---

name: training description: "Operate CUDA-based train.py adaptation runs for Medical SAM Adapter across SAM, EfficientSAM, MobileSAM, Adapter, LoRA, AdaLoRA, 2D, 3D, and multi-class configurations." metadata: disco-role: operating disable-model-invocation: true license: GPL 3.0

Training

Use this route to construct and diagnose the repository's train.py adaptation workflow. It is a self-contained operating guide, not a copied launcher: it never downloads weights or data and never requires the source checkout at runtime.

Route gate

1. If the question is about dataset names, file layouts, sample dictionaries, prompts, tensor shapes, or whether a checkpoint is a base model, start with data preparation.

2. Read the exact parser reference. In particular, inspect the effective values of the type=bool options instead of trusting their spelling.

3. Select a network/encoder pair from model variants, then run the static helper before importing model code. It supports --help, --list, and an object or file supplied through --json:

   python scripts/inspect_model_registry.py --net sam --encoder vit_b --mod sam_adpt
   python scripts/inspect_model_registry.py --json '{"net":"mobile_sam","encoder":"tiny_vit","mod":"sam_lora"}' --output json
   

4. Check that -sam_ckpt is an existing local base-model checkpoint for the selected builder. A -weights file is a different training-wrapper checkpoint; do not substitute one for the other.

5. Confirm a CUDA-capable PyTorch environment, the visible device, data contract, output destination, positive -vis, and a unique -exp_name. Use the workflow patterns without adding any download step.

What the route covers

• Original SAM registry keys are default, vit_b, vit_h, and vit_l.
• EfficientSAM keys are default, vit_s, and vit_t.
• The MobileSAM package registry has additional image-encoder and decoder entries; only full model-shaped entries are suitable for ordinary train.py calls. The compatibility warnings are in model variants.
• sam_adpt, sam_lora, and sam_adalora change image-encoder construction and/or trainable parameters. Their behavior is network-specific; do not assume that an AdaLoRA or LoRA mode creates LoRA layers for every MobileSAM entry. See adapter mechanics.
• -thd True is the source's 3D slice/chunk path. -chunk controls the training crop/depth path, -evl_chunk controls validation windows, and -num_sample/-roi_size affect MONAI sampling. Dataset ownership remains with data preparation.
• Original SAM is the only training branch that honors -multimask_output > 1; EfficientSAM and MobileSAM pass multimask_output=False in the training loop. Independent scores belong to evaluation, while detector/box inference belongs to mobile inference.

Hard execution boundary

The actual training path creates cuda:<gpu_device> tensors, initializes CUDA AMP, and moves the network and data to CUDA. -gpu False does not make it CPU safe. CPU parser/help checks are diagnostics only, not successful training. -distributed is an opt-in DataParallel path and still needs all listed CUDA devices. Do not claim a run succeeded without a real CUDA execution.

Handoff and shared routing

Preserve the exact command, effective parser values, network/encoder, mode, checkpoint provenance, data root, device, and output paths in any handoff. For root-level routing and shared failures, use the Medical SAM Adapter root skill and its shared troubleshooting when present. For input preparation use data preparation; for checkpoint scoring use evaluation; for the separate object-aware route use mobile inference.

Read the detailed CLI reference, model variants, adapter mechanics, workflow patterns, and troubleshooting before a long run.