MAX comes bundled with a super-handy max CLI tool that lets you perform a
couple of common tasks on many ML models without writing a single line of code.
We currently provide two tools:
benchmark: Benchmark models with MLPerf load generation.visualize: Show a model graph as interpreted by MAX Engine in Netron.
Try one out with benchmark-pytorch/run.sh or visualize-pytorch/run.sh to
benchmark or visualize a PyTorch model.
Either of these scripts start by downloading a ResNet50 model (a
model typically used for image classification) and converting it into the
proper format -- for PyTorch, this is TorchScript. This model will be stored
in ../models/resnet50.torchscript.
TorchScript does not include any metadata for shapes the model expects as
input, so we need to provide an --input-data-schema. For this ResNet-50
model, we've already written one for you in
common/resnet50-pytorch/input-spec.yaml, which looks like this:
inputs:
- input_name: pixel_values
shape: 1x3x224x224xf32
compile_shape: ?x3x224x224xf32Benchmarking and visualization works as follows:
% max benchmark --input-data-schema=common/resnet50-pytorch/input-spec.yaml ../models/resnet50.torchscript
% # Benchmarking results will be printed by above command
% max visualize --input-data-schema=common/resnet50-pytorch/input-spec.yaml ../models/resnet50.torchscript
% # Above command will generate a file that can be opened by netron.appLearn more about max benchmark, max visualize, and input data schemas in
the MAX documentation.