---
license: other
task_categories:
- image-feature-extraction
tags:
- imagenet
- pae
- dinov2
- latent-cache
- safetensors
pretty_name: PDM PAE DINOv2-L d32 ImageNet-256 Train Latent Cache
size_categories:
- 1M<n<10M
---

# PDM PAE DINOv2-L d32 ImageNet-256 Train Latent Cache

This repository contains the PAE latent cache produced for `/workspace/PDM`.

## Contents

- Cache type: **PAE DINOv2-L d32 latent cache**
- Source split: **ImageNet-1k train** after 256x256 ADM-style cropping
- Local source path during creation: `/workspace/PDM/data/pae_latents/PAE_DINOv2L_d32/imagenet256_train_full`
- Number of shards: **313**
- Total samples: **1,281,167**
- Total size: **41.992 GB** decimal
- Latent tensor shape per sample: **[32, 16, 16]**
- File format: **safetensors**
- Shard keys: **labels**, **latents**, **latents_flip**

Data files are under:

```text
pae_latents/imagenet256_train_full/latents_rank00_shard*.safetensors
```

Additional metadata:

- `cache_summary.json`: summary generated before upload
- `manifest.jsonl`: one JSON record per shard with size/sample/shape metadata

## Provenance

- PAE repo commit: `51f8fa6290e6c1f4ffb9bb9aea500bfb707ffd12`
- PAE config: `PAE_DINOv2L_d32.yaml`
- Expected total samples from local full-cache scan: `1,281,167`
- Local full-cache scan status before upload: `ready_for_full_train = true`
- Local smoke4096 status before upload: `practical_gate_pass = true`

## Loading example

```python
from safetensors import safe_open

path = "pae_latents/imagenet256_train_full/latents_rank00_shard000000.safetensors"
with safe_open(path, framework="pt", device="cpu") as f:
    labels = f.get_tensor("labels")
    latents = f.get_tensor("latents")
    latents_flip = f.get_tensor("latents_flip")
    print(labels.shape, latents.shape, latents_flip.shape)
```

## License / upstream data note

This is a derived latent cache from ImageNet-1k training images. Use of this cache should comply with the upstream ImageNet terms and any applicable model/checkpoint terms. This repository does not include the raw ImageNet images.