scm2cpp-lasso 0.13.2__py3-none-win_amd64.whl

scm2cpp_lasso-0.13.2.dist-info/METADATA

@@ -0,0 +1,920 @@
+Metadata-Version: 2.4
+Name: scm2cpp-lasso
+Version: 0.13.2
+Summary: scikit-learn estimators (lasso, elastic net, ridge, ARD, SVM, SVR, logistic regression, GLMs, Huber, Theil-Sen, F-score selection, RFE, NMF, Gaussian mixture, FastICA, PCA, graphical lasso, RANSAC, SGD, OMP, LARS) translated from Scheme to C++ by scm2cpp; GPU batch solver and CUDA kernels included
+Author: Hirotaka Niitsuma
+License: MIT
+Project-URL: Homepage, https://github.com/niitsuma/scm2cpp
+Project-URL: Source, https://github.com/niitsuma/scm2cpp/tree/master/python/scm2cpp-lasso
+Keywords: lasso,elastic-net,ridge,svm,logistic-regression,coordinate-descent,cuda,compiler
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: C++
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Classifier: Intended Audience :: Science/Research
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+Requires-Dist: numpy>=1.17
+
+# scm2cpp-lasso
+
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/niitsuma/scm2cpp/blob/master/notebooks/compare-sklearn.ipynb)
+
+*[Japanese (README.ja.md)](README.ja.md)*
+
+Lasso by coordinate descent over a Gram matrix, with an optional GPU
+path.  The solver is Scheme, translated to C++ by
+[scm2cpp](https://github.com/niitsuma/scm2cpp).
+
+```console
+$ pip install scm2cpp-lasso
+```
+
+On Linux x86_64, macOS arm64 (macOS 14 or later) and Windows AMD64
+this installs a wheel: compiled, nothing to build (with OpenBLAS
+inside on Linux and macOS; Windows has the loop variants).  Elsewhere pip builds the sdist,
+which needs a C++17 compiler and nothing else.  The wheels carry no
+GPU code (a wheel cannot know the device); for the GPU builds, install
+from the sdist on a machine with `nvcc` on the path:
+
+```console
+$ pip install --no-binary scm2cpp-lasso scm2cpp-lasso
+```
+
+The batched GPU solver, the cuBLAS variants and the `--cuda`
+estimators are then built as well; without `nvcc` the package installs
+and works exactly the same, minus those.
+
+Three of a hundred candidate columns carry the signal; the solver is
+asked which, and answers with them and nothing else:
+
+```python
+import numpy as np
+from scm2cpp_lasso import CovLasso
+
+rng = np.random.default_rng(0)
+X = rng.standard_normal((500, 100))          # 500 rows, 100 candidates
+beta = np.zeros(100)
+beta[[7, 23, 61]] = [2.0, -1.5, 0.8]         # only three of them matter
+y = X @ beta + 0.1 * rng.standard_normal(500)
+
+model = CovLasso(X, y)                       # the Gram matrix, built once
+path = model.fit_path(model.lambda_grid())   # 100 lambdas, warm-started
+
+fit = path[60]
+for j in np.flatnonzero(np.abs(fit) > 1e-6):
+    print(f"  x{j:<3} {fit[j]:+.3f}")
+```
+
+```console
+  x7   +1.965
+  x23  -1.474
+  x61  +0.770
+```
+
+The other 97 are exactly zero -- that is what the penalty is for -- and
+the coefficients are the planted ones, shrunk toward zero by it.
+
+The rest of the interface:
+
+```python
+from scm2cpp_lasso import cuda_available
+
+lambdas = model.lambda_grid(num=100)        # from lambda_max down
+path = model.fit_path(lambdas)              # warm-started, sequential
+grid = model.fit_path_batch(lambdas)        # every lambda from zero
+one = model.fit(lambdas[60])                # a single lambda, if you must
+print("GPU:", cuda_available())
+```
+
+`fit` is the one-element special case of `fit_path` -- the Gram matrix
+was built in the constructor, so a whole warm-started path costs barely
+more than one penalty, which is why the path is the primary interface.
+
+The objective is scikit-learn's, with `fit_intercept=False`:
+
+    (1 / 2 nobs) ||y - X b||^2 + lam ||b||_1
+
+Penalties are compared against correlations on the features' own
+scale, so the useful range of `lam` depends on the data.
+`lambda_max()` is the smallest penalty that leaves every coefficient
+at zero, and `lambda_grid()` walks down from it -- the construction
+scikit-learn uses.  If your columns differ wildly in scale, standardize
+them first; this solver does not do it for you.
+
+## Estimators at a glance
+
+Every estimator, the scikit-learn estimator it reproduces, the Scheme
+kernel under `examples/kernel-only/` it is translated from, the builds
+it runs on, and whether Intel's sklearnex accelerates the scikit-learn
+side. Builds: *loops* is the plain translation, *BLAS* / *cuBLAS* the
+products handed to (cu)BLAS, *--cuda* the translator's GPU kernels
+(`gpu_cuda/`), *GPU batch* the hand-written CUDA batch solver; the
+package picks the fastest one present. None of the classes derives
+from scikit-learn's `BaseEstimator` yet, so `clone`, `GridSearchCV`
+and `check_estimator` do not apply to them.
