Skip to content

Latest commit

 

History

History
70 lines (52 loc) · 2.56 KB

tutorial.md

File metadata and controls

70 lines (52 loc) · 2.56 KB
Raw

Tutorial

In this tutorial, you will learn how to run a simple handwritten digit recognition DNN model (MNIST) in OPML.

Train

First train a DNN model using Pytorch, the training part is shown in mlgo/examples/mnist/training/mnist.ipynb and then save the model at mlgo/examples/mnist/models/mnist/mnist-small.state_dict

Model Format Conversion

Convert the Pytorch model to GGML format, only save the hyperparameters used plus the model weights and biases. Run mlgo/examples/mnist/convert-h5-to-ggml.py to convert your pytorch model. The output format is:

  • magic constant (int32)
  • repeated list of tensors
  • number of dimensions of tensor (int32)
  • tensor dimension (int32 repeated)
  • values of tensor (int32)

Note that the model is saved in big-endian, making it easy to process in the big-endian MIPS-32 VM.

Construct ML Program in MIPS VM

Write a program for DNN model inference, the source code is at mlgo/examples/mnist_mips

Note that the model, input, and output are all stored in the VM memory.

Go supports compilation to MIPS. However, the generated executable is in ELF format. We'd like to get a pure sequence of MIPS instructions instead. To build a ML program in MIPS VM, just run mlgo/examples/mnist_mips/build.sh

Construct VM Image

The user who proposes a ML inference request should first construct an initial VM image

mlvm/mlvm --outputGolden --basedir=/tmp/cannon --program="$PROGRAM_PATH" --model="$MODEL_PATH" --data="$DATA_PATH" --mipsVMCompatible

The initial VM image is saved at /tmp/cannon/golden.json, organized in a Merkle trie tree format.

On-chain Dispute

(For more details, please refer to demo/challenge_simple.sh)

First start a local node

npx hardhat node

Then deploy the smart contract with the initial VM image

npx hardhat run scripts/deploy.js --network localhost

Then the challenger can start the dispute game

BASEDIR=/tmp/cannon_fault npx hardhat run scripts/challenge.js --network localhost

Then the submitter and the challenger will interactively find the dispute point using bisection protocol

for i in {1..25}; do
    echo ""
    echo "--- STEP $i / 25 ---"
    echo ""
    BASEDIR=/tmp/cannon_fault CHALLENGER=1 npx hardhat run scripts/respond.js --network localhost
    BASEDIR=/tmp/cannon CHALLENGER=0 npx hardhat run scripts/respond.js --network localhost
done

Finally, the bisection protocol will help to locate the dispute step, the step will be sent to the arbitration contract on the blockchain.

npx hardhat run scripts/assert.js  --network localhost