predict-SIREN-PSNR/COIN-collection

COIN collection dataset

This is the dataset for our paper "Predicting the Encoding Error of SIRENs". 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 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.