+
+| class | scikit-learn | Scheme kernel | builds | sklearnex accelerates |
+|---|---|---|---|---|
+| `CovLasso` | Lasso, ElasticNet, lasso_path(precompute=True) | `lasso-cov.scm` | CPU; GPU batch (fit_path_batch, bootstrap) | no |
+| `CovLassoCV` | LassoCV, ElasticNetCV | `lasso-cov.scm` | CPU; GPU batch | no |
+| `DerivedLasso` | lasso_path, from the design | `lasso-design.scm (--derive)` | loops / BLAS / cuBLAS | no |
+| `CovMultiTaskLasso` | MultiTaskLasso, MultiTaskElasticNet | `lasso-cov.scm` | CPU | no |
+| `CovMultiTaskLassoCV` | MultiTaskLassoCV, MultiTaskElasticNetCV | `lasso-cov.scm` | CPU; GPU batch | no |
+| `CovGroupLasso` | (none; skglm GroupLasso) | `lasso-cov.scm` | CPU | - |
+| `CovLogistic` | LogisticRegression(penalty='l1'), by majorization | `lasso-cov.scm` | CPU | no |
+| `CovRidge` | Ridge(fit_intercept=False), a whole alpha path | `lasso-cov.scm` | CPU | Ridge |
+| `CovBayesianRidge` | BayesianRidge | `lasso-cov.scm` | CPU | no |
+| `CovARDRegression` | ARDRegression (n >= p) | `ard.scm` | loops / BLAS | no |
+| `GramLars` | Lars | `lars.scm` | CPU | no |
+| `GramLassoLars` | LassoLars | `lars.scm` | CPU | no |
+| `GramLassoLarsIC` | LassoLarsIC | `lars.scm` | CPU; device Gram | no |
+| `GramLarsCV` | LarsCV | `lars.scm` | CPU | no |
+| `GramLassoLarsCV` | LassoLarsCV | `lars.scm` | CPU; device Gram | no |
+| `GramOMP` | OrthogonalMatchingPursuit | `omp.scm` | loops / BLAS | no |
+| `GramOMPCV` | OrthogonalMatchingPursuitCV | `omp.scm` | loops / BLAS | no |
+| `LBFGSLogisticRegression` | LogisticRegression(solver='lbfgs') | `logistic-lbfgs.scm, lbfgs.scm` | loops / BLAS / --cuda | LogisticRegression |
+| `LBFGSLogisticRegressionCV` | LogisticRegressionCV | `logistic-lbfgs.scm, lbfgs.scm` | loops / BLAS / --cuda | no |
+| `LBFGSHuberRegressor` | HuberRegressor | `huber.scm, lbfgs.scm` | loops / BLAS / --cuda | no |
+| `LBFGSPoissonRegressor` | PoissonRegressor | `glm.scm, lbfgs.scm` | loops / BLAS / --cuda | no |
+| `LBFGSGammaRegressor` | GammaRegressor | `glm.scm, lbfgs.scm` | loops / BLAS / --cuda | no |
+| `LBFGSTweedieRegressor` | TweedieRegressor | `glm.scm, lbfgs.scm` | loops / BLAS / --cuda | no |
+| `LinearSVM` | LinearSVC(loss='hinge') | `svm-liblinear.scm` | CPU | no |
+| `L1LinearSVM` | LinearSVC(penalty='l1', dual=False) | `svm-l1.scm` | CPU | no |
+| `KernelSVM` | SVC | `svm-smo.scm` | CPU | SVC |
+| `LinearSVR` | LinearSVR | `svr-liblinear.scm` | CPU | no |
+| `KernelSVR` | SVR | `svm-smo.scm` | CPU | SVR |
+| `L1LogisticRegression` | LogisticRegression(penalty='l1', solver='liblinear') | `logistic-l1.scm` | CPU | no |
+| `SGDLinearClassifier` | SGDClassifier(shuffle=False) | `sgd.scm` | CPU | no |
+| `SGDLinearRegressor` | SGDRegressor(shuffle=False) | `sgd.scm` | CPU | no |
+| `SGDPerceptron` | Perceptron(shuffle=False) | `sgd.scm` | CPU | no |
+| `LinearRANSAC` | RANSACRegressor (given subsets) | `ransac.scm` | CPU | no |
+| `WeiszfeldTheilSen` | TheilSenRegressor | `theilsen.scm` | CPU | no |
+| `CDNMF` | NMF(solver='cd') | `nmf-cd.scm (--derive)` | loops / BLAS / cuBLAS | no |
+| `CDGraphicalLasso` | GraphicalLasso | `graphical-lasso.scm` | loops / BLAS | no |
+| `EMGaussianMixture` | GaussianMixture(covariance_type='full') | `gmm.scm` | loops / BLAS / --cuda | no |
+| `ParallelFastICA` | FastICA | `fastica.scm` | loops / BLAS / --cuda | no |
+| `CovPCA` | PCA(svd_solver='covariance_eigh') | `pca.scm` | loops / BLAS / cuBLAS | PCA |
+| `f_regression, f_classif` | f_regression, f_classif | `fscore.scm` | loops / BLAS / cuBLAS | no |
+| `SelectKBest, SelectPercentile` | SelectKBest, SelectPercentile | `fscore.scm` | loops / BLAS / cuBLAS | no |
+| `GramRFE` | RFE(Ridge), RFE(LinearRegression) | `rfe.scm` | loops / BLAS / cuBLAS | inside: Ridge |
+
+## Two kernels: `CovLasso` and `DerivedLasso`
+
+Everything above, and the cross-validation, bootstrap and GPU methods
+below, runs on one kernel: numpy forms the Gram matrix, and a
+hand-written coordinate-descent sweep over it
+(`examples/kernel-only/lasso-cov.scm` in the repository, translated to
+C++) does the rest.  That is the usable one.
+
+`DerivedLasso` is the other kernel, and it is there to show the
+translator's work rather than to be the better estimator.  Its source
+is `examples/kernel-only/lasso-design.scm`: the residual-form
+coordinate descent -- the algorithm inside scikit-learn's
+`Lasso(precompute=False)` -- written as flat loops over X exactly as
+numpy holds it (n rows of p, `x[i*p+j]`), the residual copied from y
+and the column norms taken inside, with no Gram matrix anywhere in
+it.  The translator's `--derive` turns
+that into the Gram form: it raises the loops to an array algebra,
+notices that the residual is only ever updated by a scaled column and
+read in dot products against columns, replaces it by the memo `c =
+X'r` kept current through a Gram matrix it hoists in front of the
+penalty path, folds the column norms into that Gram's diagonal, folds
+the three products into one `dsyrk` and two `dgemv` calls (`--blas`), and puts the whole thing under
+`(if (< p n) ..)` with the sweep as written in the else branch --
+scikit-learn's `precompute='auto'` rule, derived rather than written.
+The C++ of that derivation, as the translator emitted it, is what
+`DerivedLasso` calls; numpy prepares nothing -- X and y go in as they
+are, no transposed copy, no norms.
+
+```python
+from scm2cpp_lasso import DerivedLasso, derived_backend
+
+model = DerivedLasso(X, y)                   # X and y; the kernel builds its own Gram
+path = model.fit_path(model.lambda_grid())   # one call, the whole warm-started path
+print(derived_backend())                     # 'cublas', 'blas' or 'loops': what this machine built
+gpu = DerivedLasso(X, y, backend="cublas")   # insist on one
+```
+
+A lambda is done when no coefficient moved by more than `tol`
+(absolute, default 1e-8) in a sweep, or at `max_sweeps`; `tol=0.0`
+asks for the exact fixed point.  Three translations of the one
+derivation are committed and `pip` builds what the machine allows:
+`--derive --cublas` where the CUDA toolkit is (the three products on
+the device through cuBLAS, X uploaded once per `fit_path` call --
+taken by default when a device answers), `--derive --blas` where
+`cblas.h` and an OpenBLAS are, and `--derive` alone otherwise, the
+same program with its Gram matrix as a scalar loop nest, correct but
+slow on large designs.  `check-derived.py` in the package directory runs
+the two kernels and scikit-learn on the same path and prints the
+differences and the times; on AR(1) designs (rho 0.9), 20 penalties
+from lambda_max to 0.01 lambda_max, one core, best of three (load
+about 3), whole calls with the Gram matrix wherever it is built:
+
+| n x p                                   | 1,800 x 200 | 5,000 x 1000 | 100,000 x 200 | 100,000 x 500 |
+|-----------------------------------------|-------------|--------------|---------------|---------------|
+| `DerivedLasso`, cuBLAS (upload included) | 0.003 s     | 0.024 s      | 0.052 s       | 0.11 s        |
+| `DerivedLasso`, CBLAS                   | 0.003 s     | 0.076 s      | 0.12 s        | 0.44 s        |
+| `CovLasso(X, y).fit_path`, tol 1e-8     | 0.004 s     | 0.080 s      | 0.11 s        | 0.42 s        |
+| sklearn `lasso_path(precompute=True)`, tol 1e-4 | 0.005 s     | 0.10 s       | 0.24 s        | 0.76 s        |
+
+The CBLAS run is the same `dsyrk` and the same sweeps as `CovLasso`'s
+numpy Gram and path, and times the same; the cuBLAS run is the same
+program with the three products on the device, the upload of X
+counted, and is 4.2x, 4.5x and 6.7x ahead of scikit-learn at the
+three larger shapes -- the figures of the repository's equal-work
+table, now reached from the package with nothing in front of the
+kernel: the constructor keeps X and y and the kernel takes the
+residual and the column norms itself (the norms folded into the Gram
+diagonal by the derivation).  Threads change the picture less than one might expect: scikit-learn's
+sweeps are single-threaded Cython and only its Gram product scales,
+so with 20 BLAS threads `lasso_path` takes 0.051 / 0.227 / 0.450 s on
+the three larger shapes, against 0.023 / 0.050 / 0.113 s for the
+cuBLAS build (2.2x, 4.5x, 4.0x) and 0.025 / 0.036 / 0.123 s for the
+CBLAS build (2.0x, 6.3x, 3.7x), whose `dsyrk` scales with the
+threads; the GPU figures do not move.  Against a reference driven to
+tol=1e-12, `DerivedLasso` lands within 5e-8 on every backend (its
+tolerance is absolute, on one sweep's largest move), `CovLasso`
+within 2e-10 (it tests every 20 sweeps and overshoots), and
+scikit-learn at its default within 1e-3.
