Blame for README.md

---
library_name: pytorch
tags:
- vision-transformer
- cifar-100
- imagenet-200
- in-synapse-activation
---

# ISA Vision checkpoints

Model-only PyTorch checkpoints for CIFAR-100 and ImageNet-200. The collection
contains **18 checkpoints**: two datasets, three model scales (tiny, mid and
small), and three variants:

- `standard`: the Digital baseline with ordinary digital FFNs;
- `hybrid`: a physical Flash-transistor up projection and a digital down projection;
- `physical`: physical Flash-transistor up and down projections.

Source code and experiment configurations: [ISA on GitHub](https://github.com/Vigilanceee/ISA-).
Related collections: [language Transformers](https://huggingface.co/Viligance/ISA-Language)
and [five-device MLP/VGG8 studies](https://huggingface.co/Viligance/ISA-Device-Checkpoints).

## Layout

```text
{cifar100,imagenet200}/{standard,hybrid,physical}/{tiny,mid,small}/model.pt
```

Digital baselines additionally have a matching `config.json` and
`artifact_validation.json`. Vision embedding dimensions are 192/256/384 for
tiny/mid/small. Use the actual checkpoint configuration when constructing a model.

## Load a Digital baseline

Install the ISA source package, then use the stored configuration and strict loading:

```python
import torch
from huggingface_hub import HfApi, hf_hub_download
from isa.vision.config import ModelConfig
from isa.vision.models import StandardViT

repo_id = "Viligance/ISA-Vision"
revision = HfApi().model_info(repo_id).sha
path = hf_hub_download(repo_id, "cifar100/standard/tiny/model.pt", revision=revision)
checkpoint = torch.load(path, map_location="cpu", weights_only=True)
model = StandardViT(ModelConfig(**checkpoint["model_config"]))
model.load_state_dict(checkpoint["model"], strict=True)
model.eval()
print("Hugging Face revision:", revision)
```

For repeatable downloads, retain that immutable revision and pass it to subsequent
`hf_hub_download` calls. Hybrid and Physical checkpoints also use the `model`
state-dictionary key; construct the corresponding ISA model and experiment
configuration before strictly loading it. Their threshold and TIA parameters are
already stored in the tensors.

## Physical device equation

For Hybrid and Physical projections, the fitted Flash-transistor operator evaluates

\[
I(V,V_{th}) = I_S\frac{\operatorname{softplus}^2\!\left(\frac{V-V_{th}}{2nU_T}\right)-\operatorname{softplus}^2\!\left(\frac{V-V_{th}-V_D}{2nU_T}\right)}{1+V/V_{sat}}.
\]

This device equation does not apply to the Digital baseline FFNs.

## Digital baseline records

| Dataset | Tiny | Mid | Small |
|---|---:|---:|---:|
| CIFAR-100 | 67.98% | 69.45% | 69.30% |
| ImageNet-200 | 72.43% | 72.57% | 73.64% |

These are recorded training-validation/checkpoint-selection accuracies, not new
full-dataset evaluations performed during publication. The model-only baseline
artifacts preserve all source model tensors exactly and pass finite-value checks,
strict loading, and a CPU forward check. Optimizer and scheduler state are not
required for inference and are not part of these model-only artifacts.

`checkpoint_manifest.json` covers all 18 weights with file sizes, SHA256 digests
and validation metadata. `baseline_manifest.json` records the six Digital
baselines and their configurations. `release_validation.json` distinguishes
physical-parameter verification from Digital baseline artifact checks.