haphazard 1.2.4

A modular framework for registering and running haphazard datasets and models.

Haphazard

A Python package for haphazard dataset and model management.
Provides a standardized interface for loading datasets (with online normalization) and models, running experiments, and extending with custom datasets or models.

---

Table of Contents

• Installation
• Quick Start
• Datasets
• Models
• Normalization
• Versions
• Contributing
• License

---

Installation

Install via pip (after packaging):

pip install haphazard

---

Project Structure

The Haphazard package follows a modular, extensible design:

haphazard/
├── __init__.py
├── data/                          # Dataset management
│   ├── __init__.py
│   ├── base_dataset.py
│   ├── mask.py
│   └── datasets/
│       ├── __init__.py
│       ├── dummy_dataset/
│       ├── magic04/
│       ├── a8a/
│       ├── imdb/
│       ├── susy/
│       ├── higgs/
│       ├── dry_bean/
│       └── gas/
├── models/                        # Model management
│   ├── __init__.py
│   ├── base_model.py
│   └── model_zoo/
│       ├── __init__.py
│       ├── dummy_model/
│       ├── dynfo/
│       ├── fae/
│       ├── nb3/
│       ├── ocds/
│       ├── olifl/
│       ├── olvf/
│       ├── orf3v/
│       └── ovfm/
├── normalization/                 # New in v1.1.0
│   ├── __init__.py
│   ├── base_normalizer.py
│   └── normalizer_zoo/
│       ├── __init__.py
│       ├── decimal_scale.py
│       ├── mean.py
│       ├── minmax.py
│       ├── no_normalization.py
│       ├── unit_vector.py
│       └── zscore.py
└── utils/                         # Utilities
    ├── __init__.py
    ├── file_utils.py
    ├── metrics.py
    └── seeding.py

Notes

• data/base_dataset.py defines the BaseDataset class and integrates normalization support.
• normalization/base_normalizer.py defines a universal BaseNormalizer base class.
• normalizer_zoo/ provides built-in normalizers (e.g., zscore, mean, decimal_scale, no_normalization).
• models/base_model.py defines BaseModel, used by all models in model_zoo/.
• Dynamic registration of datasets, models, and now normalizers is handled via decorators.

---

Quick Start

from haphazard import load_dataset, load_model

# Load dataset
dataset = load_dataset("dummy", n_samples=100, n_features=10, norm="zscore")

# Load model
model = load_model("dummy")

# Run model
outputs = model(dataset)
print(outputs)

---

Datasets

• All datasets inherit from BaseDataset.
• Example dataset: DummyDataset.
• Main interface:

from haphazard import load_dataset
dataset = load_dataset(
   "magic04", 
   base_path="./data", 
   scheme="probabilistic", 
   availability_prob=0.5,
   norm="none"
   )

x, y = dataset.x, dataset.y
mask = dataset.mask

Dataset Attributes

• name: str - dataset name
• task: "classification" | "regression"
• haphazard_type: "controlled" | "intrinsic"
• n_samples, n_features: int
• num_classes: int (for classification)
• normalizer: optional (default="none") defines a normalization scheme, if used

Available Datasets

• WPBC ("wpbc")
• Ionosphere ("ionosphere")
• WDBC ("wdbc")
• Australian ("australian")
• WBC ("wbc")
• Diabetes-f ("diabetes_f")
• Crowdsense(c3) ("crowdsense_c3")
• Crowdsense(c5) ("crowdsense_c5")
• German ("german")
• IPD ("ipd")
• Svmguide3 ("svmguide3")
• KR-vs-KP ("krvskp")
• Spambase ("spambase")
• Spamassasin ("apamassasin")
• Magic04 ("magic04")
• IMDB ("imdb")
• A8A ("a8a")
• Diabetes US ("diabetes_us")
• SUSY ("susy")
• HIGGS ("higgs")
• Dry Bean (dry_bean"")
• GAS ("gas")

---

Models

• All models inherit from BaseModel.
• Example model: DummyModel.

from haphazard import load_model
model = load_model("dummy")
model_params = {}  # Hyperparameters of the model
outputs = model(dataset, model_params)

Output

• Classification: labels, preds, logits, time_taken, is_logit
• Regression: targets, preds, time_taken

Available Models

• Dummy - testing/prototyping.
• NB3, FAE - Naive Bayes based models.
• DynFo, ORF3V - Decision stump based models.
• OLVF, OLIFL, OVFM, OCDS - Linear classifier based models.

---

Normalization

Overview

Introduced in v1.1.0, the normalization module provides standardized interfaces for online feature normalization across datasets and models.

