eshanli/media-agentic-ai-fp16

MedIA-Agentic-AI, FP16 optimized

Half-precision (FP16) copies of the nnU-Net checkpoints from Xiang-mira/MedIA-Agentic-AI.

What was done

For each model the network weights were cast from FP32 to FP16, and the training-only optimizer and grad-scaler state were dropped. The result is a smaller, inference-ready checkpoint. The per-model plans.json, dataset.json, label_mapping.json, and inference_config.yaml are copied alongside each weight so every folder is a drop-in replacement.

Layout

Each folder contains:

• checkpoint_final_fp16.pth — the FP16 network weights
• the original metadata files (plans.json, dataset.json, label_mapping.json, inference_config.yaml)

How to load

import torch
ckpt = torch.load("checkpoint_final_fp16.pth", map_location="cpu", weights_only=False)
weights = ckpt["network_weights"]  # already FP16
# load into the matching nnU-Net PlainConvUNet / ResEncUNet from plans.json,
# then model.half() for FP16 inference, or cast weights back with .float() for FP32.

Masks are numerically equivalent to the FP32 originals for inference; FP16 mainly reduces file size and inference memory. See conversion_summary.json for the exact per-model size before and after.