Model Metadata Validation

Model Metadata Validation

Summary

Validate that a TensorFlow Serving endpoint correctly exposes model metadata including input and output layer names. This skill confirms that a deployed neural network model's schema matches expected specifications before downstream inference.

When to use

After deploying a TensorFlow Serving container (especially within a Dockerized stack like NP-Classifier), before running classification or inference pipelines, to confirm that input layers are named 'input_2048' and 'input_4096' and output layer is named 'output'. Use this skill to catch layer name mismatches early, which would otherwise cause API calls to fail.

When NOT to use

• Model has not yet been containerized or TensorFlow Serving is not running.
• Layer names are custom or project-specific and do not follow the NP-Classifier convention (input_2048, input_4096, output).
• Only offline metadata inspection is needed without testing the live endpoint.

Inputs

• Running TensorFlow Serving container (via docker-compose)
• HTTP endpoint URL: /model/metadata

Outputs

• JSON metadata response object (containing input/output layer schema)
• Persisted metadata JSON file (for audit trail)

How to apply

Start the Dockerized NP-Classifier server via docker-compose or make server-compose. Send an HTTP GET request to the /model/metadata endpoint on the TensorFlow Serving container. Parse the returned JSON response and extract the input and output layer name fields. Validate that the two input layers are exactly named 'input_2048' and 'input_4096', and the single output layer is named 'output'. Record the complete metadata response to a JSON file for audit and debugging. If layer names do not match, the model configuration in the served artifact must be corrected before the inference API can be reliably used.

Related tools

• docker (Container runtime for launching the TensorFlow Serving instance)
• docker-compose (Orchestration tool to bring up the complete NP-Classifier stack (TensorFlow Serving + nginx) in one command)
• TensorFlow Serving (REST/gRPC server exposing the /model/metadata endpoint and serving inference requests) — https://github.com/tensorflow/serving

Examples

curl -X GET http://localhost:8501/v1/models/npc/metadata | jq '.inputs[].name, .outputs[].name'

Evaluation signals

• HTTP 200 response received from /model/metadata endpoint.
• Parsed JSON contains 'inputs' array with exactly two entries named 'input_2048' and 'input_4096'.
• Parsed JSON contains 'outputs' array with exactly one entry named 'output'.
• Persisted metadata JSON file is valid JSON and contains all layer definitions.
• Layer names match expected schema; if they differ, the validation fails and model artifact must be redeployed.

Limitations

• Metadata validation only confirms layer names; it does not verify layer shapes, data types, or functional correctness of the model.
• Requires TensorFlow Serving to be running; intermittent container failures will cause the skill to fail.
• Layer names are hard-coded expectations specific to NP-Classifier; other models with different layer naming schemes require this skill to be re-parameterized.
• The /model/metadata endpoint availability depends on TensorFlow Serving configuration; some deployments may disable it for security reasons.

Evidence

• [other] Does the TensorFlow Serving endpoint at /model/metadata successfully return model metadata including the correct input and output layer names when queried on the running Dockerized NP-Classifier server?: "Does the TensorFlow Serving endpoint at /model/metadata successfully return model metadata including the correct input and output layer names when queried on the running Dockerized NP-Classifier"
• [other] The expected model layer names that should be returned by the /model/metadata endpoint are input layers 'input_2048' and 'input_4096', and output layer 'output'.: "The expected model layer names that should be returned by the /model/metadata endpoint are input layers 'input_2048' and 'input_4096', and output layer 'output'."
• [other] 1. Start the Dockerized NP-Classifier server using docker-compose. 2. Send an HTTP GET request to the /model/metadata endpoint on the running TensorFlow Serving container. 3. Parse the JSON response and extract input/output layer name fields. 4. Validate that input layers are named 'input_2048' and 'input_4096' and output layer is named 'output'. 5. Write the complete metadata response to a JSON file.: "1. Start the Dockerized NP-Classifier server using docker-compose. 2. Send an HTTP GET request to the /model/metadata endpoint on the running TensorFlow Serving container. 3. Parse the JSON response"
• [readme] We pass through tensorflow serving at this url: /model/metadata: "We pass through tensorflow serving at this url: /model/metadata"
• [readme] Input layers' names should be "input_2048" and "input_4096". Output layer's name should be "output": "Input layers' names should be "input_2048" and "input_4096". Output layer's name should be "output""
• [readme] To bring everything up, you need docker and docker-compose: "To bring everything up, you need docker and docker-compose"