Page History

(authored by Kellen Sunderland)

This document described how to use This document provides a detailed description of the MXNet-TensorRT runtime integration to accelerate model inference. feature.  This document covers advanced techniques, contains a roadmap reflecting the current state of the feature and future directions, and also contains up-to-date benchmarks.  If you'd like a quick overview of the feature with a tutorial describing a simple use-case please refer to this MXNet hosted tutorial.  For more information you may also visit the original design proposal page.

Why is TensorRT integration useful?

TensorRT can greatly speed up inference of deep learning models. One experiment on a Titan V (V100) GPU shows that with MXNet 1.2, we can get an approximately 3x speed-up when running inference of the ResNet-50 model on the CIFAR-10 dataset in single precision (fp32). As batch sizes and image sizes go up (for CNN inference), the benefit may be less, but in general, TensorRT helps especially in cases which have:

• Many bandwidth-bound or latency-bound layers (e.g. pointwise operations) that benefit from GPU kernel fusion.
• Inference use cases which have tight latency requirements and where the client application can't wait for large batches to be queued up.
• Embedded systems, where memory constraints are tighter than on servers.
• When performing inference in reduced precision, especially for integer (e.g. int8) inference.

...

The above points ensure that we find a compromise between the flexibility of MXNet, and fast inference in TensorRT. We do this with no additional burden to the user. Users do not need to learn how TensorRT APIs work, and do not need to write their own client application or data pipeline.

How do I

...

use TensorRT integration?

...

Building MXNet together with TensorRT is somewhat complex. The recipe will hopefully be simplified in the near future, but for now, it's easiest to build a Docker container with a Ubuntu 16.04 base. This Dockerfile can be found under the ci subdirectory of the MXNet repository. You can build the container as follows:

docker build -t ci/docker/Dockerfile.build.ubuntu_gpu_tensorrt mxnet_with_tensorrt

...

A full tutorial is provided here but we'll summarize for a simple use case below.

Installation

Installing MXNet with TensorRT integration is an easy process. First ensure that you are running Ubuntu 16.04, that you have updated your video drivers, and you have installed CUDA 9.0 or 9.2. You’ll need a Pascal or newer generation NVIDIA gpu. You’ll also have to download and install TensorRT libraries instructions here. Once your these prerequisites installed and up-to-date you can install a special build of MXNet with TensorRT support enabled via PyPi and pip. Install the appropriate version by running:

To install with CUDA 9.0:

pip install mxnet-tensorrt-cu90

To install with CUDA 9.2:

pip install mxnet-tensorrt-cu92

If you are running an operating system other than Ubuntu 16.04, or just prefer to use a docker image with all prerequisites installed you can instead run

...

:

nvidia-docker run -ti --rm mxnet_with_tensorrt

After starting the container, you will find yourself in the /opt/mxnet directory by default.

Running a "hello, world" model / unit test (LeNet-5 on MNIST)

You can then run the LeNet-5 unit test, which will train LeNet-5 on MNIST using the symbolic API. The test will then run inference in MXNet both with, and without MXNet-TensorRT runtime integration. Finally, the test will display a comparison of both runtime's accuracy scores. The test can be run as follows:

python ${MXNET_HOME}/tests/python/tensorrt/test_tensorrt_lenet5.py

...

Running inference in MXNet
[03:31:18] src/operator/nn/./cudnn/./cudnn_algoreg-inl.h:107: Running performance tests to find the best convolution algorithm, this can take a while... (setting env variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
Running inference in MXNet-TensorRT
MXNet accuracy: 98.680000
MXNet-TensorRT accuracy: 98.680000

Running more complex models

The unit test directory also provides a way to run models from the Gluon model zoo after slight modifications. The models that are tested are CNN classification models from the Gluon zoo. They are mostly based on ResNet, but include ResNeXtas well:

• cifar_resnet20_v1
• cifar_resnet56_v1
• cifar_resnet110_v1
• cifar_resnet20_v2
• cifar_resnet56_v2
• cifar_resnet110_v2
• cifar_wideresnet16_10
• cifar_wideresnet28_10
• cifar_wideresnet40_8
• cifar_resnext29_16x64d

Please note that even those examples are based on CIFAR-10 due to the ease of accessing the dataset without formal registration and preprocessing, everything should work fine with models trained on ImageNet, using MXNet's ImageNet iterators, based on the RecordIO representation of the ImageNet dataset.

The script can be run simply as:

python ${MXNET_HOME}/tests/python/tensorrt/test_tensorrt_resnet_resnext.py

...

