API Reference¶

All public classes and functions are importable directly from the text4gcn top-level package.

from text4gcn import (
    GNN,
    TextPipeline,
    Builder,
    Layer,
    GraphBuildStats,
    RefinementConfig,
    TextGCNTrainer,
    TrainingResult,
    apply_refinement,
    compute_adjacency_stats,
    load_adjacency_stats,
    set_verbosity,
)

Builder classes live in text4gcn.builder:

from text4gcn.builder import Frequency, Semantic, Syntactic

Module map¶

Class / Function	Import path	Description
`GNN`	`text4gcn`	Train a GCN on a text graph
`TrainingResult`	`text4gcn`	Structured output of `GNN.fit()`
`TextPipeline`	`text4gcn`	Preprocess raw text into node features
`Builder`	`text4gcn`	Enum of available adjacency strategies
`Layer`	`text4gcn`	Enum of available GNN layer types (`GCN` only)
`TextGCNTrainer`	`text4gcn`	Lower-level training loop behind `GNN.fit()`
`GraphBuildStats`	`text4gcn`	Edge/memory stats from adjacency builders
`compute_adjacency_stats`	`text4gcn`	`GraphBuildStats` of an adjacency matrix in memory
`load_adjacency_stats`	`text4gcn`	`GraphBuildStats` of a saved adjacency file
`Frequency`	`text4gcn.builder`	TF-IDF / PMI adjacency
`Semantic`	`text4gcn.builder`	Word2Vec / BERT / cosine adjacency
`Syntactic`	`text4gcn.builder`	Dependency parsing / LIWC adjacency
`RefinementConfig`	`text4gcn`	Graph refinement parameters
`apply_refinement`	`text4gcn`	Apply refinement to loaded corpus
`NodeMetadata`	`text4gcn.graph`	Per-node metadata returned by refinement
`save_corpus`	`text4gcn.datasets`	Write your own corpus in the input format
`twenty_newsgroups`	`text4gcn.datasets`	Download 20 Newsgroups in the input format
`download`	`text4gcn.datasets`	Download R8, R52, Ohsumed or PubMed
`set_verbosity`	`text4gcn`	Show diagnostic logs (shapes, model structure)

Other public helpers, for advanced use: GraphRefinement and build_node_metadata (text4gcn.graph, what apply_refinement uses), plot_training_history (text4gcn.models, writes the accuracy/loss plots), and the device functions get_device_manager, set_global_device, get_global_device, check_pytorch_installation and DeviceManager (text4gcn.utils, see Device Management).

Output and diagnostics¶

Text4GCN separates three kinds of output, as most current ML libraries do:

What	Where	Controlled by
Progress and summaries of each step	printed to stdout	`verbose=` of `TextPipeline`, the builders and `GNN`: `0` silent, `1` (default) progress bars and summaries, `2` no progress bars (see below)
Results	returned objects: `TrainingResult` (metrics, per-class report, confusion matrix, per-epoch history) and `GraphBuildStats`	—
Diagnostics: array shapes, model structure, file details	the standard `logging` module, logger `"text4gcn"`	`text4gcn.set_verbosity(...)` or your own logging configuration

verbose per class (any other value raises ValueError):

Class	`1` (default)	`2`
`TextPipeline`	One summary line per stage (no progress bars)	Same as `1`
`Frequency` / `Semantic` / `Syntactic`	Header, a progress bar during slow steps, summary line	Header and summary line, no progress bar
`GNN`	Progress bar over the epochs, then the summary	One line per epoch instead of the bar, then the summary

Diagnostics¶

The "text4gcn" logger has only a NullHandler, so it is quiet unless you ask for it:

import text4gcn

text4gcn.set_verbosity("debug")    # "debug", "info", "warning", "error", "critical"

set_verbosity also accepts a logging level number (e.g. logging.DEBUG); an unknown name raises ValueError.

set_verbosity sends the library's messages to stderr; the environment variable TEXT4GCN_VERBOSITY=debug does the same at import time. If your application already configures logging, the text4gcn loggers follow that configuration instead.