deepbox

Conventions & Limits

Defaults that differ from NumPy, pandas, scikit-learn and PyTorch, and what Deepbox 1.x does not do yet.

Defaults that differ from the Python libraries

Deepbox follows the Python libraries closely, with a few exceptions. These defaults are kept in 1.x for compatibility and may change in 2.0. When you need to match a reference library, pass the option in the last column.

API	Deepbox default	Reference library
`tensor()` default dtype	`float32`	NumPy: `float64`. PyTorch: `float32`.
`gelu`, `GELU`	tanh approximation	PyTorch: exact (`approximate: "none"`)
`mape`	percentage	scikit-learn returns a fraction. Use `meanAbsolutePercentageError` for that.
`PowerTransformer` `standardize`	`false`	scikit-learn: `true`
`DataFrame.ewm` `adjust` and `bias`	`false` and `true`	pandas: `true` and `false`
`DataFrame.groupBy`	keys in first-appearance order, missing keys kept	pandas: sorted, missing dropped. Pass `{ sort: true, dropna: true }`.
`DataFrame.valueCounts`	missing values counted	pandas: dropped
`Series.str.replace` `regex`	`true`	pandas: `false`
`InstanceNorm` `affine`	`true`	PyTorch: `false`
`Upsample` `alignCorners` (bilinear)	`true`	PyTorch: `false`
`RNN`, `LSTM`, `GRU` `batchFirst`	`true`	PyTorch: `false`
`RMSNorm` `eps`	`1e-5`	PyTorch: machine epsilon of the dtype
`FullTransformer` final `LayerNorm`	none	PyTorch `nn.Transformer`: yes. Pass `{ finalNorm: true }`.
`DecisionTree`, `RandomForest` `maxDepth`	`10`	scikit-learn: unlimited
`LinearSVC` `loss`	`"hinge"`	scikit-learn: squared hinge
`LinearSVR` `epsilon`	`0.1`	scikit-learn: `0`
`GaussianMixture` `covarianceType`	`"diag"`	scikit-learn: `"full"`
`KernelRidge` `gamma`	`1.0`	scikit-learn: `1 / nFeatures`
`calibrationCurve` `nBins`	`10`	scikit-learn: `5`
Averaged metrics without `average`	`"binary"` for two classes, `"weighted"` for more	scikit-learn: `"binary"`, and an error for multiclass input
`solveTriangular` `lower`	`true`	SciPy: upper (`lower=False`)
`corrcoef` on a matrix	rows are observations	NumPy: rows are variables (`rowvar=True`)
`expon` and `gamma` scale	a rate	SciPy: `scale = 1 / rate`
`imshow` row 0	bottom of the y axis	matplotlib: top. Pass `{ origin: "upper" }`.
`clip(t, min, max)` with `min > max`	throws `InvalidParameterError`	NumPy and PyTorch: return `max`

Naming differences

Fitted values are getters without scikit-learn's trailing underscore: model.coef, model.intercept, kmeans.labels, pca.components.
Options are camelCase (nEstimators, randomState, maxDepth), and optimizers take flat options such as beta1 and beta2 instead of a betas tuple.
crossValScore, crossValidate and GridSearchCV take a number of folds as cv, not a splitter object. Splitters in deepbox/preprocess return arrays of { trainIndex, testIndex }.
Seeded random streams are reproducible inside Deepbox. They are not bit-identical to NumPy or PyTorch.

Known limits

complex64 and complex128 appear in the DType type, but tensors cannot be created with them yet. FFT functions return { real, imag } tensors.
WebGPU and WASM accelerate a subset of operations. The rest throw a DeviceError on WebGPU. See Devices & Execution.
Deepbox is plain TypeScript. Operations bound by BLAS, such as large matrix multiplies and decompositions, are slower than NumPy with LAPACK.
eig supports real eigenvalues only, and the linalg functions work on 2-D matrices.
WorkerPool runs its chunks in order on the calling thread, so it adds no parallelism. sparseSolve expands the matrix into a dense workspace, so memory grows as O(N^2).
Dates are local-time JavaScript Date values, and there is no time zone support.

Errors

Every module throws subclasses of DeepboxError: ShapeError, DTypeError, InvalidParameterError, NotFittedError, DeviceError, ConvergenceError and others. Catch the specific class you care about, and use instanceof DeepboxError to catch them all.

errors.tsTypeScript

import { DeepboxError, ShapeError } from "deepbox/core";import { add, tensor } from "deepbox/ndarray"; try {  add(tensor([1, 2, 3]), tensor([1, 2]));} catch (error) {  console.log(error instanceof ShapeError, error instanceof DeepboxError);}

Found a mistake? Open an issue.