model_zoo_executor

README

About

This is a simple utility that runs a model from the model zoo thanks to the Gorgonia backend

Example

Download a pre-trained model from the zoo (for now, only MNIST is known to work)

then smply run:

go run main.go -model /tmp/mnist/model.onnx -input /tmp/mnist/test_data_set_0/input_0.pb -output /tmp/mnist/test_data_set_0/output_0.pb

The utility evaluates the model and check if the computed output is equal to the expected output (within a delta of 5e-3). If the result is ok, it displays the result:

[975.67035 -618.7244 6574.5684 668.0278 -917.27057 -1671.6357 -1952.7606 -61.54949 -777.17645 -1439.5311]

Test files

You can test (and bench a model) with go test. Simply download a model from the zoo into a subdirectory of the ./models dir; then run:

MODELDIR=./models/YOURMODEL go test

If you have several models inside the ./models dir, you can run them all by typing make

Bench

A simple benchmark that measures the execution time of the model is also included; to run it:

MODELDIR=./models/YOURMODEL go test -bench=.

Profiling

You can also use this benchmark for profiling the application:

MODELDIR=models/YOURMODEL go test -bench=. -benchmem -memprofile memprofile.out -cpuprofile profile.out -benchtime=10s

The use the pprof tool:

go tool pprof -http=localhost:3999 memprofile.out or go tool pprof -http=localhost:3999 profile.out