Skip to content

Commit c9dcd5e

Browse files
jaybdubjaybdub
authored
Merge pull request #399 from jaybdub/docs
Docs
2 parents f5fb752 + 62afaf1 commit c9dcd5e

23 files changed

+566
-0
lines changed

‎.gitignore‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -15,3 +15,4 @@ __pycache__/
1515
*.pyc
1616
*.ipynb_checkpoints
1717
*.pth
18+
docs/converters.md

‎CHANGELOG.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1 @@
1+
# Changes

‎CONTRIBUTING.md‎

Lines changed: 104 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,104 @@
1+
# Contributing
2+
3+
## Forms of contribution
4+
5+
### Submit an Issue
6+
7+
torch2trt is use case driven. We originally created it to solve
8+
use cases related to NVIDIA Jetson, but the layer support has grown
9+
largely since it's release and we've found that it has
10+
helped many other developers as well.
11+
12+
The growth of torch2trt has been largely driven by issues submitted on [GitHub](https://github.com/NVIDIA-AI-IOT/torch2trt/issues).
13+
We learn a lot from the reported issues. Submitting an issue it is one of the best ways to begin contributing to torch2trt.
14+
15+
The reported issues typically are one of the following,
16+
17+
* A bug or unexpected result
18+
* A model with unsupported layers
19+
20+
If you report an issue, we typically find the following information helpful
21+
22+
* PyTorch version
23+
* TensorRT version
24+
* Platform (ie: Jetson Nano)
25+
* The PyTorch Module you're attempting to convert
26+
* The steps taken to convert the PyTorch module
27+
28+
If you're not sure how to provide any of these pieces of information, don't worry. Just open the issue
29+
and we're happy to discuss and help work out the details.
30+
31+
### Ask a Question
32+
33+
Another great way to contribute is to ask a question on [GitHub](https://github.com/NVIDIA-AI-IOT/torch2trt/issues).
34+
There are often other developers who share your question, and they may find the discussion helpful. This also
35+
helps us gauge feature interest and identify gaps in documentation.
36+
37+
### Submit a Pull Request
38+
39+
torch2trt is use case driven and has limited maintainence, for this reason we value community contributions greatly.
40+
Another great way to contribute is by submitting a pull request. Pull requests which are most likely to be accepted are
41+
42+
* A new converter
43+
* A test case
44+
* A bug fix
45+
46+
If you add a new converter, it is best to include a few test
47+
cases that cross validate the converter against the original PyTorch. We provide a utility function to do this,
48+
as described in the [Custom Converter](usage/custom_converter.md) usage guide.
49+
50+
Ideally pull requests solve one thing at a time. This makes it easy
51+
to evaluate the impact that the changes have on the project step-by-step. The more confident we are that
52+
the changes will not adversely impact the experience of other developers, the more likely we are to accept them.
53+
54+
## Running module test cases
55+
56+
Before any change is accepted, we run the test cases on at least one platform. This performs a large number
57+
of cross validation checks against PyTorch. To do this
58+
59+
```bash
60+
python3 -m torch2trt.test --name=converters --tolerance=1e-2
61+
```
62+
63+
This will not hard-fail, but will highlight any build errors or max error checks. It is helpful if you include
64+
the status of this command in any pull-request, as well as system information like
65+
66+
* PyTorch version
67+
* TensorRT version
68+
* Platform (ie: Jetson Nano)
69+
70+
## Testing documentation
71+
72+
If you have a change that modifies the documentation, it is relatively straightforward to test. We
73+
use ``mkdocs-material`` for documentation, which parses markdown files in the ``docs`` folder.
74+
75+
To view the docs, simply call
76+
77+
```
78+
./scripts/test_docs.sh
79+
```
80+
81+
And then navigate to ``https://<ip_address>:8000``.
82+
83+
Please note, this will not include dynamically generated documentation pages like the converters page.
84+
These contain cross reference links to the GitHub source code. If you want to test these
85+
you can call
86+
87+
```bash
88+
./scripts/build_docs.sh <github url> <tag>
89+
```
90+
91+
Pointing to the public reflection
92+
of your local repository. For example, if we're working off the upstream master branch, we
93+
would call
94+
95+
```bash
96+
./scripts/build_docs.sh https://github.com/NVIDIA-AI-IOT/torch2trt master
97+
```
98+
99+
If your changes are pushed to your fork, you would do
100+
101+
```bash
102+
./scripts/build_docs.sh https://github.com/<user>/torch2trt my_branch
103+
```
104+

‎docs/CHANGELOG.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1 @@
1+
../CHANGELOG.md

‎docs/CONTRIBUTING.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1 @@
1+
../CONTRIBUTING.md

