Lossy Compression for Lossless Prediction

Authors: Yann Dubois, Benjamin Bloem-Reddy, Karen Ullrich, Chris J. Maddison

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Using these objectives, we train a generic image compressor that achieves substantial rate savings (more than 1000 on Image Net) compared to JPEG on 8 datasets, without decreasing downstream classification performance. We evaluated our framework focusing on two questions: (i) What compression rates can our framework achieve at what cost? (ii) Can we train a general purpose predictive image compressor?
Researcher Affiliation Collaboration Yann Dubois Vector Institute yanndubois96@gmail.com Benjamin Bloem-Reddy The University of British Columbia benbr@stat.ubc.ca Karen Ullrich Facebook AI Research karenu@fb.com Chris J. Maddison University of Toronto Vector Institute cmaddis@cs.toronto.edu
Pseudocode Yes Algorithm 1 BINCE s forward pass for x
Open Source Code Yes Code is at github.com/Yann Dubs/lossyless. The code to train our main compressor is in Appx. E.7, the code to replicate all our results is at anonymous.
Open Datasets Yes Using these objectives, we train a generic image compressor that achieves substantial rate savings (more than 1000 on Image Net) compared to JPEG on 8 datasets... We compressed samples from a 2D banana source distribution [36], assuming rotation invariant tasks... We compressed the STL10 dataset [37]... on the small MSCOCO dataset [43].
Dataset Splits Yes We optimized hyper-parameters on validation using random search.
Hardware Specification No The paper mentions training on 'a single GPU' but does not specify the model, manufacturer, or other detailed specifications of the hardware used for experiments.
Software Dependencies No The paper mentions using specific models like 'Res Net18' and 'Ballé et al. s [28] hyperprior' but does not list programming languages (e.g., Python) or library versions (e.g., PyTorch 1.x, TensorFlow 2.x) with version numbers.
Experiment Setup Yes For all experiments, we train the compressors, freeze them, train the downstream predictors, and finally evaluate both on a test set. For classical compressors, standard neural compressors (VC) and our VIC, we used either reconstructions X as inputs to the predictors or representations Z. As BINCE does not provide reconstructions, we predicted from the compressed Z using a multi-layer perceptron (MLP). We used Res Net18 [35] for encoders and image predictors. For entropy models we used Ballé et al. s [28] hyperprior, which uses uniform quantization. We optimized hyper-parameters on validation using random search.