+
+### Deriving it yourself
+
+Everything `DerivedLasso` runs is reproducible from the Scheme with
+the translator; this is what `regenerate.sh` does for the package,
+spelled out for one kernel (Racket 8 and a C++17 compiler; a CBLAS
+for `--blas`, the CUDA toolkit for `--cublas`):
+
+```console
+$ git clone https://github.com/niitsuma/scm2cpp && cd scm2cpp
+$ raco link --user vendor/rkanren                 # once per machine
+$ mkdir /tmp/k && cp examples/kernel-only/lasso-design.scm examples/kernel-only/soft-threshold.scm /tmp/k/
+$ racket scm2cpp-file.scm -t scm2c.typ --derive --cublas -M -S /tmp/k/lasso-design.scm
+derive: lasso-design: raise differencing merge matmul guard
+```
+
+`derive:` on stderr names what fired: the loops raised to the array
+algebra, the residual differenced into the memo, the column norms
+merged into the Gram diagonal, the products folded to contractions,
+the `(if (< p n) ..)` guard.  The run leaves, beside the source,
+`lasso-design.hpp` (the derived kernel as C++: one `blas_syrk`, two
+`blas_gemv`, the sweep), `lasso-design_capi.cpp` and
+`lasso_design.py` (the `-M` C ABI and its ctypes loader -- what the
+package's `_loader_derived_*.py` are), and `lasso-design.expanded.scm`
+(`-S`: the derived program as Scheme, before C++ -- read this to see
+what the derivation did).  `--blas` instead of `--cublas` gives the
+CBLAS variant, neither the loop variant.  To call it from numpy
+without the package:
+
+```console
+$ g++ -O2 -std=c++17 -shared -fPIC -I. -I$CUDA_HOME/include -o /tmp/k/liblasso-design.so \
+      /tmp/k/lasso-design_capi.cpp -L$CUDA_HOME/lib64 -lcublas -lcudart   # -lopenblas for --blas
+$ cd /tmp/k && python3 -c 'import lasso_design'   # lasso_design.lasso_design(x, y, resid, xnorm, beta, lams, betas, iters, tol, n, p, nlam)
+```
+
+(`-M` names the Python module after the file with `-` read as `_`,
+and the loader looks for `liblasso-design.so` beside it.)
+`examples/kernel-only/README.md` walks through the same `-M` flow on
+the row-major kernel, and `python/PACKAGING.md` through turning such
+a kernel into a pip package.
+
+## Against the other lassos pip can install
+
+`bench/lasso-compare.py` in the repository runs the same
+cross-validation-shaped grid -- 4096 lambdas, each solved from zero,
+p=200, n=1800 -- against every lasso a `pip install` produces:
+scikit-learn 1.9.1, celer 0.7.4, skglm 0.5, and RAPIDS cuML 26.8.
+Every solver runs at its own default tolerance (ours at the tighter
+1e-8), each gets one untimed warm-up fit, cuML's data is on the device
+before the clock starts, and the last column is each solver's
+objective minus ours at the smallest lambda -- a fast row with a loose
+answer would show as one:
+
+| solver                                 | time    | objective gap |
+|----------------------------------------|---------|---------------|
+| scm2cpp-lasso, 1 CPU core (tol 1e-8)   | 0.91 s  | 0             |
+| scm2cpp-lasso, GPU, one block/lambda   | 0.044 s | 0             |
+| sklearn `Lasso.fit` per lambda         | 9.6 s   | +1.6e-09      |
+| celer per lambda                       | 13.5 s  | 0             |
+| skglm per lambda                       | 14.5 s  | +2.8e-17      |
+| cuML per lambda, GPU                   | 53.6 s  | +9.1e-07      |
+
+(One sitting on an otherwise idle machine, best of three runs; cuML
+26.08 on CUDA 12.9.)
+
+celer and skglm are built for very large sparse designs, where their
+screening rules dominate; at this size they pay their setup per fit.
+cuML parallelises *within* one fit, which wins when a fit is large; at
+this size its launch overhead dominates, while our GPU row
+parallelises *across* lambdas -- one block of CUDA threads per
+penalty, the axis a cross-validation grid actually offers.
+
+`CovLassoCV` is scikit-learn's `LassoCV` over this machinery -- same
+grid construction, same contiguous folds, same minimum-mean-MSE choice
+-- with two structural savings.  A fold's training Gram is a
+*subtraction* (the Gram is additive over rows, so G - Xf'Xf costs one
+fold's work where rebuilding costs four folds'), and once the Grams
+exist nothing downstream ever touches the n rows again -- which is why
+the gap grows with n.  With a GPU the whole grid is one launch: one
+*block* of threads per (fold, alpha) cell, the fold's Gram read in
+place (cv Gram matrices on the device, not cv x 100), thread 0 of the
+block doing the coordinate step and the block spreading the O(p)
+update of the correlations.  cv x 100 = 500 cells are too few
+problems to fill a device one thread each, and a block per problem is
+what turns that count into enough parallel work; `bench/cv-grid-designs.py`
+times the alternatives (one thread per cell, warm runs of several
+alphas per thread or block) side by side, and the block per cell wins
+at every size tried.  The GPU is the default when a device answers
+(`force_cpu` takes the warm CPU path).  Measured at cv=5, 100 alphas,
+one CPU core, sklearn 1.9.1, best of three runs, whole estimator
+including the Grams:
+
+| n       | p    | CPU    | CUDA   | sklearn `LassoCV` |
+|---------|------|--------|--------|-------------------|
+| 1,800   | 200  | 0.21 s | 0.02 s | 0.10 s            |
+| 5,000   | 1000 | 1.48 s | 0.25 s | 0.90 s            |
+| 100,000 | 200  | 0.49 s | 0.34 s | 1.72 s            |
+| 100,000 | 500  | 1.32 s | 1.09 s | 4.82 s            |
+
+(Taken on an otherwise idle machine, load 2-3 on 20 threads.  sklearn's
+`LassoCV` at its default `precompute='auto'` takes the Gram route at
+every size here, n > p, rebuilding each fold's Gram from that fold's
+rows where we subtract it from the whole.)  The grid
+launch itself takes 9 ms at n=1,800 and 64 ms at n=5,000, p=1000; at
+n=100,000 the CUDA column is nearly all Gram products, which the CPU
+path pays as well.  Large n is where the machinery pays either way:
+nothing downstream of the Grams touches the 100,000 rows again.  At
+n=5,000, p=1000 sklearn's raw-X descent beats our Gram-heavy CPU path,
+and the GPU makes up the difference three and a half times over.  CPU and CUDA
+choose the same alpha everywhere and agree to 1e-14; against sklearn
+the pick is exact at n=5,000 and one grid step apart on a near-tie
+elsewhere (mean-MSE relative gap at most 2.6e-4), and sklearn at
+tol=1e-10 picks our alpha at every size.
