TFRecord Viewer

This tool decodes TFRecord binaries by parsing each record as a TensorFlow Example and exposing feature names, types, and sample values. It targets ML engineers, data scientists, and data engineers who need a fast, reliable view into serialized data used for training, evaluation, or model serving. It supports quick validation of feature presence and basic value checks.
Conceptually, the tool reads records in sequence, decodes features, and builds a schema representation. It recognizes common feature types (int64, float, bytes) and can infer types when a schema is not supplied. A per-record preview surfaces anomalies, missing features, and outlier values, enabling early quality gates before heavy preprocessing.
Unique value comes from exporting decoded results and schema summaries to JSON or CSV, facilitating reproducible data validation and pipeline testing. It helps identify skew, inconsistent feature naming, or unexpected data types across large datasets, reducing debugging time and enabling targeted data cleaning and standardization.
Core features include file-based input, optional schema enforcement, per-record decoding, aggregated schema, sample previews, and export options. It supports multiple files, records selection, and batch processing with memory-aware limits. This combination makes it suitable for data validation stages in ML workflows and for audits of data lineage.
Limitations include potential schema ambiguity for complex nested features and performance constraints on very large TFRecord files. The tool emphasizes read-only inspection and does not modify data. It assumes TFRecord format and TensorFlow encoding conventions and provides warnings when records cannot be decoded.