Skip to content

Commit 01750c8

Browse files
xiaozaixiaozai
committed
creat the repository for DeT
1 parent f0b6ab8 commit 01750c8

File tree

7 files changed

+1454
-0
lines changed

7 files changed

+1454
-0
lines changed

‎INSTALL.md

Lines changed: 96 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,96 @@
1+
# Installation
2+
3+
This document contains detailed instructions for installing the necessary dependencies for PyTracking. The instrustions have been tested on an Ubuntu 18.04 system. We recommend using the [install script](install.sh) if you have not already tried that.
4+
5+
### Requirements
6+
* Conda installation with Python 3.7. If not already installed, install from https://www.anaconda.com/distribution/.
7+
* Nvidia GPU.
8+
9+
## Step-by-step instructions
10+
#### Create and activate a conda environment
11+
```bash
12+
conda create --name pytracking python=3.7
13+
conda activate pytracking
14+
```
15+
16+
#### Install PyTorch
17+
Install PyTorch with cuda10.
18+
```bash
19+
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
20+
```
21+
22+
**Note:**
23+
- It is possible to use any PyTorch supported version of CUDA (not necessarily v10).
24+
- For more details about PyTorch installation, see https://pytorch.org/get-started/previous-versions/.
25+
26+
#### Install matplotlib, pandas, tqdm, opencv, scikit-image, visdom, tikzplotlib, gdown, and tensorboad
27+
```bash
28+
conda install matplotlib pandas tqdm
29+
pip install opencv-python visdom tb-nightly scikit-image tikzplotlib gdown
30+
```
31+
32+
33+
#### Install the coco and lvis toolkits
34+
```bash
35+
conda install cython
36+
pip install pycocotools
37+
pip install lvis
38+
```
39+
40+
41+
#### Install ninja-build for Precise ROI pooling
42+
To compile the Precise ROI pooling module (https://github.com/vacancy/PreciseRoIPooling), you may additionally have to install ninja-build.
43+
```bash
44+
sudo apt-get install ninja-build
45+
```
46+
In case of issues, we refer to https://github.com/vacancy/PreciseRoIPooling.
47+
48+
49+
#### Install spatial-correlation-sampler (only required for KYS tracker)
50+
```bash
51+
pip install spatial-correlation-sampler
52+
```
53+
In case of issues, we refer to https://github.com/ClementPinard/Pytorch-Correlation-extension.
54+
55+
#### Install jpeg4py
56+
In order to use [jpeg4py](https://github.com/ajkxyz/jpeg4py) for loading the images instead of OpenCV's imread(), install jpeg4py in the following way,
57+
```bash
58+
sudo apt-get install libturbojpeg
59+
pip install jpeg4py
60+
```
61+
62+
**Note:** The first step (```sudo apt-get install libturbojpeg```) can be optionally ignored, in which case OpenCV's imread() will be used to read the images. However the second step is a must.
63+
64+
In case of issues, we refer to https://github.com/ajkxyz/jpeg4py.
65+
66+
67+
#### Setup the environment
68+
Create the default environment setting files.
69+
```bash
70+
# Environment settings for pytracking. Saved at pytracking/evaluation/local.py
71+
python -c "from pytracking.evaluation.environment import create_default_local_file; create_default_local_file()"
72+
73+
# Environment settings for ltr. Saved at ltr/admin/local.py
74+
python -c "from ltr.admin.environment import create_default_local_file; create_default_local_file()"
75+
```
76+
77+
You can modify these files to set the paths to datasets, results paths etc.
78+
79+
80+
#### Download the pre-trained networks
81+
You can download the pre-trained networks from the [google drive folder](https://drive.google.com/drive/folders/1WVhJqvdu-_JG1U-V0IqfxTUa1SBPnL0O).
82+
The networks shoud be saved in the directory set by "network_path" in "pytracking/evaluation/local.py". By default, it is set to
83+
pytracking/networks.
84+
You can also download the networks using the [gdown](https://github.com/wkentaro/gdown) python package.
85+
86+
```bash
87+
# Download the default network for DiMP-50 and DiMP-18
88+
gdown https://drive.google.com/uc\?id\=1qgachgqks2UGjKx-GdO1qylBDdB1f9KN -O pytracking/networks/dimp50.pth
89+
gdown https://drive.google.com/uc\?id\=1MAjrRJDCbL0DSjUKFyDkUuYS1-cYBNjk -O pytracking/networks/dimp18.pth
90+
91+
# Download the default network for ATOM
92+
gdown https://drive.google.com/uc\?id\=1VNyr-Ds0khjM0zaq6lU-xfY74-iWxBvU -O pytracking/networks/atom_default.pth
93+
94+
# Download the default network for ECO
95+
gdown https://drive.google.com/uc\?id\=1aWC4waLv_te-BULoy0k-n_zS-ONms21S -O pytracking/networks/resnet18_vggmconv1.pth
96+
```

