalg.lstsq(X, Y)
+    params = result.solution.t()
 
     # estimate covariance according to: http://users.stat.umn.edu/~helwig/notes/mvlr-Notes.pdf (see up to slide 66)
     # hat/projection matrix - Yhat = H*Y
+    pinvXX = X.pinverse()
     H = X @ pinvXX
 
     N = X.shape[0]
@@ -231,8 +220,9 @@ def ls_cov(X, Y, weights=None, make_symmetric=True, sigreg=1e-4):
         XXXX = XXXX_sym
         error_covariance = error_covariance_sym
 
-    # TODO might be able to use cholesky decomp here since XXXX > 0
-    covariance = kronecker_product(error_covariance, XXXX.inverse())
+    # Use solve instead of explicit inverse: solve(A, I) = A^{-1}, more numerically stable
+    XXXX_inv = torch.linalg.solve(XXXX, torch.eye(XXXX.shape[0], dtype=XXXX.dtype, device=XXXX.device)).contiguous()
+    covariance = kronecker_product(error_covariance, XXXX_inv)
 
     return params, covariance
 

src/arm_pytorch_utilities/math_utils.py

@@ -11,9 +11,7 @@ def clip(a, min_val, max_val):
 
 def replace_nan_and_inf(a, replacement=0):
     """Replaces nan,inf,-inf values with replacement value in place"""
-    a[torch.isnan(a)] = replacement
-    a[a == float('inf')] = replacement
-    a[a == -float('inf')] = replacement
+    torch.nan_to_num(a, nan=replacement, posinf=replacement, neginf=replacement, out=a)
     return a
 
 
@@ -66,8 +64,8 @@ def angle_between_stable(u: torch.tensor, v: torch.tensor):
     dim = -1
     u_norm = u.norm(dim=dim, keepdim=True)
     v_norm = v.norm(dim=dim, keepdim=True)
-    uv = u.unsqueeze(1).repeat(1, v.shape[0], 1) * v_norm
-    vu = v.unsqueeze(0).repeat(u.shape[0], 1, 1) * u_norm.unsqueeze(1)
+    uv = u.unsqueeze(1) * v_norm.transpose(-2, -1).unsqueeze(-1)
+    vu = v.unsqueeze(0) * u_norm.unsqueeze(1)
     num = (uv - vu).norm(dim=dim)
     den = (uv + vu).norm(dim=dim)
     return 2 * torch.atan2(num, den)
@@ -104,9 +102,7 @@ def angular_diff(a, b):
 def angular_diff_batch(a, b):
     """Angle difference from b to a (a - b)"""
     d = a - b
-    d[d > math.pi] -= 2 * math.pi
-    d[d < -math.pi] += 2 * math.pi
-    return d
+    return ((d + math.pi) % (2 * math.pi)) - math.pi
 
 
 def angle_normalize(a):

src/arm_pytorch_utilities/optim.py

@@ -106,12 +106,12 @@ def step(self, closure=None):
             for group in self.optimizer.param_groups:
                 for p in group['params']:
                     param_state = self.state[p]
-                    p.data.mul_(self.la_alpha).add_(1.0 - self.la_alpha, param_state['cached_params'])  # crucial line
+                    p.data.mul_(self.la_alpha).add_(param_state['cached_params'], alpha=1.0 - self.la_alpha)  # crucial line
                     param_state['cached_params'].copy_(p.data)
                     if self.pullback_momentum == "pullback":
                         internal_momentum = self.optimizer.state[p]["momentum_buffer"]
                         self.optimizer.state[p]["momentum_buffer"] = internal_momentum.mul_(self.la_alpha).add_(
-                            1.0 - self.la_alpha, param_state["cached_mom"])
+                            param_state["cached_mom"], alpha=1.0 - self.la_alpha)
                         param_state["cached_mom"] = self.optimizer.state[p]["momentum_buffer"]
                     elif self.pullback_momentum == "reset":
                         self.optimizer.state[p]["momentum_buffer"] = torch.zeros_like(p.data)

src/arm_pytorch_utilities/preprocess.py

@@ -311,13 +311,15 @@ def __init__(self, **kwargs):
         super().__init__(**kwargs)
         self._m = None
         self._s = None
+        self._inv_s = None
 
     def fit(self, X):
         self._m = X.mean(0, keepdim=True)
         self._s = X.std(0, unbiased=False, keepdim=True)
+        self._inv_s = 1.0 / self._s
 
     def transform(self, X):
-        return (X - self._m) / self._s
+        return (X - self._m) * self._inv_s
 
     def inverse_transform(self, X):
         return (X * self._s) + self._m