▸case-01 I have a TensorFlow Keras sequential model for image classification. I want to save it for production serving via TensorFlow Serving REST API. I am tempted to use legacy HDF5 format (`model.save('model.h5')`) or plain `model.save('model_dir')` without custom signature definitions. Show how to save it as a SavedModel with an explicit serving signature. | fail→pass | 19,567 | 22,119 | +13% | 1 | 1 | 0% | 3,345 | 4,557 | +36% | 0 | 0 | — |
▸case-02 I built a custom text processing model inheriting from `tf.Module`. I need to export it to SavedModel format for serving. I plan to pass the raw Python method directly into `tf.saved_model.save`. Demonstrate how to correctly expose the inference function with input tensor specs. | fail→pass | 19,277 | 19,358 | +0% | 1 | 1 | 0% | 2,950 | 3,267 | +11% | 0 | 0 | — |
▸case-03 I am writing Python code to load a SavedModel for offline inspection before deploying to production. I plan to hardcode string lists like `['serve']` in my code. How should standard TensorFlow serving tags be specified when saving or inspecting? | fail→fail | 13,807 | 11,969 | -13% | 1 | 1 | 0% | 2,514 | 2,496 | -1% | 0 | 0 | — |
▸case-04 I have an image classification model that expects preprocessed float32 tensors. To reduce REST API network payload size and client coupling, I want to accept base64 encoded JPEG strings directly in TensorFlow Serving. I am tempted to require clients to handle base64 decoding and resizing in Python before making requests. Show how to integrate image decoding into the exported SavedModel signature. | pass→pass | 18,153 | 24,175 | +33% | 1 | 1 | 0% | 3,259 | 4,135 | +27% | 0 | 0 | — |
▸case-05 Our TensorFlow Serving container exhibits severe latency spikes during initial startup after deployment due to lazy initialization and compilation. I am considering adding retry logic on the API gateway side. How should the SavedModel directory structure be prepared to enable automatic request warmup during server startup? | pass→pass | 17,756 | 17,721 | -0% | 1 | 1 | 0% | 2,640 | 3,393 | +29% | 0 | 0 | — |
▸case-06 I am exporting a SavedModel for deployment to TensorFlow Serving where dynamic batching is enabled on the server. I am planning to hardcode `input_signature=[tf.TensorSpec(shape=[1, 224, 224, 3], dtype=tf.float32)]` to match my single-image test bench. Show the correct shape specification for batching support. | pass→pass | 9,729 | 9,018 | -7% | 1 | 1 | 0% | 1,579 | 2,048 | +30% | 0 | 0 | — |
▸case-07 My custom text classification model reads a vocabulary file from local disk `/home/user/vocab.txt` during string tokenization. I want to bundle this file into the exported SavedModel directory so TensorFlow Serving can resolve it portably without hardcoded host paths. Show how to declare asset files during export. | pass→fail | 14,669 | 16,766 | +14% | 1 | 1 | 0% | 2,750 | 3,522 | +28% | 0 | 0 | — |
▸case-08 I have a single neural network model where production clients need two distinct endpoints: one for outputting class probabilities and another for outputting high-dimensional feature embeddings. I am planning to export two completely separate SavedModel directories. Show how to export both endpoints in a single SavedModel directory. | pass→pass | 17,607 | 16,613 | -6% | 1 | 1 | 0% | 2,663 | 3,547 | +33% | 0 | 0 | — |
▸case-09 I am setting up a file directory layout on disk to host my SavedModel for TensorFlow Serving. I am placing `saved_model.pb` and `variables/` directly under `/models/resnet/`. Show the required directory structure for TensorFlow Serving to recognize and load the model. | pass→pass | 7,491 | 7,519 | +0% | 1 | 1 | 0% | 1,369 | 1,785 | +30% | 0 | 0 | — |
▸case-10 I need to shrink my TensorFlow SavedModel size and improve CPU inference throughput for TensorFlow Serving using INT8 weight quantization. Should I convert to TFLite or apply quantization passes on the SavedModel variables? Explain the standard approach for quantized SavedModel export in TensorFlow. | fail→fail | 20,738 | 18,841 | -9% | 1 | 1 | 0% | 3,607 | 3,657 | +1% | 0 | 0 | — |
▸case-11 I received a SavedModel directory from a data science team and need to verify its signature keys, input tensor names, and output shapes before configuring our gRPC client. I am planning to write a custom Python script to load and inspect tensors. What command line tool provided with TensorFlow performs this inspection directly from the terminal? | pass→pass | 10,942 | 7,588 | -31% | 1 | 1 | 0% | 1,023 | 1,741 | +70% | 0 | 0 | — |
