Code artifacts and data for CS 590 project.

The project involves several changes to the TVM deep learning compiler. These changes can be found at my fork. To run experiments, please clone the fork and build from source following these instructions.

Experiments include:

• exp1.sh, exp2.sh, exp3.sh: these scripts summarize paper experiments.
  • Autotuning commands are commented by default due to their long running time.
• Individual scripts were adapted from AutoTVM tutorials. Explanations:
  • tune_conv2d_cuda.py
    • Tune a specific conv2d operator configuration.
  • tune_conv2d_cuda_test.py
    • Evaluate performance of tuned conv2d configuration stored in log file.
  • feature_experiments_resnet18/tune_nnvm_cuda.py
    • Tune ResNet-18 inference (12 conv2d configurations total).
  • transfer-learning/tune_conv2d_cuda_transfer.py
    • Tune conv2d operator using pretrained data for transfer learning.
• Incomplete neural network cost model experiments in nn-cost-model and treernn-cost-model.

Autotuning result files are also included, as autotuning execution takes many hours.

• baseline, feature_experiments_c7, feature_experiments_c12: conv2d tuning results.
• feature_experiments_resnet18: end-to-end ResNet-18 tuning results.
• transfer-learning: conv2d transfer learning results.

Raw, unpolished data can be found on the raw-data branch.

Figures in paper are generated via plot.sh.

• Final report
• Research reflections