+
+The elastic net cross-validates the same way: `CovLassoCV(l1_ratio=0.5)`
+against `ElasticNetCV(l1_ratio=0.5)` on the same folds and alpha grid
+(`bench/lasso-compare.py --cv-only --l1-ratio 0.5`, same conditions,
+load under 1):
+
+| n       | p    | CPU    | CUDA   | sklearn `ElasticNetCV` |
+|---------|------|--------|--------|------------------------|
+| 1,800   | 200  | 0.35 s | 0.02 s | 0.11 s                 |
+| 5,000   | 1000 | 1.55 s | 0.26 s | 0.98 s                 |
+| 100,000 | 200  | 0.52 s | 0.35 s | 1.76 s                 |
+| 100,000 | 500  | 1.38 s | 1.15 s | 4.97 s                 |
+
+The picture is the lasso's: on the CPU, 3.4-3.6x at 100,000 rows and
+slower than sklearn at the two small shapes; with CUDA, 3.8-5.5x
+everywhere.  The chosen alpha is sklearn's at all four sizes, the
+coefficients within 8e-6 (sklearn's default tolerance).
+
+`CovMultiTaskLasso` and `CovMultiTaskLassoCV` are the multi-task
+family over the same machinery: scikit-learn's `MultiTaskLasso`,
+`MultiTaskElasticNet` (through `l1_ratio`), `MultiTaskLassoCV` and
+`MultiTaskElasticNetCV`.  The penalty ties each feature's row of W
+together across tasks, so the coordinate update becomes a block soft
+threshold on the row's L2 norm; C = X'Y - GW is maintained per task
+exactly as the single-task C is, and nothing downstream of the Grams
+touches the n rows -- the kernel is the same translated Scheme
+(`mt-descend`).  Folds are independent and both the descent (ctypes)
+and the products (BLAS) release the GIL, so `n_jobs` threads scale the
+CV across folds; the default is sequential, as scikit-learn's
+`n_jobs=None` is, and the benchmark pins BLAS to one thread on both
+sides either way.  With a GPU the CV grid is one launch exactly as
+`CovLassoCV`'s is, a block per (fold, alpha) cell with the p x T
+coefficient block in shared memory (`force_cpu` opts out).  Same
+protocol as above, 8 tasks; single fits at alpha = 0.1 lambda_max;
+the CV pairs share the 100-alpha grid:
+
+| estimator             | n       | p   | ours   | ours `n_jobs=5` | ours CUDA | sklearn |
+|-----------------------|---------|-----|--------|-----------------|-----------|---------|
+| MultiTaskLasso        | 100,000 | 200 | 0.14 s | --              | --        | 0.76 s  |
+| MultiTaskLasso        | 100,000 | 500 | 0.53 s | --              | --        | 1.7 s   |
+| MultiTaskLassoCV      | 1,800   | 200 | 1.4 s  | 0.51 s          | 0.07 s    | 0.42 s  |
+| MultiTaskLassoCV      | 100,000 | 200 | 1.6 s  | 0.64 s          | 1.3 s     | 74 s    |
+| MultiTaskElasticNetCV | 1,800   | 200 | 1.4 s  | 0.49 s          | 0.08 s    | 0.75 s  |
+| MultiTaskElasticNetCV | 100,000 | 200 | 1.6 s  | 0.68 s          | 1.3 s     | 162 s   |
+
+(`MultiTaskElasticNet` single fits time the same as `MultiTaskLasso`'s;
+every column was taken in one sitting on the idle machine of the
+`CovLassoCV` table above, sklearn 1.9.1.  scikit-learn has no Gram
+route for its multi-task estimators -- its path code sets
+`precompute=False` for 2-D targets, "No Gram variant of multi-task
+exists right now" -- so its column is raw-X descent at every n, which
+is why it grows with n where `LassoCV`'s did not.)  At n=1,800 our sequential CV
+is slower than sklearn's: the default tol=1e-8 runs more sweeps than
+sklearn's 1e-4 dual-gap stop, and at that size there are no rows to
+save -- and the GPU grid takes the whole thing in a tenth of a second.
+At n=100,000 the Gram route is 46x and 100x ahead, essentially
+unchanged from n=1,800 in absolute time; the CUDA column there is
+the Gram products, the same ones the CPU path forms.  Coefficients agree with scikit-learn's to 1e-15 at tight
+tolerance, and the CV pair picks the same alpha with coefficients to
+1e-13.
+
+## Which method
+
+`fit_path` walks a single path, each lambda starting from the previous
+solution -- the descent is exactly resumable, so this costs almost
+nothing per lambda after the first.  `fit_path_batch` solves every
+lambda from zero, which is what a cross-validation grid needs, where
+the folds differ and warm starting across lambdas is not on offer; the
+problems are independent, so they go to the GPU together.
+
+Over a 400-lambda path at p=200 and 1800 rows, on an RTX 4090 and an
+i9-10900X:
+
+| call                                  | time    |
+|---------------------------------------|---------|
+| `fit_path` (warm, sequential)         | 0.091 s |
+| `fit_path_batch` (GPU)                | 0.047 s |
+| `fit_path_batch(force_cpu=True)`      | 0.178 s |
+
+GPU and CPU agree to 2e-14, and the objectives are within 3e-17 of
+scikit-learn's `lasso_path` on the same grid.
+
+## Elastic net and ridge
+
+The same Gram matrix serves two more estimators.  `fit_path` and
+`fit_path_batch` take `l1_ratio` (scikit-learn's mixing parameter):
+the L2 share of the penalty enters only the update's denominator, so
+the elastic net runs on the identical machinery, GPU path included,
+and `l1_ratio=1` is bit-for-bit the lasso.
+
+`CovRidge` is the closed form: one symmetric eigendecomposition,
+then every alpha costs O(p^2), so thousands of alphas cost what one
+does.  Its objective matches scikit-learn's `Ridge` with
+`fit_intercept=False` (which, unlike the lasso, is not scaled by the
+number of rows), and agrees with it to machine precision.
