For enabling PyTorch launcher you need to add
framework: pytorch in launchers section of your configuration file and provide following parameters:
device- specifies which device will be used for infer (
cudaand so on).
module- PyTorch network module for loading.
checkpoint- pre-trained model checkpoint (Optional).
python_path- appendix for PYTHONPATH for making network module visible in current python environment (Optional).
module_args- list of positional arguments for network module (Optional).
module_kwargs- dictionary (
keyis argument name,
valueis argument value) which represent network module keyword arguments.
adapter- approach how raw output will be converted to representation of dataset problem, some adapters can be specific to framework. You can find detailed instruction how to use adapters here.
batch- batch size for running model (Optional, default 1).
In turn if you model has several inputs you need to specify them in config, using specific parameter:
inputs. Each input description should has following info:
name- input layer name in network
type- type of input values, it has impact on filling policy. Available options:
CONST_INPUT- input will be filled using constant provided in config. It also requires to provide
IMAGE_INFO- specific key for setting information about input shape to layer (used in Faster RCNN based topologies). You do not need to provide
value, because it will be calculated in runtime. Format value is list with
Nelements of the form
[H, W, S], where
Nis batch size,
H- original image height,
W- original image width,
S- scale of original image (default 1).
ORIG_IMAGE_INFO- specific key for setting information about original image size before preprocessing.
IGNORE_INPUT- input which should be stayed empty during evaluation.
INPUT- network input for main data stream (e. g. images). If you have several data inputs, you should provide regular expression for identifier as
valuefor specifying which one data should be provided in specific input.
shape - shape of input layer described as comma-separated of all dimensions size except batch size.
Optionally you can determine
layout in case when your model was trained with non-standard data layout (For PyTorch default layout is
precision (Supported precisions:
FP32 - float,
FP16 - signed shot,
U8 - unsigned char,
U16 - unsigned short int,
I8 - signed char,
I16 - short int,
I32 - int,
I64 - long int).
If you model has several outputs you also need to specify their names in config for ability to get their values in adapter using option
PyTorch launcher config example (demonstrates how to run AlexNet model from torchvision):