This guide provides resources for DeepStream application development in Python.
- Prerequisites
- Running Sample Applications
- Pipeline Construction
- MetaData Access
- Image Data Access
- Custom Inference Output Parsing
- FAQ and Troubleshooting
- Ubuntu 22.04
- DeepStream SDK 7.0 or later
- Python 3.10
- Gst Python v1.20.3
Gst python should be already installed on Jetson.
If missing, install with the following steps:
$ sudo apt-get install python-gi-dev
$ export GST_LIBS="-lgstreamer-1.0 -lgobject-2.0 -lglib-2.0"
$ export GST_CFLAGS="-pthread -I/usr/include/gstreamer-1.0 -I/usr/include/glib-2.0 -I/usr/lib/x86_64-linux-gnu/glib-2.0/include"
$ git clone https://github.com/GStreamer/gst-python.git
$ cd gst-python
$ git checkout 5343aeb
$ ./autogen.sh PYTHON=python3
$ ./configure PYTHON=python3
$ make
$ sudo make install
The prebuilt DeepStreamSDK python bindings for both x86 and Jetson are already available on the release section. The readme here provides instructions to customize the bindings or recompile them, if you need to.
Note: Compiling bindings now also generates a pip installable python wheel for the platform (x86 or aarch64) it is compiled on.
Clone the deepstream_python_apps repo under /sources: git clone https://github.com/NVIDIA-AI-IOT/deepstream_python_apps
This will create the following directory:
<DeepStream ROOT>/sources/deepstream_python_apps
The Python apps are under the "apps" directory.
Go into each app directory and follow instructions in the README.
NOTE: The app configuration files contain relative paths for models.
DeepStream pipelines can be constructed using Gst Python, the GStreamer framework's Python bindings.
See sample applications main functions for pipeline construction examples.
DeepStream MetaData contains inference results and other information used in analytics. The MetaData is attached to the Gst Buffer received by each pipeline component. The metadata format is described in detail in the SDK MetaData documentation and API Guide.
The SDK MetaData library is developed in C/C++. Python bindings provide access to the MetaData from Python applications. The bindings are provided as part of this repository here and can be compiled natively for x86_64 and Jetson platforms. This module, pyds.so, can also be cross-compiled for aarch64 on x86 host by using Qemu emulator. Dockerfile for the cross-compilation is provided here
The sample applications gets the import path for this module through common/utils.py. A setup.py is also included for installing the module into standard path:
cd /opt/nvidia/deepstream/deepstream/lib
python3 setup.py install
This is currently not automatically done through the SDK installer because python usage is optional.
The bindings generally follow the same API as the underlying C/C++ library, with a few exceptions detailed in sections below.
Memory for MetaData is shared by the Python and C/C++ code paths. For example, a MetaData item may be added by a probe function written in Python, and needs to be accessed by a downstream plugin written in C/C++. The deepstream-test4 app contains such usage. The Python garbage collector does not have visibility into memory references in C/C++, and therefore cannot safely manage the lifetime of such shared memory. Because of this complication, Python access to MetaData memory is typically achieved via references without claiming ownership.
When MetaData objects are allocated in Python, an allocation function is provided by the bindings to ensure proper memory ownership of the object. If the constructor is used, the object will be claimed by the garbage collector when its Python references terminate. However, the object will still need to be accessed by C/C++ code downstream, and therefore must persist beyond those Python references.
Example: To allocate an NvDsEventMsgMeta instance, use this:
msg_meta = pyds.alloc_nvds_event_msg_meta(user_event_meta) # get reference to allocated instance without claiming memory ownership
NOT this:
msg_meta = NvDsEventMsgMeta() # memory will be freed by the garbage collector when msg_meta goes out of scope in Python
Allocators are available for the following structs:
NvDsVehicleObject: alloc_nvds_vehicle_object()
NvDsPersonObject: alloc_nvds_person_object()
NvDsFaceObject: alloc_nvds_face_object()
NvDsEventMsgMeta: alloc_nvds_event_msg_meta()
NvDsEvent: alloc_nvds_event()
NvDsPayload: alloc_nvds_payload()
Generic buffer: alloc_buffer(size)
Some MetaData structures contain string fields. These are accessable in the following manner:
Setting a string field results in the allocation of a string buffer in the underlying C++ code.
obj.type = "Type"
This will cause a memory buffer to be allocated, and the string "TYPE" will be copied into it.