+
+```python
+path = model.fit_path(lambdas, l1_ratio=0.5)   # elastic net
+ridge = CovRidge(X, y)
+betas = ridge.fit_path(ridge.alpha_grid())     # a whole ridge path
+```
+
+Measured (`bench/ridge-enet-compare.py` in the repository: AR(1)
+designs, 20 penalties, whole calls with the Gram matrix wherever one
+is built, one core, best of three):
+
+| n x p                                        | 1,800 x 200 | 5,000 x 1000 | 100,000 x 200 | 100,000 x 500 |
+|----------------------------------------------|-------------|--------------|---------------|---------------|
+| elastic net, `CovLasso.fit_path(l1_ratio=0.5)`, tol 1e-8 | 0.007 s | 0.092 s | 0.11 s      | 0.40 s        |
+| elastic net, sklearn `enet_path(precompute=True)`, tol 1e-4 | 0.005 s | 0.099 s | 0.23 s   | 0.74 s        |
+| ridge, `CovRidge.fit_path(alphas)`           | 0.004 s     | 0.21 s       | 0.10 s        | 0.41 s        |
+| ridge, sklearn `Ridge(alpha).fit` per alpha  | 0.060 s     | 2.05 s       | 3.37 s        | 11.2 s        |
+| ridge, sklearn `RidgeCV(alphas)` (its LOO path) | 0.093 s  | 3.40 s       | 5.38 s        | 19.0 s        |
+
+The elastic net is the same coordinate descent on both sides, so the
+two rows differ by the Gram and the warm path, not the sweeps: level
+at small n, 1.8-2.2x at 100,000 rows, and ours lands within 2e-10 of
+a tight reference where scikit-learn's default tolerance lands within
+6e-4.  The ridge rows compare a whole path against what a user
+writes -- one `Ridge.fit` per alpha, each forming X'X and solving
+afresh, O(np^2 + p^3) every time -- so the gap is the path's: 10-33x,
+agreeing to 1e-14.  `RidgeCV` is given for scale (it picks one alpha
+by leave-one-out rather than returning the path).  Neither row runs
+on the GPU: the ridge is one eigendecomposition and the elastic net's
+sweeps are the lasso's; `fit_path_batch(l1_ratio=..)` is the GPU
+grid for the elastic net.
+
+## Bayesian ridge
+
+`CovBayesianRidge` is scikit-learn's `BayesianRidge` -- the same model,
+the same defaults, the same iteration (MacKay's updates of the noise
+precision `alpha_` and the weight precision `lambda_`, stopping when
+the posterior mean moves by less than `tol`) -- with the linear
+algebra moved into the eigenbasis of the Gram matrix.  scikit-learn
+takes the SVD of X once and multiplies by X again in every iteration
+for the squared error.  Here X enters once, through X'X (X X' when
+n <= p): one eigendecomposition, after which the posterior mean, its
+norm and the squared error are sums over the eigenvalues, O(p) per
+iteration.  `sigma_` (the posterior covariance) and
+`predict(return_std=True)` are there as in scikit-learn.
+
+```python
+from scm2cpp_lasso import CovBayesianRidge
+m = CovBayesianRidge().fit(X, y)          # fit_intercept=True, as sklearn
+mean, std = m.predict(X_new, return_std=True)
+```
+
+Measured (`bench/bayesian-ridge-compare.py` in the repository:
+Gaussian designs, a sparse truth plus noise, whole fits including
+`sigma_`, one core, best of three, sklearn 1.9.1):
+
+| n x p          | `CovBayesianRidge` | sklearn `BayesianRidge` | ratio | iterations (both) | max coefficient difference |
+|----------------|--------------------|-------------------------|-------|-------------------|----------------------------|
+| 1,800 x 200    | 0.005 s            | 0.026 s                 | 5.2x  | 3                 | 3.6e-15                    |
+| 5,000 x 1000   | 0.257 s            | 1.231 s                 | 4.8x  | 4                 | 6.7e-15                    |
+| 100,000 x 200  | 0.126 s            | 3.495 s                 | 27.8x | 3                 | 2.8e-14                    |
+| 100,000 x 500  | 0.475 s            | 8.909 s                 | 18.8x | 3                 | 1.6e-14                    |
+| 500 x 5000     | 0.662 s            | 8.689 s                 | 13.1x | 71                | 1.7e-15                    |
+
+The iteration counts are identical and the coefficients agree to
+rounding, so the gap is the linear algebra alone: X'X by one BLAS
+product against a full SVD of the tall X, and no pass over X inside
+the loop.  At 500 x 5000 scikit-learn takes the full 5000 x 5000 SVD
+for `sigma_`; here `sigma_` comes from the 500 x 500 kernel side by
+the Woodbury identity.  This estimator is numpy, not translated code:
+the eigendecomposition is LAPACK's either way.
+
+## Classifiers and SVRs: `LinearSVM`, `KernelSVM`, `LinearSVR`, `KernelSVR`, `L1LogisticRegression`
+
+Three classifiers and two regressors, each the algorithm scikit-learn runs, written out
+in the Scheme subset and translated plainly (no derivation; the C++
+is committed under `_generated/classifiers*`):
+
+| class | algorithm | scikit-learn counterpart |
+|---|---|---|
+| `LinearSVM` | liblinear's dual coordinate descent for the hinge loss (random order, shrinking) | `LinearSVC(loss="hinge", dual=True, fit_intercept=False)` |
+| `KernelSVM` | libsvm's SMO (WSS3, shrinking, the bias) over the whole kernel matrix | `SVC` (`kernel="rbf"`, `"linear"` or `"precomputed"`) |
+| `L1LogisticRegression` | liblinear's newGLMNET (Newton over coordinate descent, line search) | `LogisticRegression(l1_ratio=1, solver="liblinear", fit_intercept=False)` |
+| `LinearSVR` | liblinear's dual coordinate descent for the epsilon-insensitive loss (random order, shrinking) | `LinearSVR(loss="epsilon_insensitive", dual=True, fit_intercept=False)` |
+| `KernelSVR` | the same SMO as `KernelSVM`, set up as libsvm's epsilon-SVR (2n variables) | `SVR` (`kernel="rbf"`, `"linear"` or `"precomputed"`) |
+
+```python
+from scm2cpp_lasso import LinearSVM, KernelSVM, L1LogisticRegression
+LinearSVM(C=1.0).fit(X, y).predict(X_new)
+KernelSVM(C=1.0, gamma="scale").fit(X, y).decision_function(X_new)
+L1LogisticRegression(C=0.1).fit(X, y).predict_proba(X_new)
+```
+
+The classifiers take two classes (the second of `classes_` is the
+positive one), and `tol` / `max_iter` mean what they mean in scikit-learn.
+`KernelSVM` and `KernelSVR` build the kernel matrix in numpy and keep all of it,
+n x n doubles (800 MB at n = 10,000), where `SVC` computes rows on
+demand through a cache; that is its price.
