Faster Neural Networks Straight from JPEG
Authors: Lionel Gueguen, Alex Sergeev, Ben Kadlec, Rosanne Liu, Jason Yosinski
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We modify libjpeg to produce DCT coefficients directly, modify a Res Net-50 network to accommodate the differently sized and strided input, and evaluate performance on Image Net. We find networks that are both faster and more accurate, as well as networks with about the same accuracy but 1.77x faster than Res Net-50. Experiments described in Section 3 and 4 are conducted with the Keras framework and Tensor Flow backend. Training is performed on the Image Net dataset [4] with the standard Res Net-50 stepwise decreasing learning rates described in [10]. |
| Researcher Affiliation | Industry | Lionel Gueguen1 Alex Sergeev1 Ben Kadlec1 Rosanne Liu2 Jason Yosinski2 1Uber 2Uber AI Labs |
| Pseudocode | No | The paper provides architectural diagrams (e.g., Fig. 3, Fig. S1) but does not include any pseudocode or algorithm blocks. |
| Open Source Code | Yes | So others may reproduce experiments and benefit from speed increases found in this paper, we release our code at https://github.com/uber-research/jpeg2dct. |
| Open Datasets | Yes | Training is performed on the Image Net dataset [4] with the standard Res Net-50 stepwise decreasing learning rates described in [10]. |
| Dataset Splits | No | The paper mentions 'validation top-1 and top-5 error rates' and 'Training is performed on the Image Net dataset [4] with the standard Res Net-50 stepwise decreasing learning rates described in [10]', implying the use of a validation set and standard splits. However, it does not explicitly state the specific train/validation/test dataset splits (e.g., percentages or counts) used for the experiments. |
| Hardware Specification | Yes | The distributed training framework Horovod [21] is employed to facilitate parallel training over 128 NVIDIA Pascal GPUs. |
| Software Dependencies | No | The paper mentions using 'Keras framework and Tensor Flow backend' and 'Horovod' but does not specify their version numbers or any other software dependencies with versions. |
| Experiment Setup | Yes | Training is performed on the Image Net dataset [4] with the standard Res Net-50 stepwise decreasing learning rates described in [10]. The distributed training framework Horovod [21] is employed to facilitate parallel training over 128 NVIDIA Pascal GPUs. To accommodate the parallel training, the learning rates are multiplied by the number of parallel running instances. Each experiment trains for 90 epochs... All experiments are conducted with images which are first resized to 224 224 pixels with a random crop, and the JPEG quality used during encoding is 100, so as little information is lost as possible. |