‎docs/benchmarks/jetson_nano.md‎

Lines changed: 22 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,22 @@
1+
# Jetson Nano
2+
3+
| Name | Data Type | Input Shapes | torch2trt kwargs | Max Error | Throughput (PyTorch) | Throughput (TensorRT) | Latency (PyTorch) | Latency (TensorRT) |
4+
|------|-----------|--------------|------------------|-----------|----------------------|-----------------------|-------------------|--------------------|
5+
| torchvision.models.alexnet.alexnet | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.29E-05 | 46.4 | 69.9 | 22.1 | 14.7 |
6+
| torchvision.models.squeezenet.squeezenet1_0 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.20E-02 | 44 | 137 | 24.2 | 7.6 |
7+
| torchvision.models.squeezenet.squeezenet1_1 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 9.77E-04 | 76.6 | 248 | 14 | 4.34 |
8+
| torchvision.models.resnet.resnet18 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 5.86E-03 | 29.4 | 90.2 | 34.7 | 11.4 |
9+
| torchvision.models.resnet.resnet34 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.56E-01 | 15.5 | 50.7 | 64.8 | 20.2 |
10+
| torchvision.models.resnet.resnet50 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 6.45E-02 | 12.4 | 34.2 | 81.7 | 29.8 |
11+
| torchvision.models.resnet.resnet101 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.01E+03 | 7.18 | 19.9 | 141 | 51.1 |
12+
| torchvision.models.resnet.resnet152 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 0.00E+00 | 4.96 | 14.1 | 204 | 72.3 |
13+
| torchvision.models.densenet.densenet121 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 3.42E-03 | 11.5 | 41.9 | 84.5 | 24.8 |
14+
| torchvision.models.densenet.densenet169 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 5.86E-03 | 8.25 | 33.2 | 118 | 31.2 |
15+
| torchvision.models.densenet.densenet201 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 3.42E-03 | 6.84 | 25.4 | 141 | 40.8 |
16+
| torchvision.models.densenet.densenet161 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 4.15E-03 | 4.71 | 15.6 | 247 | 65.8 |
17+
| torchvision.models.vgg.vgg11 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 3.51E-04 | 8.9 | 18.3 | 114 | 55.1 |
18+
| torchvision.models.vgg.vgg13 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 3.07E-04 | 6.53 | 14.7 | 156 | 68.7 |
19+
| torchvision.models.vgg.vgg16 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 4.58E-04 | 5.09 | 11.9 | 201 | 85.1 |
20+
| torchvision.models.vgg.vgg11_bn | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 3.81E-04 | 8.74 | 18.4 | 117 | 54.8 |
21+
| torchvision.models.vgg.vgg13_bn | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 5.19E-04 | 6.31 | 14.8 | 162 | 68.5 |
22+
| torchvision.models.vgg.vgg16_bn | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 9.77E-04 | 4.96 | 12 | 207 | 84.3 |

