---

name: model-inference description: "Select, prepare, and run verified Gluon or PyTorch/pytorchcv image-classification models with safe CPU defaults, strict local checkpoints, ImageNet preprocessing, and output checks." disable-model-invocation: true metadata: disco-role: operating license: MIT

Model inference

Use this skill for one-batch image-classification inference, model selection, local checkpoint loading, input preprocessing, or lightweight model statistics through the Gluon or PyTorch/pytorchcv providers.

Safe default

Start from the sub-skill directory with a non-pretrained resnet18 CPU smoke:

python scripts/infer_gluon.py --model resnet18
python scripts/infer_pytorch.py --model resnet18

Both scripts construct random weights by passing pretrained=False, use one normalized (1, 3, 224, 224) input, and print the output shape, parameter count, and zero-based top-k class indices/probabilities. If --image is omitted, the input is a zero-valued RGB image before normalization, so the result is only a shape/device smoke and has no accuracy meaning.

The scripts never request or download pretrained weights. A checkpoint is accepted only as an existing local file; use --checkpoint FILE with the same provider, model, class count, and input-channel count used to create it. The PyTorch script also accepts --remove-module for a checkpoint saved under torch.nn.DataParallel.

Route by intent

1. Select a provider-supported model name. The providers lowercase the name and raise ValueError for an unsupported model; they do not infer aliases.
2. Keep pretrained=False for smoke tests. If using --checkpoint, keep the model and --classes aligned with the checkpoint and let strict loading expose missing, extra, or shape-incompatible parameters.
3. Apply the ImageNet RGB preprocessing in checkpoints-and-inputs.md.
4. Assert that the input is NCHW and the output is rank-2 with batch size one before interpreting top-k results.
5. Use api-reference.md for provider and local loading contracts. Use troubleshooting.md for recovery.

Do not turn this route into training, resume, or dataset evaluation. Send those requests to training-evaluation. Send cross-framework parameter conversion to conversion. Send TensorFlow, Keras, and Chainer requests to framework-compatibility.