Chameleon: Adaptive Code Optimization for Expedited Deep Neural Network Compilation
Authors: Byung Hoon Ahn, Prannoy Pilligundla, Amir Yazdanbakhsh, Hadi Esmaeilzadeh
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentation with real hardware shows that CHAMELEON provides 4.45 speed up in optimization time over Auto TVM, while also improving inference time of the modern deep networks by 5.6%. |
| Researcher Affiliation | Collaboration | Byung Hoon Ahn1, Prannoy Pilligundla1, Amir Yazdanbakhsh2, Hadi Esmaeilzadeh1 1 University of California, San Diego 2 Google Research |
| Pseudocode | Yes | Algorithm 1 Adaptive Sampling and Sample Synthesis |
| Open Source Code | Yes | CHAMELEON is publicly available in the project page: https://bitbucket.org/act-lab/chameleon. |
| Open Datasets | Yes | Table 4: Details of the DNN models used in evaluating CHAMELEON. NETWORK DATASET Alex Net Image Net VGG-16 Image Net Res Net-18 Image Net |
| Dataset Splits | No | The paper does not explicitly provide training/validation/test dataset splits with percentages or counts. It mentions training time reduction but not how the datasets were split for training and validation. |
| Hardware Specification | Yes | Table 6: Details of the hardware used for evaluation of CHAMELEON. SPECIFICATIONS DETAILS GPU Titan Xp Host CPU 3.4G Hz Intel Core i7 Main Memory 32GB 2400 MHz DDR3 |
| Software Dependencies | No | The paper mentions using TVM, Auto TVM, and Proximal Policy Optimization (PPO), and frameworks like Tensor Flow and Py Torch. However, it does not provide specific version numbers for these software dependencies (e.g., 'TVM vX.Y' or 'PyTorch vZ.W'). |
| Experiment Setup | Yes | Table 7: Hyper-parameters uses in CHAMELEON. HYPERPARAMETER VALUE DESCRIPTION iterationopt 16 number of iterations for optimization process (equivalent to 1000 hardware measurements) mode GBT xgb-reg type of loss used for cost model b GBT 64 maximum batch size of planning in GBT (Chen & Guestrin, 2016) cost model per iteration of optimization process episoderl 128 number of episodes for reinforcement learning steprl 500 maximum steps of one reinforcement learning episode thresholdmeta 2.5 threshold used for meta-search in sampling |