‎docs/benchmarks/jetson_xavier.md‎

Lines changed: 33 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,33 @@
1+
# Jetson Xavier
2+
3+
| Name | Data Type | Input Shapes | torch2trt kwargs | Max Error | Throughput (PyTorch) | Throughput (TensorRT) | Latency (PyTorch) | Latency (TensorRT) |
4+
|------|-----------|--------------|------------------|-----------|----------------------|-----------------------|-------------------|--------------------|
5+
| torch2trt.tests.torchvision.classification.alexnet | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 7.63E-05 | 251 | 565 | 4.96 | 2.02 |
6+
| torch2trt.tests.torchvision.classification.squeezenet1_0 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 9.77E-04 | 121 | 834 | 8.04 | 1.49 |
7+
| torch2trt.tests.torchvision.classification.squeezenet1_1 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 9.77E-04 | 125 | 1.29e+03 | 8.01 | 1.02 |
8+
| torch2trt.tests.torchvision.classification.resnet18 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 9.77E-03 | 136 | 722 | 7.33 | 1.64 |
9+
| torch2trt.tests.torchvision.classification.resnet34 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.50E-01 | 77.8 | 396 | 12.9 | 2.79 |
10+
| torch2trt.tests.torchvision.classification.resnet50 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.09E-01 | 55.8 | 326 | 17.9 | 3.37 |
11+
| torch2trt.tests.torchvision.classification.resnet101 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 0.00E+00 | 28.3 | 175 | 35.1 | 6.04 |
12+
| torch2trt.tests.torchvision.classification.resnet152 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 0.00E+00 | 18.8 | 122 | 53.2 | 8.57 |
13+
| torch2trt.tests.torchvision.classification.densenet121 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 7.81E-03 | 20.9 | 76.6 | 47.5 | 13 |
14+
| torch2trt.tests.torchvision.classification.densenet169 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 3.91E-03 | 14.8 | 41.7 | 66.7 | 23.7 |
15+
| torch2trt.tests.torchvision.classification.densenet201 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 4.88E-03 | 12.6 | 30.2 | 79.1 | 33 |
16+
| torch2trt.tests.torchvision.classification.densenet161 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 4.88E-03 | 16.1 | 43.7 | 62.1 | 23 |
17+
| torch2trt.tests.torchvision.classification.vgg11 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.56E-03 | 84.8 | 201 | 12.1 | 5.24 |
18+
| torch2trt.tests.torchvision.classification.vgg13 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.24E-03 | 71.1 | 165 | 14.3 | 6.34 |
19+
| torch2trt.tests.torchvision.classification.vgg16 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 3.78E-03 | 61.5 | 139 | 16.5 | 7.46 |
20+
| torch2trt.tests.torchvision.classification.vgg19 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.81E-03 | 54.1 | 120 | 18.7 | 8.61 |
21+
| torch2trt.tests.torchvision.classification.vgg11_bn | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.20E-03 | 81.5 | 200 | 12.5 | 5.27 |
22+
| torch2trt.tests.torchvision.classification.vgg13_bn | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.71E-03 | 67.5 | 165 | 15.1 | 6.33 |
23+
| torch2trt.tests.torchvision.classification.vgg16_bn | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.87E-03 | 58.3 | 139 | 17.4 | 7.48 |
24+
| torch2trt.tests.torchvision.classification.vgg19_bn | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.44E-03 | 51.4 | 120 | 19.7 | 8.61 |
25+
| torch2trt.tests.torchvision.classification.mobilenet_v2 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 0.00E+00 | 64.8 | 723 | 15.4 | 1.67 |
26+
| torch2trt.tests.torchvision.classification.shufflenet_v2_x0_5 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.53E-05 | 51.2 | 463 | 19.4 | 2.17 |
27+
| torch2trt.tests.torchvision.classification.shufflenet_v2_x1_0 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.53E-05 | 49.4 | 419 | 20.4 | 2.43 |
28+
| torch2trt.tests.torchvision.classification.shufflenet_v2_x1_5 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.53E-05 | 51.4 | 426 | 19.6 | 2.37 |
29+
| torch2trt.tests.torchvision.classification.shufflenet_v2_x2_0 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 1.53E-05 | 48.2 | 419 | 20.8 | 2.48 |
30+
| torch2trt.tests.torchvision.classification.mnasnet0_5 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 2.03E-06 | 67.8 | 883 | 14.9 | 1.4 |
31+
| torch2trt.tests.torchvision.classification.mnasnet0_75 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 0.00E+00 | 67.6 | 751 | 14.8 | 1.6 |
32+
| torch2trt.tests.torchvision.classification.mnasnet1_0 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 0.00E+00 | 65.7 | 667 | 15.2 | 1.77 |
33+
| torch2trt.tests.torchvision.classification.mnasnet1_3 | float16 | [(1, 3, 224, 224)] | {'fp16_mode': True} | 0.00E+00 | 67.4 | 573 | 15 | 2.02 |

‎docs/css/version-select.css‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,5 @@
1+
@media only screen and (max-width:76.1875em) {
2+
#version-selector {
3+
padding: .6rem .8rem;
4+
}
5+
}

‎docs/getting_started.md‎

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,32 @@
1+
# Getting Started
2+
3+
Follow these steps to get started using torch2trt.
4+
5+
!!! note
6+
7+
torch2trt depends on the TensorRT Python API. On Jetson, this is included with the latest JetPack. For desktop, please follow the [TensorRT Installation Guide](https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html). You may also try installing torch2trt inside one of the NGC PyTorch docker containers for [Desktop](https://ngc.nvidia.com/catalog/containers/nvidia:pytorch) or [Jetson](https://ngc.nvidia.com/catalog/containers/nvidia:l4t-pytorch).
8+
9+
### Install Without plugins
10+
11+
To install without compiling plugins, call the following
12+
13+
```bash
14+
git clone https://github.com/NVIDIA-AI-IOT/torch2trt
15+
cd torch2trt
16+
python setup.py install
17+
```
18+
19+
### Install With plugins
20+
21+
To install with plugins to support some operations in PyTorch that are not natviely supported with TensorRT, call the following
22+
23+
!!! note
24+
25+
Please note, this currently only includes the interpolate plugin. This plugin requires PyTorch 1.3+ for serialization.
26+
27+
```bash
28+
git clone https://github.com/NVIDIA-AI-IOT/torch2trt
29+
cd torch2trt
30+
sudo python setup.py install --plugins
31+
```
32+

‎docs/images/chart.svg‎

Lines changed: 1 addition & 0 deletions
Loading

0 commit comments

Comments
 (0)