Using Built-in Normalizers

from haphazard import load_dataset, load_normalizer

dataset = laod_dataset(
   "a8a",
   base_path="./",
   scheme="sudden",
   num_chunks=4,
   norm="none"  # No normalization applied internally
)

# Load z-score normalization
normalizer = load_normalizer("zscore", num_features=dataset.n_features)
for x, mask, y in dataset:
   x_norm = normalizer(x, mask)
   ...  
   # Further processing as required

# Load mean normalization
normalizer = load_normalizer("mean", num_features=dataset.n_features)
X, Mask, Y = dataset.x, dataset.y, dataset.mask
for x, mask, y in zip(X, Mask, Y):
   x_norm = normalizer(x, mask)
   ...  
   # Further processing as required

or

from haphazard import load_dataset, load_normalizer

dataset = laod_dataset(
   "a8a",
   base_path="./",
   scheme="sudden",
   num_chunks=4,
   norm="zscore"  # Apply normalization internally
)

# Load z-score normalization
# Iterating through the dataset normalizes the input at every step
for x_norm, mask, y in dataset:
   ...  
   # Further processing as required


# Un-normalized values can still be extracted using the following
X, Mask, Y = dataset.x, dataset.y, dataset.mask

# Load mean normalization
normalizer = load_normalizer("mean", num_features=dataset.n_features)
for x, mask, y in zip(X, Mask, Y):
   x_norm = normalizer(x, mask)
   ...  # Further processing as required

Available Normalizers

Normalizer Name	Description
decimal_scale	Scales feature values by powers of 10
mean	Online mean normalization by substracting running mean
minmax	Scales features by the range (max-min) of the feature value observed
no_normalization	Pass-through, no normalization applied
unit_vector	Normalizes observed values into a unit-vector (scales by L2 norm)
zscore	Online mean normalization using running mean and variance (substract mean, scale by variance)

Extending Normalization

Developers can register their own normalization schemes-see Contributing.

---

Versions

v1.2.4

• Freeze required package versions.

v1.2.3

• Remove internal normalization from HI2.
• Freeze package versions of torch and tochvision for compatibility.

v1.2.2

• Debug imports in HI2 model.

v1.2.1

• Removed check for passing empty hyperparameter dict from BaseModel to support default hyperparameter values.

v1.2.0: datasets module changed to dataset_zoo

• Added 15 dataset: WPBC, Ionosphere, WDBC, Australian, WBC, Diabetes-f, Crowdsense(c3), Crowdsense(c5), German, IPD, Svmguide3, KR-vs-KP, Spambase, Spamassasin and Diabetes US.
• Set default value of mask scheme as "probabilistic". For datasets with intrinsic haphazardness scheme value is forced to "intrinsic".
• Minor updates (name and comments) in A8A, GAS, SUSY, HIGGS and Dry Bean datasets.
• Seed the random shuffle of IMDB dataset appropriately.

v1.1.8

• Hardened seeding implementation in haphazard/utils/seeding.py.
• Immplemented optional imports in HI2 model.
• Update string regresentation of BaseDataset and BaseModel.
• Strict data-typing of x, y and mask in DummyModel.

v1.1.7

• Bug-fix: Resolve dtype of y in v2f.py for different criterion.

v1.1.6

• Add task heads to V2F: Transformer, MLP, ODL
• Available task heads for V2F: LSTM, Transformer, MLP, ODL

v1.1.5

• Bug fix: Squeezed output logit of V2F for criterion compatibility

v1.1.4

• Added model V2F

v1.1.3

• Bug fix: HI2 is not deterministic.
• Set self.determinitic = False in HI2 module.

v1.1.2

• Add model HI2

v1.1.1

• Added model HapTransformer
• Update __repr__ of base_dataset

v1.1.0

Major Features

• Added Normalization Framework

  • Introduced new module normalization/ with base and zoo submodules.
  • Built-in normalizers: mean, zscale, no_normalization, etc.
  • Unified registration via @register_normalizer.
  • Datasets and models now support integrated normalization.

Modifications

• Updated:

  • data/base_dataset.py - normalization integration.
  • models/base_model.py - normalization compatibility.
  • model_zoo and datasets modules - decorator consistency.

v1.0.9

• Added model FAE

• Bug Fix

Update the X2 calculation in NB3 model

v1.0.8

• Added model NB3.

v1.0.7

• Bug Fix

• Set RunOCDS.determministic = False as it uses random initialization.
• Not passing 'tau' (or passing None) hyperparameter in OCDS will now result in using tau=np.sqrt(1.0/t) as a varied step size, as mentioned in OCDS paper (but not GLSC paper).

v1.0.6

• Added datasets A8a, IMDB, Susy, and Higgs.

v1.0.5

• Added model OCDS.

• Bug Fixes and Improvements:

• In haphazard/models/model_zoo/dynfo/dynfo.py:
  Updated the dropLearner() method to prevent errors when attempting to remove the last remaining weak learner.

  def dropLearner(self, i):
      if len(self.learners) == 1:
          return
      self.learners.pop(i)
      self.weights.pop(i)
      self.acceptedFeatures.pop(i)
      assert len(self.weights) == len(self.learners) == len(self.acceptedFeatures)
  

  This ensures stability in low-learner configurations and prevents IndexError during runtime.

v1.0.4

• Added model ORF3V.

