This runnable package trains a small PyTorch image classifier and exports a
deployable ONNX model plus a resumable .pt checkpoint. Its platform runtime
is pytorch-2.5-cu124, which is provided by the training runtime image; the
ZIP does not install dependencies.
For platform_dataset, choose an image classification dataset and annotation
project in the UI. Each annotation must contain one class ID, and its class
order is used unchanged by the ONNX model.
For script_managed, set class names to cat,dog. The package downloads the
public CIFAR-10 archive with torchvision, selects its cat and dog samples, and
trains without a platform dataset or annotation project. This requires the
platform to enable both allow_script_managed_data: true and outbound access
to the CIFAR-10 download host. The downloaded data exists only in the current
task workspace and is removed after the run.
Build the uploadable ZIP:
python build_package_zip.py ./dist/pytorch-image-classifier-1.0.0.zipImport that ZIP from 模型管理 → 自定义模型, then choose 自定义脚本训练 when creating a training task. For a fast smoke test, use:
{"epochs": 1, "max_samples": 128, "batch_size": 32}To start from a pre-trained .pt weight, include it in the same ZIP (for
example weights/initial.pt) and load it from train.py using a path relative
to __file__. The example checks that exact optional path automatically; do
not upload it as a separate model package or add a manifest field.