‎INSTALL_win.md

Lines changed: 232 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,232 @@
1+
# Windows Installation
2+
3+
This document contains detailed instructions for installing the necessary dependencies for PyTracking on Windows. The instructions have been tested on a Windows 10 system with Visual Studio 2015. **Notice that Windows installation is much more complex. [Installation on Linux (Ubuntu) is highly recommended.](INSTALL.md)**
4+
5+
### Requirements
6+
* Conda 64 installation with Python 3.7. If not already installed, install from https://www.anaconda.com/distribution/.
7+
* Nvidia GPU.
8+
* Visual Studio 2015 or newer.
9+
* Pre install CUDA 10.0 (not necessarily v10) with VS support.
10+
11+
## Step-by-step instructions
12+
#### Create and activate a conda environment
13+
```bash
14+
conda create --name pytracking python=3.7
15+
conda activate pytracking
16+
```
17+
18+
#### Install PyTorch
19+
Install PyTorch with cuda10.
20+
```bash
21+
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
22+
```
23+
24+
**Note:**
25+
- It is possible to use any PyTorch supported version of CUDA (not necessarily v10), but better be the same version with your preinstalled CUDA (if you have one)
26+
- For more details about PyTorch installation, see https://pytorch.org/get-started/previous-versions/.
27+
28+
#### Install matplotlib, pandas, opencv, visdom and tensorboad
29+
```bash
30+
conda install matplotlib pandas
31+
pip install opencv-python visdom tb-nightly
32+
```
33+
34+
35+
#### Install the coco toolkit
36+
If you want to use COCO dataset for training, install the coco python toolkit. You additionally need to install cython to compile the coco toolkit.
37+
```bash
38+
conda install cython
39+
pip install pycocotools
40+
```
41+
42+
#### Install Precise ROI pooling
43+
44+
This is thecomplicated part. There are two options:
45+
46+
##### Install pre-build Precise ROI pooling package
47+
48+
DiMP and ATOM trackers need Precise ROI pooling module (https://github.com/vacancy/PreciseRoIPooling). You can download the [pre-build binary file](https://visionml.github.io/dimp/prroi_pool.pyd) (build on Windows 10) and install it. Or you could build your own package by following [Build Precise ROI pooling with Visual Studio (Optional)](#build-precise-roi-pooling-with-visual-studio-optional).
49+
50+
+ The package is built with VS2015, so in some cases (such as you don't have VS2015) you will need to install [Visual C++ Redistributable for Visual Studio 2015](https://www.microsoft.com/en-us/download/details.aspx?id=48145) from Microsoft.
51+
52+
+ Add `Anaconda3\envs\pytracking\Lib\site-packages\torch\lib` to users path (Right click this PC --> Properties --> Advanced System settings --> Environment Variables --> User variables --> Path).
53+
54+
+ Copy the `prroi_pool.pyd` file to the conda environment python path (such as `Anaconda3\envs\pytracking\Lib\site-packages\`). This will take action after restart the shell.
55+
56+
+ Add this code to `pytracking\ltr\external\PreciseRoIPooling\pytorch\prroi_pool\functional.py`:
57+
58+
```python
59+
...
60+
def _import_prroi_pooling():
61+
global _prroi_pooling
62+
63+
#load the prroi_pool module
64+
import imp
65+
file, path, description = imp.find_module('prroi_pool')
66+
with file:
67+
_prroi_pooling = imp.load_module('prroi_pool', file, path, description)
68+
...
69+
```
70+
71+
which should then look like:
72+
73+
```python
74+
import torch
75+
import torch.autograd as ag
76+
77+
__all__ = ['prroi_pool2d']
78+
79+
_prroi_pooling = None
80+
81+
def _import_prroi_pooling():
82+
global _prroi_pooling
83+
84+
#load the prroi_pool module
85+
import imp
86+
file, path, description = imp.find_module('prroi_pool')
87+
with file:
88+
_prroi_pooling = imp.load_module('prroi_pool', file, path, description)
89+
90+
if _prroi_pooling is None:
91+
try:
92+
from os.path import join as pjoin, dirname
93+
from torch.utils.cpp_extension import load as load_extension
94+
root_dir = pjoin(dirname(__file__), 'src')
95+
96+
_prroi_pooling = load_extension(
97+
'_prroi_pooling',
98+
[pjoin(root_dir, 'prroi_pooling_gpu.cpp'), pjoin(root_dir, 'prroi_pooling_gpu_impl.cu')],
99+
verbose=True
100+
)
101+
except ImportError:
102+
raise ImportError('Can not compile Precise RoI Pooling library.')
103+
104+
return _prroi_pooling
105+
...
106+
```
107+
108+
+ If the pre-build package don't work on your platform, you can build your own package as described in the next section.
109+
110+
##### Build Precise ROI pooling with Visual Studio (Optional)
111+
112+
To compile the Precise ROI pooling module (https://github.com/vacancy/PreciseRoIPooling) on Windows, you need Visual Studio with CUDA installed.
113+
114+
