gugabobo/neuroadapter-subject01-l120

NeuroAdapter subject01: seed42 L120

Public inference-weight backup of the half-learning-rate L120 development candidate.

Identity and selection

• —Run: subject01-author-fp32-seed42-lr5e-5-from100.
• —Epoch 120; 63,840 cumulative optimizer updates; training seed 42.
• —First 100 epochs at AdamW LR 1e-4, then full-state continuation at 5e-5 through epoch 120.
• —Selected by the existing fixed monitor100 semantic score among the L120/L130/L140 candidates: S=84.656566. Full val500 S=82.819038.
• —S=(AlexNet-5 + Inception + CLIP identification)/3. Eight metrics and separate pool scores are in evaluation-summary.json; Eff and SwAV are lower-is-better.
• —This does not assert global superiority, independent-test performance, or that L120 is best on every metric. Original mainline best remains epoch110 and is not changed by this backup. Later L130/L140 full500 point estimates were slightly higher, without confirmed superiority in the reported comparisons.

Contents and loading

inference.pt contains image_proj, ip_adapter, and guidance_generator state dictionaries. parcel_vertices.pt and metadata.json preserve the exact subject-specific parcel mapping and model configuration. Keep these three files together.

import torch
weights = torch.load("inference.pt", map_location="cpu", weights_only=True)
# After constructing matching NeuroAdapter modules:
brain_adapter.image_proj_model.load_state_dict(weights["image_proj"], strict=True)
brain_adapter.adapter_modules.load_state_dict(weights["ip_adapter"], strict=True)
guidance_generator.load_state_dict(weights["guidance_generator"], strict=True)
parcels = torch.load("parcel_vertices.pt", map_location="cpu", weights_only=True)

Use the existing author-first decode_brain_adapter.py --export_path /path/to/inference.pt --no_selection and its required data/model arguments. This is not a standalone Diffusers pipeline: matching NeuroAdapter code, pretrained Stable Diffusion v1.5 components, and the original subject preprocessing are required. No base diffusion weights, participant brain measurements, stimulus images, optimizer states, RNG states, or credentials are included. Full training resumption requires the original complete checkpoint, not this inference export. Metadata retains original provenance paths; adapt paths to your environment rather than using them literally.

Decoding protocol: two fixed candidates per image, candidate batch1, image-ID-derived random streams, 50 denoising steps, CFG4, FP32 model weights with the existing TF32/backend and author VAE decode behavior; no candidate selection or GT initialization.

Provenance and limitations

Upstream: https://github.com/kriegeskorte-lab/NeuroAdapter (pinned 7d86d8c59dc13d8c14a04709e66e4b2de1b7d3ae). Author-first project: https://github.com/GeYugong/neuroadapter-author-first (private project; access may be required). The export's source commit and exact SHA256 values are recorded in metadata.json and manifest.json.

The 500-image pool is a repeatedly used internal development set; monitor100 is a subset with a different identification pool. Standard test was not used. Fixed-model brain/HL interventions indicate model input dependence, not human-brain causality or functional specificity. Reconstruction category errors remain. This repository is a backup, not a new training run or model promotion.

No new license grant is asserted here. Consult the original project and required pretrained component terms before redistribution or downstream use.