Device Placement Optimization with Reinforcement Learning
Authors: Azalia Mirhoseini, Hieu Pham, Quoc V. Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen Kumar, Mohammad Norouzi, Samy Bengio, Jeff Dean
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our main result is that on Inception-V3 for Image Net classification, and on RNN LSTM, for language modeling and neural machine translation, our model finds non-trivial device placements that outperform hand-crafted heuristics and traditional algorithmic methods. In the following experiments, we apply our proposed method to assign computations to devices on three important neural networks in the deep learning literature: Recurrent Neural Language Model (RNNLM), Attentional Neural Machine Translation (Bahdanau et al., 2015), and Inception V3 (Szegedy et al., 2016). We compare the RL placements against strong existing baselines described in Section 4.2. |
| Researcher Affiliation | Industry | 1Google Brain 2Members of the Google Brain Residency Program (g.co/brainresidency) 3Google. Correspondence to: Azalia Mirhoseini <azalia@google.com>, Hieu Pham <hyhieu@google.com>. |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not provide an unambiguous statement or link for the open-sourcing of the code for their proposed method. |
| Open Datasets | Yes | We evaluate our approach on three established deep learning models: Recurrent Neural Network Language Model (RNNLM) (Zaremba et al., 2014; Jozefowicz et al., 2016), Attentional Neural Machine Translation (Bahdanau et al., 2015), and Inception V3 (Szegedy et al., 2016). We train our Neural MT model on the WMT14 English-German dataset.1 [...]. We train Inception-V3 on the Image Net dataset (Russakovsky et al., 2015) until the model reaches the accuracy of 72% on the validation set. |
| Dataset Splits | Yes | We train Inception-V3 on the Image Net dataset (Russakovsky et al., 2015) until the model reaches the accuracy of 72% on the validation set. |
| Hardware Specification | Yes | In our experiments, the available devices are 1 Intel Haswell 2300 CPU, which has 18 cores, and either 2 or 4 Nvidia Tesla K80 GPUs. We allow 50 GB of RAM for all models and settings. |
| Software Dependencies | No | The paper mentions software like Tensor Flow, Adam, RMSProp, and Scotch, but it does not provide specific version numbers for these dependencies (e.g., 'TensorFlow: A system for large-scale machine learning. arXiv preprint arXiv:1605.08695, 2016.', 'Adam (Kingma & Ba, 2014)', 'RMSProp (Tieleman & Hinton, 2012)', 'Scotch optimizer (Pellegrini, 2009)'). |
| Experiment Setup | Yes | For Inception-V3, each step is executed on a batch of images, each of size 299x299x3. For RNNLM and NMT, we use the model with 2 LSTM layers, with sizes of 2048 and 1024, respectively. We set the number of unrolling steps for RNNLM, as well as the maximum length for the source and target sentences of NMT, to 40. Each pass on RNNLM and NMT consists of a minibatch of 64 sequences. We train each model for 200,000 steps [...]. All settings use the learning rate of 0.045 and are trained using RMSProp. |