Get started with cost model experiments

[merrymercy]

Get started

This tutorial contains a minimal example of training a cost model and using it for search.

Dataset

Install and Download

1. Build and install this repo following the install guide of TVM.
2. Download dataset file.
   • You can download it from google drive
   • Or you can use the command line

     pip3 install gdown
     gdown https://drive.google.com/uc?id=1hciRGyXcGY9fK_owgvlJow8P_l8xYIVJ

3. Put dataset_v3.1.zip under tvm-cost-model/scripts and run unzip dataset_v3.1.zip A new folder dataset will appear in tvm-cost-model/scripts.

Dataset Content

see this readme

Example experiment

Train a cost model and use it for search

Go to tvm-cost-model/scripts.

1. Make a dataset You can either
   • create a sampled smaller dataset for fast experiments.

     python3 make_dataset.py --logs dataset/measure_records/e5-2673/*.json --sample-in-files 100

   • create a complete dataset by using all files. This takes a longer time and requires more memory.

     python3 make_dataset.py --logs dataset/measure_records/e5-2673/*.json

2. Train a cost model

   python3 train_model.py

3. Use the model for search

   python3 tune_network.py --network resnet_50 --n-trials 100 --cost-model xgb-no-update --load-model xgb.pkl

References

• Learning to Optimize Tensor Programs (NeurIPS 18)
• Ansor : Generating High-Performance Tensor Programs for Deep Learning (OSDI 20)
• Value Learning For Throughput Optimization of Deep Neural Networks (MLSys 21)
• A Deep Learning Based Cost Model for Automatic Code Optimization (MLSys 21)
• Learned TPU Cost Model for XLA Tensor Programs (MLSys 21)
• MetaTune: Meta-Learning Based Cost Model for Fast and Efficient Auto-tuning Frameworks
• Ithemal: Accurate, Portable and Fast Basic Block Throughput Estimation using Deep Neural Networks (ICML 19)
• Learning to Represent Programs with Graphs (ICLR 18)
• Learning to Optimize Halide with Tree Search and Random Programs (SIGGRAPH 19)
• CS 330: Deep Multi-Task and Meta Learning

[merrymercy]

Try search for an avx-512 CPU

1. install dataset_v3.1.zip following the instructions above.

2. python3 make_dataset.py --preset batch-size-1 --sample-in-files 500
   python3 train_model.py
   python3 tune_network.py --network resnet_50 --n-trials 100 --target "llvm -mcpu=skylake-avx512" --cost-model xgb-no-update --load-model xgb.pkl

[merrymercy]

Dataset and hardware platforms

Folder	Hardware platform	VM instance type
platinum-8272	Intel platinum-8272	Azure D32s_v4
e5-2673	Intel e5-2673	Azure F16s
epyc-7452	AMD epyc-7452	Azure D16as_v4
k80	NVIDIA K80	AWS p2

[catmaohphp]

The dataset file is not available for download now.

[merrymercy]

@catmaohphp This issue is outdated. see https://github.com/tlc-pack/tenset/blob/main/docs/get_started_with_cost_model_experiments.md

[catmaohphp]

Thank you for your quick response!