Specifying I/O Formats Using The Reformat Free I/O APIs

Table Of Contents

• Description
• How does this sample work?
• Running the sample
  • Sample --help options
• Preparing sample data
• Additional resources
• License
• Changelog
• Known issues

Description

This sample, sampleReformatFreeIO, uses a Caffe model that was trained on the MNIST dataset and performs engine building and inference using TensorRT. The correctness of outputs is then compared to the golden reference. Specifically, it shows how to use reformat free I/O APIs to explicitly specify I/O formats to TensorFormat::kLINEAR, TensorFormat::kCHW2 and TensorFormat::kHWC8 for Float16 and INT8 precision.

How does this sample work?

ITensor::setAllowedFormats is invoked to specify which format is expected to be supported so that the unnecessary reformatting will not be inserted to convert from/to FP32 formats for I/O tensors. BuilderFlag::kSTRICT_TYPES is also assigned to the builder configuration to let the builder choose a reformat free path rather than the fastest path.

Note: If a reformat free path cannot be found, then the fastest path with reformatting will be selected with the following warning message: Warning: no implementation obeys reformatting-free rules, ....

```
bool SampleReformatFreeIO::build(int dataWidth)
{
	...

	network->getInput(0)->setAllowedFormats(static_cast<TensorFormats>(1 << static_cast<int>(mTensorFormat)));
	network->getOutput(0)->setAllowedFormats(static_cast<TensorFormats>(1 << static_cast<int>(mTensorFormat)));
	...
	config->setFlag(BuilderFlag::kSTRICT_TYPES);
	...
}
```

Preparing sample data

1. Download the sample data from TensorRT release tarball, if not already mounted under /usr/src/tensorrt/data (NVIDIA NGC containers) and set it to $TRT_DATADIR.

   export TRT_DATADIR=/usr/src/tensorrt/data
   pushd $TRT_DATADIR/mnist
   pip3 install Pillow
   python3 download_pgms.py
   popd

Running the sample

1. Compile the sample by following build instructions in TensorRT README.

2. Run inference on the digit looping from 0 to 9:

   ./sample_reformat_free_io --datadir=<path/to/data> --useDLACore=N

   For example:

   ./sample_reformat_free_io --datadir $TRT_DATADIR/mnist

3. Verify that all 10 digits match correctly. If the sample runs successfully, you should see output similar to the following:

   &&&& RUNNING TensorRT.sample_reformat_free_io # ./sample_reformat_free_io
   [I] The test chooses MNIST as the network and recognizes a randomly generated digit
   [I] Firstly it runs the FP32 as the golden data, then INT8/FP16 with different formats will be tested
   [I]
   [I] Building and running a FP32 GPU inference to get golden input/output
   [I] [TRT] Detected 1 input and 1 output network tensors.
   [I] Input:
   ... (omitted message)
   &&&& PASSED TensorRT.sample_reformat_free_io
   

   This output shows that the sample ran successfully; PASSED.

Sample --help options

To see the full list of available options and their descriptions, use the -h or --help command line option.

Additional resources

The following resources provide a deeper understanding about this sample:

Models

• MNIST

Documentation

• Introduction To NVIDIA’s TensorRT Samples
• Working With TensorRT Using The C++ API
• NVIDIA’s TensorRT Documentation Library

License

For terms and conditions for use, reproduction, and distribution, see the TensorRT Software License Agreement documentation.

Changelog

June 2019 This is the first release of the README.md file and sample.

Known issues

There are no known issues in this sample.