lambdaml 1.0.3

Gradient-free machine learning for any numpy-compatible function

LambdaML

Gradient-free machine learning. Give it any function; it learns the parameters.

LambdaML lets you use any numpy-compatible function as your model and automatically fits its parameters using numerical (finite-difference) differentiation — no hand-derived gradients required. The "lambda" really can be anything: logistic regression, a neural network with custom activations, a physics equation, a learnable signal transform, or something entirely your own.

---

Quick-start

pip install lambdaml

import numpy as np
from lambdaml import LambdaClassifierModel, Optimizer, DiffMethod, LRSchedule

# 1. Write your model — anything numpy-compatible works
def my_model(x, p):
    return (np.tanh(p['w'].dot(x) + p['b']) + 1) / 2

# 2. Initial parameters (scalars or numpy arrays)
p = {'w': np.zeros(2), 'b': 0.0}

# 3. Create and fit
model = LambdaClassifierModel(
    f=my_model,
    p=p,
    diff_method=DiffMethod.COMPLEX_STEP,   # recommended
    l2_factor=0.001,
    optimizer=Optimizer.ADAM,
    lr_schedule=LRSchedule.cosine_annealing(T_max=100),
)
model.fit(X_train, Y_train, n_iter=100, lr=0.01,
          early_stopping=True, patience=10, verbose=True)

print(model.score(X_test, Y_test))       # accuracy
print(model.predict_proba(X_test))       # probabilities

For regression, swap in LambdaRegressorModel with loss='mse', 'mae', 'huber', or 'pseudo_huber'.

See the examples/ folder for runnable scripts and the LambdaML_Showcase.ipynb notebook for an interactive walkthrough with charts.

---

What is finite-difference differentiation?

The term you're looking for is finite-difference approximation (sometimes called numerical differentiation). Rather than deriving f′(θ) analytically, we estimate it by evaluating the function at nearby points:

f'(θ) ≈ [f(θ+h) - f(θ-h)] / (2h)     ← Central difference, O(h²)

LambdaML supports six methods with different accuracy/cost trade-offs:

Method	Order	f-evals/param	Notes
Forward	O(h)	1	Fast, low accuracy
Backward	O(h)	1	Fast, low accuracy
Central	O(h²)	2	Default — good balance
Five-Point	O(h⁴)	4	High accuracy, smooth f
Complex-Step	O(h²)	1 (complex)	Recommended — no cancellation error
Richardson	O(h⁴)	4	High accuracy, no complex inputs needed

Left: all six estimates on a known function. Right: absolute error vs step size h — complex-step never hits the cancellation-error floor.

Is it tractable? Yes, for models up to ~10k parameters. Each gradient step costs O(n_params) forward passes instead of O(1) for analytic backprop. For small-to-medium models on a CPU+numpy backend this is entirely practical.

---

The lambda can be any function

Six completely different model functions, one .fit() call:

From top-left: logistic regression, tanh, sine activation (non-standard), Gaussian RBF, softplus, and a physics-inspired decay+oscillation model σ(a·exp(−λ|x₀|)·cos(ω·x₁+φ)) — the kind of thing nobody derives analytically.

---

Neural network with numerically computed gradients

A 2-layer ELU network on non-linearly separable data, fitted entirely via finite-difference backprop. No autograd, no torch, no chain rule.

Clockwise from top-left: log-loss curve, final decision boundary, weight trajectories for hidden and output layers, bias evolution across epochs.

---

Regression — recovering true sine parameters

Starting from wrong values (a=2.5, ω=0.4, φ=1.8, c=−1) on outlier-corrupted data, the optimizer converges back to the true parameters using pseudo-Huber loss (complex-step safe).