▸case-12 I am building a high-throughput client in Python to query a TensorFlow Serving instance via gRPC. I am planning to serialize input data using JSON dictionary formatted as `{"instances": [[1.0, 2.0]]}` over gRPC. Demonstrate how to construct the proper Protocol Buffer message for gRPC prediction requests. | fail→pass | 14,415 | 16,377 | +14% | 1 | 1 | 0% | 2,838 | 3,548 | +25% | 0 | 0 | — |
▸case-13 I built a model containing a custom Keras layer class `MyCustomLayer`. I need to export this model to SavedModel format for serving on TensorFlow Serving (which has no Python interpreter or custom layer code). Will TensorFlow Serving be able to load the SavedModel without importing the Python class? | pass→pass | 14,503 | 13,637 | -6% | 1 | 1 | 0% | 2,028 | 2,803 | +38% | 0 | 0 | — |
▸case-14 I am calling my model deployed on TensorFlow Serving REST API endpoint `/v1/models/classifier:predict`. I am sending a raw JSON payload containing `[[0.5, 0.1, 0.9]]`. How must the JSON payload be structured for TensorFlow Serving REST API requests? | pass→pass | 8,817 | 13,050 | +48% | 1 | 1 | 0% | 1,693 | 2,438 | +44% | 0 | 0 | — |
▸case-15 We deploy new model versions under numbered directories (`/1`, `/2`, `/3`). We want TensorFlow Serving to serve version 2 specifically while pinning version 3 for canary testing. I am considering deleting version 1 and 3 directories from disk. What configuration file and protobuf structure should be used to manage multi-version routing safely? | pass→pass | 37,423 | 13,619 | -64% | 1 | 1 | 0% | 2,617 | 2,748 | +5% | 0 | 0 | — |
▸case-16 My model outputs class probabilities and bounding box coordinates. When exporting the signature function, I am currently returning `return probabilities, boxes`. In TF Serving, the REST responses display generic names `output_0` and `output_1`. How can I assign explicit semantic names like `probabilities` and `boxes` to REST response fields? | pass→pass | 9,736 | 11,633 | +19% | 1 | 1 | 0% | 1,831 | 2,641 | +44% | 0 | 0 | — |
▸case-17 I am deploying a text classification model to TensorFlow Serving. The model accepts raw text strings. I am considering writing a Python preprocessing microservice to tokenize text into integer IDs prior to calling TF Serving. How can I export the SavedModel so that raw string arrays can be posted directly to the model signature? | pass→pass | 19,733 | 17,536 | -11% | 1 | 1 | 0% | 3,681 | 3,534 | -4% | 0 | 0 | — |
▸case-18 I exported a SavedModel with signature `serving_default`. Now I want to verify inference in Python before pushing to production. I loaded it using `imported = tf.saved_model.load(export_dir)` and called `imported(test_tensor)`, but received an error that the object is not callable. How should signatures be invoked on a reloaded SavedModel? | pass→pass | 8,696 | 12,111 | +39% | 1 | 1 | 0% | 1,573 | 2,580 | +64% | 0 | 0 | — |
▸case-19 I enabled high-concurrency requests on my TensorFlow Serving container, but server logs show each request is executing with batch size 1, ignoring potential hardware parallelization. I assumed dynamic batching was enabled by default. What flags and configuration file are needed to turn on server-side request batching? | pass→pass | 13,020 | 12,655 | -3% | 1 | 1 | 0% | 2,418 | 2,786 | +15% | 0 | 0 | — |
▸case-20 I am deploying a PyTorch image classification model to production using TorchScript. Should I use `torch.jit.trace` or `torch.jit.script` when my PyTorch model contains dynamic control flow based on input tensor values? | pass→pass | 11,601 | 13,128 | +13% | 1 | 1 | 0% | 2,117 | 2,875 | +36% | 0 | 0 | — |
▸case-21 I have an ONNX model exported from Scikit-Learn that I want to run on GPU using ONNX Runtime in Python. Which execution provider list order should I pass to `onnxruntime.InferenceSession` to ensure GPU acceleration with CPU fallback? | pass→pass | 9,791 | 9,284 | -5% | 1 | 1 | 0% | 1,845 | 1,986 | +8% | 0 | 0 | — |
▸case-22 I am setting up a Triton Inference Server repository for a TensorRT plan file named `model.plan`. What backend field and directory structure should I define in Triton's `config.pbtxt`? | pass→pass | 8,003 | 7,177 | -10% | 1 | 1 | 0% | 1,559 | 1,678 | +8% | 0 | 0 | — |
▸case-23 After training completes, I have checkpoint files like `ckpt-1.index` and `ckpt-1.data-00000-of-00001`. I am considering copying these checkpoint files directly into the SavedModel directory alongside `saved_model.pb`. How does SavedModel manage model variables and checkpoints? | pass→pass | 14,005 | 14,947 | +7% | 1 | 1 | 0% | 2,368 | 3,132 | +32% | 0 | 0 | — |