Mirrors/blakeblackshear.frigate
1
0
Fork 0
mirror of https://github.com/blakeblackshear/frigate.git synced 2024-11-21 19:07:46 +01:00
Code Issues Projects Releases Wiki Activity

Reimplement support for rknn detector (#11365)

Browse Source 
* initial support for rknn detector

* remove purge_model_cache option

* update rknn

* support rk3576

* fix post_process_yolonas call

* add yolonas models

* update config

* exclude yolonas from image

* remove code
This commit is contained in:
Marc Altmann 2024-05-22 00:50:03 +02:00 committed by GitHub
parent 910c85b1c0
commit e91f3d8d9b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
7 changed files with 231 additions and 90 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
.github/workflows/ci.yml vendored
View File

@ -69,15 +69,15 @@ jobs:
rpi.tags=${{ steps.setup.outputs.image-name }}-rpi
*.cache-from=type=registry,ref=${{ steps.setup.outputs.cache-name }}-arm64
*.cache-to=type=registry,ref=${{ steps.setup.outputs.cache-name }}-arm64,mode=max
#- name: Build and push RockChip build
# uses: docker/bake-action@v3
# with:
# push: true
# targets: rk
# files: docker/rockchip/rk.hcl
# set: |
# rk.tags=${{ steps.setup.outputs.image-name }}-rk
# *.cache-from=type=gha
- name: Build and push Rockchip build
uses: docker/bake-action@v3
with:
push: true
targets: rk
files: docker/rockchip/rk.hcl
set: |
rk.tags=${{ steps.setup.outputs.image-name }}-rk
*.cache-from=type=gha
jetson_jp4_build:
runs-on: ubuntu-latest
name: Jetson Jetpack 4

5
docker/rockchip/Dockerfile
View File

@ -18,10 +18,7 @@ RUN --mount=type=bind,from=rk-wheels,source=/rk-wheels,target=/deps/rk-wheels \
WORKDIR /opt/frigate/
COPY --from=rootfs / /
ADD https://github.com/MarcA711/rknpu2/releases/download/v1.5.2/librknnrt_rk356x.so /usr/lib/
ADD https://github.com/MarcA711/rknpu2/releases/download/v1.5.2/librknnrt_rk3588.so /usr/lib/
# TODO removed models, other models support may need to be added back in
ADD https://github.com/MarcA711/rknn-toolkit2/releases/download/v2.0.0/librknnrt.so /usr/lib/
RUN rm -rf /usr/lib/btbn-ffmpeg/bin/ffmpeg
RUN rm -rf /usr/lib/btbn-ffmpeg/bin/ffprobe

3
docker/rockchip/requirements-wheels-rk.txt
View File

@ -1,2 +1 @@
hide-warnings == 0.17
rknn-toolkit-lite2 @ https://github.com/MarcA711/rknn-toolkit2/releases/download/v1.5.2/rknn_toolkit_lite2-1.5.2-cp39-cp39-linux_aarch64.whl
rknn-toolkit-lite2 @ https://github.com/MarcA711/rknn-toolkit2/releases/download/v2.0.0/rknn_toolkit_lite2-2.0.0b0-cp39-cp39-linux_aarch64.whl

