predict-SIREN-PSNR/COIN-collection

---
license: mit
---

# COIN collection dataset

This is the dataset for our paper ["Predicting the Encoding Error of SIRENs"](https://openreview.net/forum?id=iKPC7N85Pf). It consists of 300,000 small SIREN networks trained to encode images from the MSCOCO dataset.

We will publish a loading script for this dataset soon, but until then, see the following instructions:

## How to Use

First, download this repository using:

`huggingface-cli repo download predict-SIREN-PSNR/COIN-collection --repo_type datasets`

There are two types of files in this dataset:

1. `.json.gz` files containing data about the SIRENs we have trained,
2. the images from the MSCOCO dataset that those SIRENs are trained on.

### MSCOCO images
To download the MSCOCO images:

1. `pip install img2dataset`
2. `cd data/mscoco`
3. `bash download_mscoco.sh`

This will download around 80Gb of images in `data/mscoco/mscoco`.

### SIREN run records
The sirens are organized into two sub-datasets, `single-architecture` and `many-architecture`. Each `.json.gz` file contains one SIREN per line, which can be loaded as a JSON object. Each SIREN record contains the following fields:

- `config`: The starting configuration of the SIREN training run. contains the following subfields:
- `image_id` corresponds to the filename of the corresponding MSCOCO image, as downloaded using `download_mscoco.sh`. e.g. `image_id=123` corresponds to the filename `000000123.png`.
- `image_size` indicates what size the image was downsampled to, using PIL's `resize()` function with `BOX` resampling.
- The other arguments in `config` specify the arguments to be used in the [COIN training script](https://github.com/EmilienDupont/coin) to reproduce this SIREN run.
- `psnr_history`: record of the PSNR curve during training time. PSNR is recorded once every 10 training iterations.
- `best_psnr_history`: Similar to `psnr_history`, but stores the maximum value of `psnr history` seen up until this point during training.
- `iteration_history`: Parallel to the psnr_history and best_psnr_history; the training iteration at wich those PSNRs are recorded.
- `hp_bpp`: bits per pixel of the SIREN encoding of the image, with weights stored at half-precision (16-bit floats)
- `fp_bpp`: bits per pixel of the SIREN encoding of the image, with weights stored at full-precision (32-bit floats)
- `fp_psnr`: PSNR of the SIREN-based image reconstruction.
- `best_model`: Binary blob of the SIREN's `state_dict`.

原始信息汇总

COIN collection dataset 概述

数据集描述

• 名称: COIN collection dataset
• 目的: 用于论文 "Predicting the Encoding Error of SIRENs" 的研究。
• 内容: 包含300,000个小型SIREN网络，这些网络用于编码来自MSCOCO数据集的图像。

数据集组成

• 文件类型:
  1. .json.gz 文件：包含训练的SIREN网络数据。
  2. MSCOCO数据集的图像。

数据集使用指南

• 下载: 使用 huggingface-cli repo download predict-SIREN-PSNR/COIN-collection --repo_type datasets 命令下载数据集。
• MSCOCO图像下载:
  1. 安装 img2dataset。
  2. 进入 data/mscoco 目录。
  3. 运行 bash download_mscoco.sh 下载约80Gb的图像。

SIREN网络记录

• 组织方式: 分为 single-architecture 和 many-architecture 两个子数据集。
• 记录内容: 每个 .json.gz 文件包含一条SIREN记录，包含以下字段：
  • config: SIREN训练运行的初始配置，包括 image_id, image_size 和其他训练参数。
  • psnr_history: 训练期间的PSNR曲线记录。
  • best_psnr_history: 训练期间PSNR的最大值记录。
  • iteration_history: 与PSNR记录对应的训练迭代次数。
  • hp_bpp: 半精度（16位浮点数）下SIREN图像编码的每像素位数。
  • fp_bpp: 全精度（32位浮点数）下SIREN图像编码的每像素位数。
  • fp_psnr: SIREN图像重建的PSNR值。
  • best_model: SIREN模型的二进制状态字典。