+ First make a DLL project by the following step.
115+
116+
1. Download the Precise ROI pooling module with `git clone https://github.com/vacancy/PreciseRoIPooling `.
117+
2. Download pybind11 `git clone https://github.com/pybind/pybind11 `
118+
3. Open Visual Studio and start a new C++ `Empty project`.
119+
4. Add `PreciseRoIPooling\src\prroi_pooling_gpu_impl.cu` and `PreciseRoIPooling\pytorch\prroi_pool\src\prroi_pooling_gpu.c` to the `Source File` and change the name `prroi_pooling_gpu.c` to `prroi_pooling_gpu.cpp`.
120+
5. Add `PreciseRoIPooling\src\prroi_pooling_gpu_impl.cuh` and `PreciseRoIPooling\pytorch\prroi_pool\src\prroi_pooling_gpu.h` to the `Header File`.
121+
6. Right click the project --> Property. **Change Configuration to `Release` and `x64`**.
122+
Then Configuration Properties --> General --> change Configuration Type to `.dll` and Target Extension to `.pyd` .
123+
124+
+ Set the VC++ Directories.
125+
126+
1. Find the following dirs and add them to VC++ Directories --> Include Directories.
127+
```
128+
Anaconda3\envs\pytracking\Lib\site-packages\torch\include\torch\csrc\api\include
129+
Anaconda3\envs\pytracking\Lib\site-packages\torch\include\THC
130+
Anaconda3\envs\pytracking\Lib\site-packages\torch\include\TH
131+
Anaconda3\envs\pytracking\Lib\site-packages\torch\include
132+
Anaconda3\envs\pytracking\include
133+
CUDA\v10.0\include
134+
pybind11\pybind11\include
135+
```
136+
137+
2. Find the following dirs and add them to VC++ Directories --> Lib Directories.
138+
139+
```
140+
Anaconda3\envs\pytracking\Lib\site-packages\torch\lib
141+
Anaconda3\envs\pytracking\libs
142+
```
143+
144+
+ Set the Linker.
145+
146+
1. Find and add them to Linker --> General -->Additional Library Directories.
147+
148+
```
149+
CUDA\v10.0\lib\x64
150+
Anaconda3\envs\pytracking\libs
151+
Anaconda3\envs\pytracking\Lib\site-packages\torch\lib
152+
```
153+
154+
2. Add them to Linker --> Input -->Additional Dependencies
155+
156+
```
157+
python37.lib
158+
python3.lib
159+
cudart.lib
160+
c10.lib
161+
torch.lib
162+
torch_python.lib
163+
_C.lib
164+
c10_cuda.lib
165+
```
166+
167+
+ Set the CUDA dependence.
168+
169+
1. Right click the project --> Build dependencies --> Build Customizations --> click CUDA
170+
2. Right click the `*.cu` and `*.cuh` files --> Property. And change the type from `C/C++` to `CUDA C/C++`
171+
172+
+ Set the package name and build.
173+
174+
Change `prroi_pooling_gpu.cpp` file in `line 109`
175+
176+
from
177+
178+
```
179+
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
180+
```
181+
182+
to
183+
184+
```
185+
PYBIND11_MODULE(prroi_pool, m) {
186+
```
187+
188+
then build the package with **`Release` and `x64`**. You will get a `*.pyd` file. Rename it as `prroi_pool.pyd`.
189+
190+
+ Last but not least, follow the step in [Install pre-build Precise ROI pooling package](#install-pre-build-precise-roi-pooling-package).
191+
192+
In case of issues, we refer to https://github.com/vacancy/PreciseRoIPooling.
193+
194+
#### Install jpeg4py
195+
In order to use [jpeg4py](https://github.com/ajkxyz/jpeg4py) for loading the images instead of OpenCV's imread(), install jpeg4py in the following way,
196+
```bash
197+
pip install jpeg4py
198+
```
199+
200+
In case of issues, we refer to https://github.com/ajkxyz/jpeg4py.
201+
202+
203+
#### Setup the environment
204+
Create the default environment setting files.
205+
```bash
206+
# Environment settings for pytracking. Saved at pytracking/evaluation/local.py
207+
python -c "from pytracking.evaluation.environment import create_default_local_file; create_default_local_file()"
208+
209+
# Environment settings for ltr. Saved at ltr/admin/local.py
210+
python -c "from ltr.admin.environment import create_default_local_file; create_default_local_file()"
211+
```
212+
213+
You can modify these files to set the paths to datasets, results paths etc.
214+
215+
216+
#### Download the pre-trained networks
217+
You can download the pre-trained networks from the [google drive folder](https://drive.google.com/drive/folders/1WVhJqvdu-_JG1U-V0IqfxTUa1SBPnL0O).
218+
The networks shoud be saved in the directory set by "network_path" in "pytracking/evaluation/local.py". By default, it is set to
219+
pytracking/networks.
220+
You should download them manually and copy to the correct directory.
221+
222+
```bash
223+
# directory of the default network for DiMP-50 and DiMP-18
224+
pytracking/networks/dimp50.pth
225+
pytracking/networks/dimp18.pth
226+
227+
# directory of the default network for ATOM
228+
pytracking/networks/atom_default.pth
229+
230+
# directory of the default network for ECO
231+
pytracking/networks/resnet18_vggmconv1.pth
232+
```

0 commit comments

Comments
 (0)