NOTE:

• ORF3V also requires an initial buffer, which works similarly to DynFo.

• ORF3V depends on the optional package tdigest, which requires Microsoft Visual C++ Build Tools.

• To install with this dependency:

  1. Visit: https://visualstudio.microsoft.com/visual-cpp-build-tools/

  2. Download and install Build Tools for Visual Studio. During installation:

     • Select “Desktop development with C++” workload.
     • Ensure MSVC v143 or later, Windows 10/11 SDK, and CMake tools are checked.

• After installation, restart your terminal and re-run:

  pip install haphazard[orf3v]
  

  or

  pip install haphazard[all]   # installs all optional dependencies
  

• The package can still be used without installing tdigest; only the ORF3V model will be unavailable.

• Bug Fixes and Improvements:

• In haphazard/models/model_zoo/dynfo/__init__.py: corrected docstring from "Initialize the OLVF runner class." -> "Initialize the DynFo runner class."

• In haphazard/models/model_zoo/dynfo/dynfo.py: changed

  return int(np.argmax(wc)), float(max(wc))
  

  to

  return int(np.argmax(wc)), float(wc[1])
  

  for correct AUROC/AUPRC compatibility.

v1.0.3

• Added model DynFo

NOTE:

• DynFo requires an initial buffer.
• If no initial buffer size is provided, it is set to 1.
• The length of the output labels/preds/logits is reduced by the initial buffer size.

v1.0.2

• Added model OVFM

v1.0.0

(Considered to be the base version, ignore versions before this)

• Includes models OLVF and OLIFL natively.
• Includes datasets Magic04, Dry Bean and Gas. (Does not include raw files to read from, please use base_path argument to point to relevant path containing the raw files).

---

Contributing

Haphazard supports easy extensibility for new datasets, models, and normalizers.

Adding a new dataset

1. Create a new folder under haphazard/data/dataset_zoo/, e.g., my_dataset/.
2. Add __init__.py:

from ...base_dataset import BaseDataset
from ...dataset_zoo import register_dataset
import numpy as np

@register_dataset("my_dataset")
class MyDataset(BaseDataset):
    def __init__(self, base_path="./", **kwargs):
        self.name = "my_dataset"
        self.haphazard_type = "controlled"
        self.task = "classification"
        super().__init__(base_path=base_path, **kwargs)

    def read_data(self, base_path="./"):
        # Load or generate x, y
        x = np.random.random((100, 10))
        y = np.random.randint(0, 2, 100)
        return x, y

3. The dataset is automatically registered and can be loaded with load_dataset("my_dataset").

Adding a new model

1. Create a new folder under haphazard/models/model_zoo/, e.g., my_model/.
2. Add __init__.py:

from ...base_model import BaseModel, BaseDataset
from ...model_zoo import register_model
import numpy as np

@register_model("my_model")
class MyModel(BaseModel):
    def __init__(self, **kwargs):
        self.name = "MyModel"
        self.tasks = {"classification", "regression"}
        self.deterministic = True
        self.hyperparameters = set()
        super().__init__(**kwargs)

    def fit(self, dataset: BaseDataset, model_params=None, seed=42):
        # Dummy implementation
        preds = []
        for x, mask, y in dataset:
            preds.append(int(np.random.randint(0, 2)))
        if dataset.task == "classification":
            return {
                "labels": y,
                "preds": preds,
                "logits": preds,
                "time_taken": 0.0,
                "is_logit": True
            }
        elif dataset.task == "regression":
            return {
                "targets": dataset.y,
                "preds": preds,
                "time_taken": 0.0,
            }

3. The model is automatically registered and can be loaded with load_model("my_model").

Adding a New Normalizer

1. Create a folder under:

   haphazard/normalization/normalizer_zoo/my_normalizer/
   

2. Add __init__.py:

   from ...base_normalizer import OnlineNormalization
   from ...normalizer_zoo import register_normalizer
   import numpy as np
   from numpy.typing import NDArray
   
   @register_normalizer("my_normalizer")
   class MyNormalizer(OnlineNormalization):
       def __init__(self, num_features: int, replace_with: float | str = "nan"):
           # initialize required parameters
           super().__init__(num_features, replace_with)
   
       def update_params(self, x: NDArray[np.float64], indices: NDArray[np.int64]) -> None:
           # Update parameters
   
       def normalize(self, x: NDArray[np.float64], indices: NDArray[np.int64]) -> NDArray[np.float64]:
           # normalize x
           x_norm = ...
           return x_norm
   

3. Load dynamically with:

   from haphazard import load_normalizer
   normalizer = load_normalizer("my_normalizer", num_features=10)
   

---

License

MIT License.