Training DNNs with Hybrid Block Floating Point

Authors: Mario Drumond, Tao LIN, Martin Jaggi, Babak Falsafi

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We experiment with a set of popular image classification tasks with the CIFAR-100, SVHN, and Image Net datasets.
Researcher Affiliation Academia Mario Drumond Ecocloud EPFL mario.drumond@epfl.ch Tao Lin Ecocloud EPFL tao.lin@epfl.ch Martin Jaggi Ecocloud EPFL martin.jaggi@epfl.ch Babak Falsafi Ecocloud EPFL babak.falsafi@epfl.ch
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository for the described methodology.
Open Datasets Yes We experiment with a set of popular image classification tasks with the CIFAR-100 [19], SVHN [20], and Image Net [21] datasets. We also evaluate language modeling tasks with Penn Tree Bank(PTB) dataset [24].
Dataset Splits No For the image classification experiments, we report training loss and validation top-1 error. For the language modeling models, we report training loss and validation perplexity. While these datasets have standard splits, the paper doesn't explicitly state what splits were used for reproducibility.
Hardware Specification Yes We synthesize the accelerator on a Stratix V 5SGSD5 FPGA at a clock rate of 200MHz.
Software Dependencies No The paper mentions modifying Py Torch's layers but does not specify the version number of Py Torch or any other software dependencies with their versions.
Experiment Setup No The paper states, 'We trained all models using the same hyperparameters reported in their respective original papers,' which means the specific hyperparameter values are not directly provided within this paper.