+
+Measured with `check-classifiers.py` (in this directory): the same
+problem and stopping rule on both sides, one core, best of three,
+sklearn 1.9.1, load average under 2. Ours includes building the kernel
+matrix (`KernelSVM`) or transposing X into columns
+(`L1LogisticRegression`):
+
+| problem | ours | scikit-learn | ratio | objective (ours / sklearn) | same predictions |
+|---|---|---|---|---|---|
+| `LinearSVM`, 5,000 x 200 | 0.379 s | 0.598 s | 1.6x | 1541.06 / 1541.29 | 99.5% |
+| `LinearSVM`, 100,000 x 200 | 23.0 s | 41.7 s | 1.8x | 34503.5 / 34500.7 | 98.9% |
+| `LinearSVM`, 2,000 x 20,000 | 0.975 s | 1.894 s | 1.9x | 0.052484 / 0.052428 | 100% |
+| `KernelSVM` rbf, 2,000 x 50 | 0.068 s | 0.139 s (precomputed 0.081 s) | 2.1x (1.2x) | -1065.5026 / -1065.5026 | 100% |
+| `KernelSVM` rbf, 5,000 x 50 | 0.501 s | 1.032 s (precomputed 0.588 s) | 2.1x (1.2x) | -2628.1422 / -2628.1422 | 100% |
+| `KernelSVM` rbf, 10,000 x 50 | 2.12 s | 4.15 s (precomputed 2.61 s) | 2.0x (1.2x) | -4994.9982 / -4994.9982 | 100% |
+| `L1LogisticRegression`, 5,000 x 200, C=0.1 | 0.064 s | 0.103 s | 1.6x | 61.676698 / 61.676708 | 100% |
+| `L1LogisticRegression`, 20,000 x 500, C=0.05 | 1.20 s | 2.23 s | 1.9x | 106.77913 / 106.77915 | 100% |
+| `L1LogisticRegression`, 100,000 x 200, C=0.05 | 9.06 s | 18.7 s | 2.1x | 349.61154 / 349.61216 | 100% |
+
+The regressors, from the same script (C = 1, epsilon 0.1):
+
+| problem | ours | scikit-learn | ratio | agreement |
+|---|---|---|---|---|
+| `LinearSVR`, 5,000 x 200 | 1.32 s | 2.53 s | 1.9x | both at the 1,000-sweep cap; objective 779.99 / 790.72 |
+| `LinearSVR`, 20,000 x 500 | 22.1 s | 36.6 s | 1.7x | both at the cap; objective 3741.4 / 3712.2 |
+| `KernelSVR` rbf, 1,000 x 10 | 0.023 s | 0.043 s | 1.9x | iterations 2,172 / 2,177; predictions within 4e-4 |
+| `KernelSVR` rbf, 3,000 x 10 | 0.246 s | 0.332 s | 1.3x | iterations 7,471 / 7,581; within 5e-4 |
+| `KernelSVR` rbf, 6,000 x 10 | 1.18 s | 1.38 s | 1.2x | iterations 16,805 / 16,630; within 5e-4 |
+
+`LinearSVR` on these data does not reach its tolerance in 1,000 sweeps on
+either side, and at the cap the random order alone moves the objective
+by about 1% (scikit-learn's own seeds give 776-789 at 5,000 x 200);
+with 20,000 sweeps both land on 728.58.  `KernelSVR`'s gain shrinks
+with n here because the 10 features make each kernel row cheap for
+libsvm to compute, while ours reads the rows of a matrix of n^2
+doubles; the predictions agree to within scikit-learn's tolerance.
+
+The objectives agree to the stopping tolerance on both sides, not to
+rounding. The linear SVM on tall data stops at the 1,000-sweep cap on
+both sides, so there the two land a hair apart and 1% of the
+predictions differ. The gain is the same in all three: the same
+algorithm, run as dense loops that vectorise, against libsvm's and
+liblinear's sparse rows. The classifiers are compiled at `-O3` with
+`-fstrict-overflow`: Python's build flags carry
+`-fno-strict-overflow`, which keeps the compiler from treating the
+flat index `i*p + j` as an induction variable and makes the linear
+SVM 25% slower.
+
+## NMF: `CDNMF`
+
+`CDNMF` is scikit-learn's `NMF(solver="cd")` -- Frobenius loss, no
+regularization, no shuffling -- as the translator derives it.
+`examples/kernel-only/nmf-cd.scm` writes the half step with the
+residual R = X - W H and no Gram matrix. `--derive` memoises R H' over
+the hoisted H H' and distributes the residual through the memo, so the
+compiled half step is scikit-learn's `XHt - W HHt` with R never formed.
+Like `DerivedLasso`, it is built in three variants (`nmf_backend()`
+says which is taken: cublas when a device answers, else blas, else
+loops).
+
+```python
+from scm2cpp_lasso import CDNMF
+model = CDNMF(n_components=20, random_state=0)   # init="random", as sklearn's
+W = model.fit_transform(X)                       # H in model.components_
+```
+
+`init="random"` draws the start as scikit-learn does, so the same seed
+gives the same W and H. scikit-learn's default `"nndsvda"` is not
+offered; pass `init="custom"` with its W and H for that start.
+Measured with `check-classifiers.py --only nmf` against
+`NMF(solver="cd", init="random")` from the same seed, tol 1e-4, at most
+200 iterations, one core:
+
+| n x p, k | `CDNMF` cuBLAS | `CDNMF` BLAS | sklearn | iterations | max \|H - H_sklearn\| |
+|---|---|---|---|---|---|
+| 1,000 x 500, 10 | 0.441 s (0.6x) | 0.212 s (1.15x) | 0.244 s | 200 / 200 | 1.3e-13 |
+| 5,000 x 2,000, 20 | 7.01 s (1.6x) | 8.27 s (1.36x) | 11.25 s | 200 / 200 | 4.1e-12 |
+| 20,000 x 1,000, 30 | 16.9 s (1.8x) | 20.6 s (1.46x) | 30.1 s | 200 / 200 | 2.4e-12 |
+
+The iterates are scikit-learn's to rounding. The cuBLAS variant pays
+from the middle size on; at 1,000 x 500 its transfers cost more than
+the products save, and `backend="blas"` is the faster choice there.
+
+## Graphical lasso and RANSAC: `CDGraphicalLasso`, `LinearRANSAC`
+
+Two more of scikit-learn's algorithms, written out in the Scheme subset
+and translated plainly.
+
+- `CDGraphicalLasso` is `GraphicalLasso(mode="cd")`: `_graphical_lasso`
+  with `enet_coordinate_descent_gram` inside it, its two matrix-vector
+  products as dgemv. `alpha`, `tol`, `enet_tol`, `max_iter` and
+  `assume_centered` mean what they mean in scikit-learn. scikit-learn's
+  `costs_` (a log-determinant every iteration) is not computed.
+- `LinearRANSAC` is `RANSACRegressor` around `LinearRegression`, with
+  the same defaults. With scikit-learn installed, the subsets are drawn
+  exactly as it draws them, so `random_state` gives scikit-learn's fit.
+
+```python
+from scm2cpp_lasso import CDGraphicalLasso, LinearRANSAC
+prec = CDGraphicalLasso(alpha=0.1).fit(X).precision_
+fit = LinearRANSAC(random_state=0).fit(X, y)       # fit.coef_, fit.inlier_mask_
+```
+
+Measured through the package, one core, best of three, against
+scikit-learn on the same data (an AR(1) design with n = 20p for the
+graphical lasso, 100 iterations on both sides; a linear model with
+outliers for RANSAC, max_trials 1000):
+
+| problem | ours | scikit-learn | ratio | agreement |
+|---|---|---|---|---|
+| `CDGraphicalLasso`, p = 50 | 0.0069 s | 0.306 s | 44.5x | 100 / 100 iterations, precision to 1.8e-14 |
+| `CDGraphicalLasso`, p = 200 | 0.286 s | 1.62 s | 5.7x | precision to 2.6e-14 |
+| `CDGraphicalLasso`, p = 800 | 33.9 s | 43.8 s | 1.3x | precision to 5.0e-14 |
+| `LinearRANSAC`, 10,000 x 5 | 0.0254 s | 0.0916 s | 3.6x | 98 / 98 trials, same inliers |
+| `LinearRANSAC`, 100,000 x 10 | 0.153 s | 0.241 s | 1.6x | 97 / 97 trials, same inliers |
+| `LinearRANSAC`, 20,000 x 20 | 0.470 s | 1.19 s | 2.5x | 1,000 / 1,000 trials, same inliers |
+
+Both gains are mostly scikit-learn's Python between its compiled steps:
+the per-column bookkeeping of the graphical lasso, and the per-trial
+validation and fitting of RANSAC. They shrink as the compiled part
+grows (the graphical lasso at large p).