This memory is owned by the C code and will be freed later. To free the buffer in Python code, use:
pyds.free_buffer(obj.type)
NOTE: NvOSD_TextParams.display_text string now gets freed automatically when a new string is assigned.
Directly reading a string field returns C address of the field in the form of an int, e.g.:
obj = pyds.NvDsVehicleObject.cast(data);
print(obj.type)
This will print an int representing the address of obj.type in C (which is a char*).
To retrieve the string value of this field, use pyds.get_string()
, e.g.:
print(pyds.get_string(obj.type))
Some MetaData instances are stored in GList form. To access the data in a GList node, the data field needs to be cast to the appropriate structure. This casting is done via cast() member function for the target type:
NvDsBatchMeta.cast
NvDsFrameMeta.cast
NvDsObjectMeta.cast
NvDsUserMeta.cast
NvDsClassifierMeta.cast
NvDsDisplayMeta.cast
NvDsLabelInfo.cast
NvDsEventMsgMeta.cast
NvDsVehicleObject.cast
NvDsPersonObject.cast
In version v0.5, standalone cast functions were provided. Those are now deprecated and superseded by the cast() functions above:
glist_get_nvds_batch_meta
glist_get_nvds_frame_meta
glist_get_nvds_object_meta
glist_get_nvds_user_meta
glist_get_nvds_classifier_meta
glist_get_nvds_display_meta
glist_get_nvds_label_info
glist_get_nvds_event_msg_meta
glist_get_nvds_vehicle_object
glist_get_nvds_person_object
Example:
l_frame = batch_meta.frame_meta_list
frame_meta = pyds.NvDsFrameMeta.cast(l_frame.data)
Custom MetaData added to NvDsUserMeta require custom copy and release functions. The MetaData library relies on these custom functions to perform deep-copy of the custom structure, and free allocated resources. These functions are registered as callback function pointers in the NvDsUserMeta structure.
Previously, Callback functions were registered using these functions:
pyds.set_user_copyfunc(NvDsUserMeta_instance, copy_function)
pyds.set_user_releasefunc(NvDsUserMeta_instance, free_func)
These are now DEPRECATED and are replaced by similar implementation in the binding itself. These are event_msg_meta_copy_func() and event_msg_meta_release_func() respectively. These can be found inside bindschema.cpp
NOTE: Previously, callbacks needed to be unregistered with the bindings library before the application exits. The bindings library currently keeps global references to the registered functions, and these cannot last beyond bindings library unload which happens at application exit. Use the following function to unregister all callbacks:
pyds.unset_callback_funcs()
These callbacks are automatically set inside the alloc_nvds_event_msg_meta() function and should NOT be set from the python application (e.g. deepstream-test4)
The deepstream-test4 sample application has been updated to show an example of removal of these callback registration and unregistration.
Limitation: the bindings library currently only supports a single set of callback functions for each application. The last registered function will be used.
Python interpretation is generally slower than running compiled C/C++ code. To provide better performance, some operations are implemented in C and exposed via the bindings interface. This is currently experimental and will expand over time.
The following optimized functions are available:
This is a simple function that performs the same operations as the following:
txt_params.text_bg_clr.red = red
txt_params.text_bg_clr.green = green
txt_params.text_bg_clr.blue = blue
txt_params.text_bg_clr.alpha = alpha
These are performend on each object in deepstream_test_4.py, causing the aggregate processing time to slow down the pipeline. Pushing this function into the C layer helps to increase performance.
This function populates the input buffer with a timestamp generated according to RFC3339:
%Y-%m-%dT%H:%M:%S.nnnZ\0
Decoded images are accessible as NumPy arrays via the get_nvds_buf_surface
function. This function is documented in the API Guide.
Please see the deepstream-imagedata-multistream sample application for an example of image data usage.