ParoleLM/socialjax-harvest-tokenizer-512

SocialJax Harvest 512-code tokenizer

Tokenizer used to produce ParoleLM/socialjax-harvest-frame-aligned-512.

• Repository: ParoleLM/socialjax-harvest-tokenizer-512
• Training step: 5,000
• Input: categorical SocialJax grids [B,T,16,22]
• Window: 16 frames
• Patch size: 2x2
• Codes per frame: 88
• Codebook size: 512
• Latent dimension: 32
• Exported inference checkpoint size: 270.7 MB

tokenizer.pt contains the exact model_state_dict used for the published dataset, without optimizer state. config.json and the relevant source files are included for reproducible loading.

import json
import torch

from src.architectures.genie.tokenizers.video_tokeniser import VideoTokenizer

config = json.load(open("config.json"))
model = VideoTokenizer(**config["model"], window=config["data"]["window"])
checkpoint = torch.load("tokenizer.pt", map_location="cpu")
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()