Awesome

L3: Accelerator-Friendly Lossless Image Format for High-Resolution, High-Throughput DNN Training

This is the source code repository for L3, an accelerator-friendly lossless image format.

The source code consists of independent encoder, decoder module, header file, and patch file for NVIDIA DALI.

The current version is a prototype version, and many parts are not automated in various aspects:

• To encode the image, the data of R, G, and B channels of the raw image must be separately saved.
• Manually, the file to be encoded or decoded must be specified in the "encoder_main.cu" or "decoder_main.cu" file.
• Manually, the factor "N" must be defined to determine the number of patches in "l3.cuh".
• L3 format can encode and decode the RGB-formatted image only.

Prepare

Before the using L3 encoder, the user needs to prepare the raw data of R, G, and B channel of image from other image formats. For example, the user can use PIL and numpy package on Python3 by folloing commands:

>>> from PIL import Image
>>> import numpy as np

>>> im = Image.open("some_example_your_image.png")
>>> pixels = np.array(im)

# Save "R" channel data to file
>>> pix_r = pixels[:, :, 0]
>>> pix_r.tofile(open("pixel_r.dat", "wb"))

# Save "G" channel data to file
>>> pix_g = pixels[:, :, 1]
>>> pix_g.tofile(open("pixel_g.dat", "wb"))

# Save "B" channel data to file
>>> pix_b = pixels[:, :, 2]
>>> pix_b.tofile(open("pixel_b.dat", "wb"))

>>> im.close()

Encoding

Before the encoding is started, the user makes sure that the input files (raw data of R, G, and B) and output file path are predefined in encoder/encoder_main.cu, and factor N is defined in l3.cuh

$ cd src/
$ nvcc encoder/encoder.cu encoder/encoder_main.cu -o encoder_test
$ ./encoder_test

Decoding

Before the decoding is started, the user makes sure that the input files (L3-encoded data) is predefined in decoder/decoder_main.cu.

$ cd src/
$ nvcc decoder/decoder.cu decoder/decoder_main.cu -o decoder_test
$ ./decoder_test

L3 with NVIDIA DALI

We support Github patch file to use L3 decoder with DALI on version 1.1.0 (commit number: 25b99fa703e4971906321e9360e357d74975de6e).

To apply the patch file to DALI:

# Download the DALI version 1.1.0
$ git clone -b release_v1.1 https://github.com/NVIDIA/DALI.git
$ cd $DALI_HOME

# Test patch command
$ patch -p1 --dry-run < ${L3 directory}/src/nvidia-dali/l3-integrated-dali.patch

# Apply patch to DALI
$ patch -p1 < ${L3 directory}/src/nvidia-dali/l3-integrated-dali.patch

Citation

Please cite the following paper if you use L3:

L3: Accelerator-Friendly Lossless Image Format for High-Resolution, High-Throughput DNN Training. Jonghyun Bae, Woohyeon Baek, Tae Jun Ham, and Jae W. Lee. In European Conference on Computer Vision (ECCV), October 2022.

@inproceedings {XXXXXX,
  author = {Jonghyun Bae and Woohyeon Baek and Tae Jun Ham and Jae W. Lee},
  title = {L3: Accelerator-Friendly Lossless Image Format for High-Resolution, High-Throughput DNN Training},
  booktitle = {European Conference on Computer Vision ({ECCV} 22)},
  year = {2022},
  publisher = {European Computer Vision Association},
  month = oct,
  address = {Tel-Aviv, Israel},
}