A Hierarchical Model for Device Placement

Authors: Azalia Mirhoseini, Anna Goldie, Hieu Pham, Benoit Steiner, Quoc V. Le, Jeff Dean

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments with widely-used computer vision and natural language models show that our algorithm can find optimized, non-trivial placements for Tensor Flow computational graphs with over 80,000 operations.
Researcher Affiliation Industry {azalia,agoldie,hyhieu,bsteiner,qvl,jeff}@google.com
Pseudocode No No structured pseudocode or algorithm blocks were found in the paper.
Open Source Code No No explicit statement or link providing access to the source code for the methodology was found.
Open Datasets Yes For a fair comparison to previous state-of-the-art deep RL methods (Mirhoseini et al., 2017), we use the same model architectures (Inception-V3, RNNLM, and 2-Layer NMT models), hyperparameters and input data. In addition, we evaluate our model on a 152-layer Res Net (He et al., 2016) with Image Net data (Deng et al., 2009), as well as more complex NMT models with 4 and 8 layers.
Dataset Splits No The paper uses well-known models (e.g., Inception-V3, ResNet) and mentions ImageNet data, but does not provide specific training/validation/test dataset split percentages, sample counts, or explicit instructions for how to partition the data for reproducibility.
Hardware Specification Yes Our experiments are run on machines with 1 Intel Haswell 2300 CPU and up to 8 Nvidia Tesla K40 GPUs.
Software Dependencies Yes We use TensorFlow r1.3 to run our experiments.
Experiment Setup Yes We train both policies using Adam (Kingma & Ba, 2015) optimizer with a fixed learning rate of 0.1, gradient clipping of norm 1.0, tanh constant C = 5.0, and temperature T = 10.0. The number of Grouper and Placer samples in Eqs. 4 and 6 are m = 1 and k = 4, respectively.