===========================================
Model: cifar_resnet56_v1
===========================================

*** Running inference using pure MXNet ***
MXNet: time elapsed: 2.463s, accuracy: 94.19%

*** Running inference using MXNet + TensorRT ***
TensorRT: time elapsed: 1.652s, accuracy: 94.19%

TensorRT speed-up (not counting compilation): 1.49x
Absolute accuracy difference: 0.000000


===========================================
Model: cifar_resnet110_v1
===========================================

*** Running inference using pure MXNet ***
MXNet: time elapsed: 4.000s, accuracy: 95.20%

*** Running inference using MXNet + TensorRT ***

TensorRT: time elapsed: 2.085s, accuracy: 95.20%

TensorRT speed-up (not counting compilation): 1.92x
Absolute accuracy difference: 0.000000

As you can see, the speed-up varies by model. ResNet-110 has more layers that can be fused than ResNet-56, hence the speed-up is greater.

Running TensorRT with your own models with the symbolic API

When building your own models, feel free to use the above ResNet-50 model as an example. Here, we highlight a small number of issues that need to be taken into account.

1. When loading a pre-trained model, the inference will be handled using the Symbol API, rather than the Module API.

2. In order to provide the weights from MXNet (NNVM) to the TensorRT graph converter before the symbol is fully bound (before the memory is allocated, etc.), the arg_params and aux_params need to be provided to the symbol's simple_bind method. The weights and other values (e.g. moments learned from data by batch normalization, provided via aux_params) will be provided via the shared_buffer argument to simple_bind as follows:

executor = sym.simple_bind(ctx=ctx, data = data_shape,
softmax_label=sm_shape, grad_req='null', shared_buffer=all_params, force_rebind=True)

...

mxnet/tensorrt bash

Model Initialization

import mxnet as mx
from mxnet.gluon.model_zoo import vision
import time
import os

batch_shape = (1, 3, 224, 224)
resnet18 = vision.resnet18_v2(pretrained=True)
resnet18.hybridize()
resnet18.forward(mx.nd.zeros(batch_shape))
resnet18.export('resnet18_v2')
sym, arg_params, aux_params = mx.model.load_checkpoint('resnet18_v2', 0)

Baseline MXNet Network Performance

# Create sample input
input = mx.nd.zeros(batch_shape)

# Execute with MXNet
os.environ['MXNET_USE_TENSORRT'] = '0'
executor = sym.simple_bind(ctx=mx.gpu(0), data=batch_shape, grad_req='null', force_rebind=True)
executor.copy_params_from(arg_params, aux_params)

# Warmup
print('Warming up MXNet')
for i in range(0, 10):
	y_gen = executor.forward(is_train=False, data=input)
	y_gen[0].wait_to_read()

# Timing
print('Starting MXNet timed run')
start = time.process_time()
for i in range(0, 10000):
	y_gen = executor.forward(is_train=False, data=input)
	y_gen[0].wait_to_read()
end = time.time()
print(time.process_time() - start)

TensorRT Integrated Network Performance

# Execute with TensorRT
print('Building TensorRT engine')
os.environ['MXNET_USE_TENSORRT'] = '1'
arg_params.update(aux_params)
all_params = dict([(k, v.as_in_context(mx.gpu(0))) for k, v in arg_params.items()])
executor = mx.contrib.tensorrt.tensorrt_bind(sym, ctx=mx.gpu(0), all_params=all_params,
                                             data=batch_shape, grad_req='null', force_rebind=True)
# Warmup
print('Warming up TensorRT')
for i in range(0, 10):
	y_gen =

def merge_dicts(*dict_args):
    result = {}
    for dictionary in dict_args:
        result.update(dictionary)
        return result
sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, epoch)
all_params = merge_dicts(arg_params, aux_params)

...

for idx, dbatch in enumerate(test_iter):
    data = dbatch.data[0]
    executor.arg_dict["data"][:] = data
    executor.forward(is_train=False)
    preds = executor.outputs, data=input)
	y_gen[0].asnumpywait_to_read()

# Timing
print('Starting TensorRT   top1timed run')
start = nptime.argmax(preds, axis=1)

...

def set_use_tensorrt(status = False):
        os.environ["MXNET_USE_TENSORRT"] = str(int(status))

...

print("Running inference in MXNet")
set_use_tensorrt(False)
mx_pct = run_inference(sym, arg_params, aux_params, mnist, all_test_labels, batch_size=batch_size)

print("Running inference in MXNet-TensorRT")
set_use_tensorrt(True)
trt_pct = run_inference(sym, arg_params, aux_params, mnist, all_test_labels,  batch_size=batch_size)

...

