How DeepInsight Transforms Deep Learning Model Debugging with Real-Time Visualization

Background

Deep neural networks have revolutionized machine‑learning research and applications, yet they remain black boxes that are difficult to interpret. Visualization helps humans understand and optimize these models, especially in domains such as advertising where existing tools are scarce.

DeepInsight System Architecture

DeepInsight is a distributed micro‑service platform for deep‑learning visualization, consisting of a front‑end web UI, back‑end services, and deep‑learning components. It supports TensorFlowRS and native TensorFlow training lifecycle management, exposing raw training data for debugging and improving model interpretability.

Multi‑Dimensional Visualization Based on Data Exposure

Deep learning components (e.g., TF‑Tracer, TF‑Profiler) expose raw data from the training graph. Back‑end services and the web UI provide online/offline interactive visual analysis, forming an ecosystem that spans the entire model lifecycle.

Deep Learning Components

TF‑Tracer is a plug‑in built on TensorFlow’s graph API that can trace all variables, filter them with regular expressions, and output data in NumPy or binary formats. Configuration options include enabling/disabling components without affecting training performance, specifying sampling intervals (every_steps, every_secs, step_range), and controlling output destinations (HDFS, ODPS, Logview, TensorBoard+).

"trace": "true", // enable TF‑Tracer
    "trace_config": {
        "graphkeys": "TRAINABLE_VARIABLES,<custom Graph Key>",
        "scopes": {"TRAINABLE_VARIABLES": "^layer3.*"},
        "every_steps": 5000,
        "every_secs": 60,
        "step_range": {"start": 5000, "end": 5010},
        "save_config": {
            "dump_dir": "hdfs://ns1/data/xxx/tftracer/trace_output_dir",
            "data_format": "csv",
            "limit_size": 90000
        },
        "chief_only": "false",
        "logview_level": "detail",
        "at_begin": "true",
        "at_begin_config": {
            "graphkeys": "default_summary_collection",
            "scopes": {"default_summary_collection": ".*Relu.*"},
            "limit_size": 100000000
        }
    }

TF‑Tracer also supports dynamic updates of the exposed dataset without restarting training, allowing users to modify the variable list on‑the‑fly and compare outputs from different workers.

Backend Docker Micro‑services

Notebook+ (interactive analysis) and TensorBoard+ (real‑time visualization) run as Docker containers orchestrated by the front‑end. Containers are isolated per user, support automatic resource release, and enforce per‑user access control.

Front‑End Web Platform

The UI offers cluster management, dashboards, configuration, permission, and data management. Lifecycle management covers model development, distributed training, component integration, log viewing, and visual analytics.

External Visualization Services

DeepInsight provides micro‑service images and data APIs to third‑party platforms such as PAI, XDL, and Lotus, enabling them to consume TensorBoard+ and Notebook+ capabilities.

Conclusion

Future work will extend visualization dimensions to full model‑level visual analytics, linking features to weights and supporting Facets for dataset exploration, thereby improving model interpretability and debugging efficiency.