+
+## SGD, OMP and LARS: `SGDLinearClassifier`, `SGDLinearRegressor`, `GramOMP`, `GramLars`, `GramLassoLars`
+
+scikit-learn's own algorithms, translated:
+
+- `SGDLinearClassifier` and `SGDLinearRegressor` are
+  `SGDClassifier` / `SGDRegressor` at their defaults, run by
+  `_plain_sgd` with the points in order. Against
+  `shuffle=False` the weights are bit-identical. scikit-learn's
+  shuffling generator is not reproduced.
+- `GramOMP` is `OrthogonalMatchingPursuit` over the Gram matrix (its
+  route for n > p, taken here always; X'X and X'y by dsyrk and dgemv).
+- `GramLars` and `GramLassoLars` are `Lars` and `LassoLars` over the
+  Gram (their route for n > p, taken here always; the Gram formed in
+  numpy).
+
+```python
+from scm2cpp_lasso import SGDLinearClassifier, GramOMP, GramLassoLars
+SGDLinearClassifier().fit(X, y).predict(X_new)
+GramOMP(n_nonzero_coefs=10).fit(X, y).coef_
+GramLassoLars(alpha=0.01).fit(X, y).coef_
+```
+
+Each gives scikit-learn's coefficients, to the bit for SGD and to 1e-15
+for the rest, with the same number of epochs or steps. The speed is the
+repository README's table: 1.0-3.3x for SGD, 1.1-2.8x for OMP and
+1.2-5.0x for LARS, largest on small problems, where scikit-learn's
+Python between steps dominates.
+
+## L2 logistic regression and Huber: `LBFGSLogisticRegression`, `LBFGSHuberRegressor`
+
+`LogisticRegression(solver="lbfgs")` (the default; two classes
+binomial, more multinomial) and `HuberRegressor`, with scipy's
+L-BFGS-B written out in the subset under them. The products are dgemm
+or dgemv where a CBLAS links.
+
+```python
+from scm2cpp_lasso import LBFGSLogisticRegression, LBFGSHuberRegressor
+LBFGSLogisticRegression(C=1.0).fit(X, y).predict_proba(X_new)
+LBFGSHuberRegressor(epsilon=1.35).fit(X, y).coef_
+```
+
+Both take scikit-learn's iterations one for one and give its
+coefficients to rounding (1e-13 at worst). The speed is the repository
+README's table: 1.1-8x for the logistic regression (near 1x on a large
+binary problem, where both sides are two matrix-vector products per
+evaluation) and 2.5-18x for Huber.
+
+## Gaussian mixture, FastICA, PCA: `EMGaussianMixture`, `ParallelFastICA`, `CovPCA`
+
+`GaussianMixture(covariance_type="full")`, `FastICA` at its defaults
+and `PCA` by the covariance's eigenvectors (scikit-learn's
+"covariance_eigh", what "auto" picks for a tall design), each
+scikit-learn's algorithm translated, the products dgemm / dsyrk where a
+CBLAS links and the eigenproblem written in the subset.
+
+```python
+from scm2cpp_lasso import EMGaussianMixture, ParallelFastICA, CovPCA
+EMGaussianMixture(4, random_state=0).fit(X).predict(X)   # KMeans start: scikit-learn's
+ParallelFastICA(random_state=0).fit(X).transform(X)
+CovPCA(n_components=10).fit(X).components_
+```
+
+From the same `random_state` each takes scikit-learn's iterations and
+lands on its parameters to rounding. The speed is the repository
+README's table: 1.6-2.4x for the mixture, 1.1-1.4x for FastICA, and
+for PCA 1.3-4x up to about 100 features, 0.6-0.8x at 200-500, where
+LAPACK's eigensolver wins.
+
+## Perceptron, ARD, LassoLarsIC, L1 linear SVM: `SGDPerceptron`, `CovARDRegression`, `GramLassoLarsIC`, `L1LinearSVM`
+
+- `SGDPerceptron` is `Perceptron`. `SGDLinearClassifier` /
+  `SGDLinearRegressor` also take `loss="perceptron"` /
+  `"epsilon_insensitive"`, `learning_rate="constant"` / `"pa1"` /
+  `"pa2"` and `penalty=None`, which covers scikit-learn's replacements
+  for the passive-aggressive estimators. All run with shuffle=False.
+- `CovARDRegression` is `ARDRegression` (n >= p). With `hoist=True`,
+  the default, the kept features' Gram comes from X'X formed once;
+  `hoist=False` forms it every iteration, as scikit-learn does.
+- `GramLassoLarsIC` is `LassoLarsIC`. The whole lasso path comes from
+  the translated LARS, and the criterion is evaluated from the Gram.
+- `L1LinearSVM` is `LinearSVC(penalty="l1", dual=False)`, liblinear's
+  solve_l1r_l2_svc. It agrees with scikit-learn to the tolerance,
+  because the coordinate order comes from another generator.
+
+```python
+from scm2cpp_lasso import SGDPerceptron, CovARDRegression, GramLassoLarsIC, L1LinearSVM
+SGDPerceptron().fit(X, y)
+CovARDRegression().fit(X, y).coef_
+GramLassoLarsIC("bic").fit(X, y).alpha_
+L1LinearSVM(C=1.0).fit(X, y).coef_
+```
+
+Speed, from the repository README's table: SGD 1.2-3x, ARD 1.6-1.9x
+(3x with the Gram hoisted), LassoLarsIC 8-20x, the L1 SVM 1.1-1.5x.
+
+## GLMs, CV and Theil-Sen: `LBFGSPoissonRegressor`, `LBFGSGammaRegressor`, `LBFGSTweedieRegressor`, `LBFGSLogisticRegressionCV`, `GramLarsCV`, `GramLassoLarsCV`, `GramOMPCV`, `WeiszfeldTheilSen`
+
+- The GLMs are `PoissonRegressor`, `GammaRegressor` and
+  `TweedieRegressor` (solver "lbfgs"), over the same L-BFGS-B as
+  `LBFGSLogisticRegression`.
+- The CV classes reproduce `LogisticRegressionCV`, `LarsCV`,
+  `LassoLarsCV` and `OrthogonalMatchingPursuitCV`: the same folds (KFold
+  / StratifiedKFold without shuffling), the same choice of C, alpha or
+  number of columns, and the same refit.
+- `WeiszfeldTheilSen` is `TheilSenRegressor` on scikit-learn's own
+  subsets.
+
+```python
+from scm2cpp_lasso import LBFGSPoissonRegressor, LBFGSLogisticRegressionCV, GramLassoLarsCV
+LBFGSPoissonRegressor(alpha=1e-3).fit(X, counts).predict(X_new)
+LBFGSLogisticRegressionCV().fit(X, labels).C_
+GramLassoLarsCV().fit(X, y).alpha_
+```
+
+Speed, from the repository README's table: the GLMs 1.1-6x, the CV
+estimators 1.6-6.6x, Theil-Sen 1.3-1.8x (its subset drawing dominates).
+
+## Feature selection: `f_regression`, `f_classif`, `SelectKBest`, `SelectPercentile`, `GramRFE`
+
+scikit-learn's univariate F statistics and its selectors over them,
+and `RFE` around `Ridge` / `LinearRegression`, translated.
