TA-SAE Model Card

This repository contains the trained Temporal-Aware Sparse AutoEncoder (TA-SAE) models for different layers.

Model Description

TA-SAE is a specialized autoencoder model designed for temporal feature extraction and compression. Each layer model represents a different level of feature abstraction in the network.

Usage

Installation

pip install huggingface_hub

Loading Models

Download a specific file:

from huggingface_hub import hf_hub_download

# Download specific layer model
file_path = hf_hub_download(
    repo_id="jeix/TA-SAE",
    filename="PixArt/SAE-Layer0/model.safetensors"
)

Download all files for a specific layer:

from huggingface_hub import snapshot_download

# Download all files for layer0
local_dir = snapshot_download(
    repo_id="jeix/TA-SAE",
    repo_type="model",
    allow_patterns="PixArt/SAE-Layer0/*"
)

Download all layers:

local_dir = snapshot_download(
    repo_id="jeix/TA-SAE",
    repo_type="model",
    allow_patterns="PixArt/SAE-Layer*/*"
)

Using Command Line

Install CLI tool

pip install -U huggingface_hub

Download specific file

huggingface-cli download jeix/TA-SAE --local-dir ./download --include "PixArt/SAE-Layer0/model.safetensors"

Model Files Description

Each layer directory contains the following files:

• model.safetensors: The main model weights
• optimizer.bin: Optimizer state
• scheduler.bin: Learning rate scheduler state
• random_states_0.pkl: Random state information
• scaler.pt: Data scaling parameters