---

Optimizer comparison

SGD vs Momentum vs RMSProp vs Adam on the same logistic task:

---

Derivative method benchmark

All 6 methods on the same problem — speed, accuracy, and Pareto trade-off:

---

Regularization — L1 vs L2

With the corrected L1 formula (Σ|θ| not Σθ — a bug in the original), L1 now induces true sparsity on a 10-feature problem where only features 0 and 1 matter:

---

Learning rate schedules

Five schedules visualised and compared for convergence speed:

---

Gradient accuracy verification

Per-component absolute error vs an analytically known gradient — complex-step and Richardson hit near-machine-precision:

---

API reference

LambdaClassifierModel(f, p, **kwargs)

Parameter	Default	Description
f	—	Model: f(x, p) → float ∈ (0,1)
p	—	Parameter dict (scalars or numpy arrays)
diff_method	DiffMethod.CENTRAL	Finite-difference method
diff_h	None	Custom step size (None = optimal default per method)
l1_factor	0.0	L1 regularization strength
l2_factor	0.01	L2 regularization strength
regularize_bias	False	Whether to regularize b* params
optimizer	Optimizer.ADAM	sgd, momentum, rmsprop, adam
lr_schedule	None (constant)	Learning rate schedule callable

Methods: .fit(X, Y, n_iter, lr, batch_size, early_stopping, patience, verbose, validation_data) · .predict(X) · .predict_proba(X) · .score(X, Y) · .compute_loss(X, Y) · .loss_history

LambdaRegressorModel(f, p, loss='mse', **kwargs)

Parameter	Default	Description
loss	'mse'	'mse', 'mae', 'huber', 'pseudo_huber'
huber_delta	1.0	Threshold for Huber / pseudo-Huber

Methods: .fit(...) · .predict(X) · .score(X, Y) (R²)

DiffMethod · Optimizer · LRSchedule

# Derivative methods
DiffMethod.FORWARD | BACKWARD | CENTRAL | FIVE_POINT | COMPLEX_STEP | RICHARDSON

# Optimizers
Optimizer.SGD | MOMENTUM | RMSPROP | ADAM

# LR schedules
LRSchedule.constant()
LRSchedule.step_decay(drop=0.5, epochs_drop=10)
LRSchedule.exponential_decay(k=0.01)
LRSchedule.cosine_annealing(T_max=100)
LRSchedule.warmup_cosine(warmup=10, T_max=100)

---

Bug fixes from the original library

Bug	Original	Fixed
Epsilon	float16.eps ≈ 0.001 — catastrophically large	Float64-optimal per method (~6e-6 for central)
L1 regularization	Summed raw θ — negative weights reduced penalty	Summed |θ| using smooth complex-safe approximation
Closure-in-loop	Array gradient loop captured last index for all closures	Fixed with factory functions
L1/L2 complex-step safety	float() cast stripped imaginary part	Uses v*v and sqrt(v*v+eps) to preserve imaginary parts
No test split	Accuracy reported on training data	Train/test split in all examples

---

Is LambdaML useful for Kaggling?

As a primary model for large nets — rarely. As a prototyping and ensembling tool — genuinely yes.

The core insight: LambdaML decouples your model definition from gradient computation. Anywhere you want a custom functional form but don't want to derive its gradients by hand, LambdaML fills that gap.

Konkret use cases for Kaggling: fitting domain equations with unknown parameters (physics-based pricing, pharmacokinetics, decay curves); directly optimising non-differentiable competition metrics (NDCG, F-beta, Cohen's kappa) as the loss function; building exotic meta-learners in stacking ensembles; small-data + custom hypothesis problems where sklearn doesn't have your model form.

---

Project structure

LambdaML/
├── lambdaml/                # Installable package (pip install lambdaml)
│   ├── __init__.py
│   ├── lambda_model.py      # LambdaClassifierModel, LambdaRegressorModel, Optimizer
│   └── lambda_utils.py      # NumericalDiff, GradientComputer, Regularization, LossFunctions, LRSchedule
├── pyproject.toml           # Package metadata
├── LambdaML_Showcase.ipynb  # Interactive notebook with all charts
├── examples/
│   ├── example_tanh_regression.py
│   ├── example_neural_network.py
│   ├── example_diff_methods.py
│   └── example_regressor.py
├── assets/
│   ├── fig_decision_boundaries.png
│   ├── fig_derivative_methods.png
│   ├── fig_diff_benchmark.png
│   ├── fig_gradient_accuracy.png
│   ├── fig_lr_schedules.png
│   ├── fig_neural_network.png
│   ├── fig_optimizers.png
│   ├── fig_regularization.png
│   └── fig_sine_regression.png
├── data/
│   └── circles.csv
└── legacy/                  # Original library files (pre-rewrite)

---

License

See LICENSE.