+`f_regression` also takes an n x q array of targets: the q results in
+one product, on the device when one answers.
+
+```python
+from scm2cpp_lasso import SelectKBest, GramRFE, f_regression
+SelectKBest("f_regression", k=20).fit(X, y).get_support()
+GramRFE(alpha=1.0, n_features_to_select=10).fit(X, y).ranking_
+F, p = f_regression(X, Y)          # Y: n x q, every target at once
+```
+
+The same scores, ranking and features as scikit-learn. RFE is 80-1,000x
+faster (one Gram and one Cholesky factor, downdated, instead of a refit
+per round), and f_regression over many targets 30-50x. The Gram-route
+estimators (`GramLassoLarsIC`, `GramLarsCV`, `GramLassoLarsCV`,
+`GramLars`, `GramLassoLars`, `CovPCA`) form a large X'X on the device
+when one answers. `SCM2CPP_NO_DEVICE=1` keeps everything on the CPU.
+
+## Against Intel's sklearnex
+
+sklearnex (scikit-learn-intelex 2026.2, oneDAL underneath) replaces
+some scikit-learn estimators with its own. `bench/sklearnex-compare.py`
+times the ones with a counterpart here, sklearnex's estimators
+imported directly (not patched in) so that scikit-learn itself is
+timed in the same process. The ratios are scikit-learn's or
+sklearnex's time over the package's best (its CPU build, or the device
+build where one answers); the load average was 8 to 11, so the times
+move by some tens of percent.
+
+| estimator, n x p | iterations (sk / sklearnex / ours) | vs scikit-learn, one core | vs sklearnex, one core | vs sklearnex, 20 cores |
+|---|---|---|---|---|
+| LogisticRegression, 200,000 x 100 | 6 / 12 / 6 | 10.6x (device) | 8.4x (device) | 4.4x (device); on the CPU 0.5x |
+| Ridge (alpha 1), 200,000 x 200 | - | 2.1x | 1.2x | 0.3x |
+| PCA (10 components), 100,000 x 50 | - | 1.5x | 2.4x | 1.7x |
+| PCA (10 components), 50,000 x 500 | - | 1.1x | 1.4x | 0.5x |
+| SVC (rbf), 5,000 x 20 | 4,858 / 7 / 4,882 | 1.2x | 1.4x | 0.5x |
+| SVR (rbf), 5,000 x 20 | 12,821 / 10 / 12,824 | 1.3x | 1.4x | 0.4x |
+| RFE(Ridge), 20 of 200 kept, 20,000 x 200 | - | 216x | 128x | 78x |
+
+On one core the package is ahead everywhere. sklearnex uses every
+core by default, and the package's kernels are single-threaded, so
+with 20 cores sklearnex is faster on Ridge, the larger PCA and the
+SVMs; what stays ahead is the device build (logistic regression) and
+an algorithm that does less work (RFE: the Gram once, the Cholesky
+factor downdated per drop, where scikit-learn refits). sklearnex is
+not scikit-learn's arithmetic: its logistic regression takes 12
+iterations to scikit-learn's 6 and its coefficients differ from
+scikit-learn's by 6e-3, where the package's agree to 4e-15; its SVMs
+report a different iteration count. This sklearnex accelerates no
+lasso, so the lasso tables above stand against scikit-learn alone.
+
+## The GPU build
+
+Where the CUDA toolkit is at install time, `ParallelFastICA`,
+`EMGaussianMixture`, `LBFGSLogisticRegression`,
+`LBFGSLogisticRegressionCV`, `LBFGSHuberRegressor` and the `LBFGS*`
+GLMs also get the same Scheme translated with `--cuda`
+(`libscm2cpp_gpu.so`). Its loops are CUDA kernels and its products
+cuBLAS, on arrays that stay on the device. The classes take it when a
+device answers and the data has a million entries or more, and give
+the CPU build's iterations and results to the last digits. Measured
+against scikit-learn: 10-22x for the L-BFGS fits, 6.7x FastICA, 3x the
+mixture and LogisticRegressionCV (the repository README's table).
+`SCM2CPP_NO_DEVICE=1` keeps everything on the CPU.
+
+## Logistic regression with L1
+
+`CovLogistic` solves L1-penalized logistic regression by
+majorization: the logistic Hessian is bounded by X'X/4, so the
+quadratic term is the same Gram matrix, fixed once, and each outer
+round costs one gradient pass before handing the majorizer to the
+same coordinate descent the lasso uses.  Objectives agree with
+scikit-learn's `LogisticRegression(penalty="l1",
+fit_intercept=False)` at `C = 1/(n lam)` to 9e-15.
+
+## Group lasso
+
+`CovGroupLasso` penalizes whole groups -- `lam * sum_g sqrt(|g|)
+||b_g||` -- so correlated features enter or leave together.  Block
+coordinate descent on the same Gram machinery: each block visit is
+one majorized proximal step (the group's Gram block dominated by its
+top eigenvalue, found once), descending monotonically.  With size-one
+groups it reduces exactly to the lasso -- verified against sklearn to
+9e-16 -- and at convergence the group KKT conditions hold to 2e-12.
+
+## Bootstrap on the GPU
+
+`bootstrap` draws pairs-bootstrap resamples and refits them all at
+one lambda.  Each resample's Gram matrix is `X' diag(m) X` for its
+multiplicity counts `m`, written as `Xs' Xs` with `Xs = diag(sqrt m) X`
+so that numpy hands it to the symmetric rank-k product (`dsyrk`,
+half the flops of `X' (diag(m) X)`) -- one BLAS call -- and because
+the problems are independent, the descents run as one batch: on the
+GPU, one block of threads per resample, each reading its own Gram
+matrix.  The products are the cost: at n=5000, p=1000 a hundred
+resamples took 18.1 s as general products and 11.2 s as `dsyrk`,
+descents included.
+
+```python
+betas = model.bootstrap(lam, n_boot=500, seed=0)   # (500, p)
+freq = (abs(betas) > 1e-9).mean(axis=0)            # selection frequency
+```
+
+Requires constructing the model from `X, y` (a Gram matrix alone
+cannot be resampled by rows).  GPU and CPU agree to machine
+precision.
+
+## A design with structure
+
+When the design matrix has structure, forming X'X the general way is
+the wrong move.  `kernel` exposes the translated functions directly,
+and [`scm2cpp-tfs`](https://pypi.org/project/scm2cpp-tfs/) does exactly
+that for moving-average designs: it builds the Gram matrix from a
+series' prefix sums in O(n p) time, never forming the design.  That
+package stands alone -- it carries its own copy of this descent -- so
+neither installs the other.
+
+## Where this comes from
+
+The C++ this compiles is committed to `python/scm2cpp-lasso/` in the
+scm2cpp repository.  `CovLasso` and the rest run on the translation of
+`examples/kernel-only/lasso-cov.scm`, a covariance-update sweep
+written by hand at the Gram level; `DerivedLasso` runs on the
+translation of `examples/kernel-only/lasso-design.scm`, the naive
+residual-form descent over X as numpy holds it, with the translator's `--derive --blas`
+deriving the same covariance form from it.  The two are the same
+sweep over the same `dsyrk`, and they time the same.  To refresh the
+committed C++ after changing the Scheme, run `regenerate.sh` -- that
+step, and only that step, needs Racket.
+
+## License
+
+MIT, the same as scm2cpp.