TAPAS: Train-Less Accuracy Predictor for Architecture Search

Authors: R. Istrate, F. Scheidegger, G. Mariani, D. Nikolopoulos, C. Bekas, A. C. I. Malossi3927-3934

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we demonstrate TAPAS performance over a wide range of experiments. Results are compared with reference works from the literature.
Researcher Affiliation Collaboration R. Istrate,1,2 F. Scheidegger,1 G. Mariani,1 D. Nikolopoulos,2 C. Bekas,1 A. C. I. Malossi1 1IBM Research Zurich, Switzerland 2University of Belfast, United Kingdom
Pseudocode No The paper describes procedures and architectures but does not include any explicit pseudocode blocks or figures labeled as 'Algorithm'.
Open Source Code No The paper states 'TAPAS is one of the AI engines in IBM s new breakthrough capability called Neu Net S, that will be available to users as part of the AI Open Scale (Smith 2018).' This indicates commercial availability rather than open-source code release for the described methodology.
Open Datasets Yes All the experiments are based on a LDE populated with nineteen datasets, ranked by difficulty in Figure 2. Eleven of them are publicly available. The other eight are generated by sub-sampling the Image Net dataset (Deng et al. 2009) varying the number of classes and the number of images per class.
Dataset Splits Yes We perform ten-fold cross validation and present the results for Peephole, LCE, and TAPAS in the first row of Figure 4. (...) we consider the list of datasets in Figure 2 and perform eleven leave-one-out cross-validation benchmarks, considering only the real datasets.
Hardware Specification Yes All runs involve single-precision arithmetic and are performed on IBM1 POWER8 compute nodes, equipped with four NVIDIA P100 GPUs.
Software Dependencies No The paper mentions optimizers and initialization methods like 'RMSprop' and 'He Normal weight initialization' but does not provide specific software library names with version numbers (e.g., TensorFlow 2.x, PyTorch 1.x).
Experiment Setup Yes To facilitate the TAP, we train all networks with the same hyperparameters, i.e., same optimizer, learning rate, batch size, and weights initiallizer. (...) TAP is trained with RMSprop (Tieleman and Hinton 2012), using a learning rate of 10 3, a He Normal weight initialization (He et al. 2015), and a batch size of 512. (...) All networks are trained under the same settings: RMSprop optimizer with a learning rate of 10 3, weight decay 10 4, batch size 64, and He Normal weight initialization.