Running TensorRT with your own models with the Gluon API

Note: Please first read the previous section titled "Running TensorRT with your own models with the symbolic API" - it contains information that will also be useful for Gluonusers.

Note: If the user wishes to use the Gluon vision models, it's necessary to install the gluoncv pip package:

pip install gluoncv

The above package is based on a separate repository.

For Gluon models specifically, we need to add a data symbol to the model to load the data, as well as apply the softmax layer, because the Gluon models only present the logits that are to be presented for softmax. This is shown in python ${MXNET_HOME}/tests/python/tensorrt/test_tensorrt_resnet_resnext.py. Here's the relevant code:

net = gluoncv.model_zoo.get_model(model_name, pretrained=True)
data = mx.sym.var('data')
out = net(data)
softmax = mx.sym.SoftmaxOutput(out, name='softmax')

...

net = gluoncv.model_zoo.get_model(model_name, pretrained=True)
all_params = dict([(k, v.data()) for k, v in net.collect_params().items()])
executor = softmax.simple_bind(ctx=ctx, data=(batch_size, 3, 32, 32), softmax_label=(batch_size,), grad_req='null',
                                   shared_buffer=all_params, force_rebind=True)

...

gluon.data.DataLoader(
        gluon.data.vision.CIFAR10(train=False).transform_first(transform_test),
        batch_size=batch_size, shuffle=False, num_workers=num_workers)

...

process_time()
for i in range(0, 10000):
	y_gen = executor.forward(is_train=False, data=input)
	y_gen[0].wait_to_read()
end = time.time()
print(time.process_time() - start)

The output should be the same both when using an MXNet executor and when using a TensorRT executor.  The performance speedup should be roughly 1.8x depending on the hardware and libraries used.

Roadmap

Finished Items

Initial integration has been completed and launched as of MXNet 1.3.  We've tested this integration against a variety of models, including all the gluonCV models, Wavenet and some custom computer vision models.  Performance is roughly in line with expectations, but we're seeing a few regressions over earlier measurements that require investigation.

Continuous Integration support is enabled and running continually for all active PRs opened with MXNet.

PIP packages and Docker images have been published along with the MXNet 1.3 release.

Future work

FP16 Integration

The current integration of TensorRT into MXNet supports only FP32 float values for tensors.  Allowing FP16 values would enable many further optimizations on Jetson and Volta devices.

https://jira.apache.org/jira/browse/MXNET-1084

Subgraph Integration

The new subgraph API is a natural fit for TensorRT.  To help make the codebase consistent we'd like to port the current TensorRT integration to use the new API.  The experimental integration into MXNet requires us to use contrib API calls.  Once integration has moved to use the subgraph API users will be able to use TensorRT with a consistent API.  Porting should also enable acceleration of gluon and module base models.

https://jira.apache.org/jira/browse/MXNET-1085

Conditional Checkout and Compilation of Dependencies

TensorRT integration required us to add a number of third party code sub-repositories to the project.  This is not ideal for users who would like to checkout and build MXNet without using the TensorRT feature.  In the future we should migrate the feature to be CMake only, and checkout the project at pre-compilation time to avoid forcing all users to checkout these subrepos.  We can also model these dependencies using CMake such that they're automatically built and linked against when required, which would make building from scratch easier for those that do want to use TensorRT integration.

Make use of Cached TRT Engines

Similar to the cudnn auto-tuning feature we've received requests from users that we cache TensorRT engine compilations so that we avoid the delay of building the engine each time we start the process.

Increased Operator (/Layer) Coverage

The current operator coverage is fairly limited.  We'd like to enable all models that TensorRT is able to work with. 

Decouple  NNVM to ONNX from NNVM to TensorRT in MXNet

The current nnvm_to_onnx classes are tightly coupled to TensorRT.  We could extract all of the TensorRT specific functionality and have a proper separation between nnvm_to_onnx and onnx_to_tensorrt.  When structuring nnvm_to_onnx we should make use of object hierarchy to convert to specific opsets of onnx to help us maintain compatibility with different toolsets.  We should create a base class that performs generic onnx conversions.  We should then specialized objects that inherit from the base onnx class and take care of the  differences between opsets.  We should also create unit tests on a per-op basis to make sure we're introducing regressions.

Currently supported operators:

Operators to be added:

Benchmarks

TensorRT is still an experimental feature, so benchmarks are likely to improve over time.  As of Oct 11, 2018 we've measured the following improvements which have all been run with FP32 weighted networks.

Related articles

https://mxnet.incubator.apache.org/tutorials/tensorrt/inference_with_trt.html

Runtime Integration with TensorRT

Examples

Image classification examples are available here.

...

...