2
docs/docs/configuration/hardware_acceleration.md
View File

@ -366,7 +366,7 @@ Hardware accelerated video de-/encoding is supported on all Rockchip SoCs using
### Prerequisites
Make sure that you use a linux distribution that comes with the rockchip BSP kernel 5.10 or 6.1 and supports VPU. To check, enter the following commands:
Make sure that you use a linux distribution that comes with the rockchip BSP kernel 5.10 or 6.1 and rkvdec2 driver. To check, enter the following commands:
```
$ uname -r

106
docs/docs/configuration/object_detectors.md
View File

@ -302,3 +302,109 @@ Replace `<your_codeproject_ai_server_ip>` and `<port>` with the IP address and p
To verify that the integration is working correctly, start Frigate and observe the logs for any error messages related to CodeProject.AI. Additionally, you can check the Frigate web interface to see if the objects detected by CodeProject.AI are being displayed and tracked properly.
# Community Supported Detectors
## Rockchip platform
Hardware accelerated object detection is supported on the following SoCs:
- RK3562
- RK3566
- RK3568
- RK3576
- RK3588
This implementation uses the [Rockchip's RKNN-Toolkit2](https://github.com/airockchip/rknn-toolkit2/) Currently, only [Yolo-NAS](https://github.com/Deci-AI/super-gradients/blob/master/YOLONAS.md) is supported as object detection model.
### Prerequisites
Make sure that you use a linux distribution that comes with the rockchip BSP kernel 5.10 or 6.1 and rknpu driver. To check, enter the following commands:
```
$ uname -r
5.10.xxx-rockchip # or 6.1.xxx; the -rockchip suffix is important
$ ls /dev/dri
by-path card0 card1 renderD128 renderD129 # should list renderD129
$ sudo cat /sys/kernel/debug/rknpu/version
RKNPU driver: v0.9.2 # or later version
```
I recommend [Joshua Riek's Ubuntu for Rockchip](https://github.com/Joshua-Riek/ubuntu-rockchip), if your board is supported.
### Setup
Follow Frigate's default installation instructions, but use a docker image with `-rk` suffix for example `ghcr.io/blakeblackshear/frigate:stable-rk`.
Next, you need to grant docker permissions to access your hardware:
- During the configuration process, you should run docker in privileged mode to avoid any errors due to insufficient permissions. To do so, add `privileged: true` to your `docker-compose.yml` file or the `--privileged` flag to your docker run command.
- After everything works, you should only grant necessary permissions to increase security. Add the lines below to your `docker-compose.yml` file or the following options to your docker run command: `--security-opt systempaths=unconfined --security-opt apparmor=unconfined --device /dev/dri:/dev/dri`:
```yaml
security_opt:
- apparmor=unconfined
- systempaths=unconfined
devices:
- /dev/dri:/dev/dri
```
### Configuration
This `config.yml` shows all relevant options to configure the detector and explains them. All values shown are the default values (except for two). Lines that are required at least to use the detector are labeled as required, all other lines are optional.
```yaml
detectors: # required
rknn: # required
type: rknn # required
# number of NPU cores to use
# 0 means choose automatically
# increase for better performance if you have a multicore NPU e.g. set to 3 on rk3588
num_cores: 0
model: # required
# name of model (will be automatically downloaded) or path to your own .rknn model file
# possible values are:
# - deci-fp16-yolonas_s
# - deci-fp16-yolonas_m
# - deci-fp16-yolonas_l
# - /config/model_cache/your_custom_model.rknn
path: deci-fp16-yolonas_s
# width and height of detection frames
width: 320
height: 320
# pixel format of detection frame
# default value is rgb but yolo models usually use bgr format
input_pixel_format: bgr # required
# shape of detection frame
input_tensor: nhwc
```
### Choosing a model
:::warning
yolo-nas models use weights from DeciAI. These weights are subject to their license and can't be used commercially. For more information, see: https://docs.deci.ai/super-gradients/latest/LICENSE.YOLONAS.html
:::
The inference time was determined on a rk3588 with 3 NPU cores.
| Model | Size in mb | Inference time in ms |
| ------------------- | ---------- | -------------------- |
| deci-fp16-yolonas_s | 24 | 25 |
| deci-fp16-yolonas_m | 62 | 35 |
| deci-fp16-yolonas_l | 81 | 45 |
:::tip
You can get the load of your NPU with the following command:
```bash
$ cat /sys/kernel/debug/rknpu/load
>> NPU load: Core0: 0%, Core1: 0%, Core2: 0%,
```
:::
- By default the rknn detector uses the yolonas_s model (`model: path: default-fp16-yolonas_s`). This model comes with the image, so no further steps than those mentioned above are necessary and no download happens.
- The other choices are automatically downloaded and stored in the folder `config/model_cache/rknn_cache`. After upgrading Frigate, you should remove older models to free up space.
- Finally, you can also provide your own `.rknn` model. You should not save your own models in the `rknn_cache` folder, store them directly in the `model_cache` folder or another subfolder. To convert a model to `.rknn` format see the `rknn-toolkit2` (requires a x86 machine). Note, that there is only post-processing for the supported models.

2
frigate/detectors/detection_api.py
View File

@ -72,7 +72,7 @@ class DetectionApi(ABC):
def post_process(self, output):
if self.detector_config.model.model_type == ModelTypeEnum.yolonas:
return self.yolonas(output)
return self.post_process_yolonas(output)
else:
raise ValueError(
f'Model type "{self.detector_config.model.model_type}" is currently not supported.'

185
frigate/detectors/plugins/rknn.py
View File

@ -1,118 +1,157 @@
import logging
import os.path
import re
import urllib.request
from typing import Literal
try:
from hide_warnings import hide_warnings
except: # noqa: E722
def hide_warnings(func):
pass
from pydantic import Field
from frigate.detectors.detection_api import DetectionApi
from frigate.detectors.detector_config import BaseDetectorConfig
from frigate.detectors.detector_config import BaseDetectorConfig, ModelTypeEnum
logger = logging.getLogger(__name__)
DETECTOR_KEY = "rknn"
supported_socs = ["rk3562", "rk3566", "rk3568", "rk3588"]
supported_socs = ["rk3562", "rk3566", "rk3568", "rk3576", "rk3588"]
supported_models = {ModelTypeEnum.yolonas: "^deci-fp16-yolonas_[sml]$"}
model_chache_dir = "/config/model_cache/rknn_cache/"
class RknnDetectorConfig(BaseDetectorConfig):
type: Literal[DETECTOR_KEY]
core_mask: int = Field(default=0, ge=0, le=7, title="Core mask for NPU.")
num_cores: int = Field(default=0, ge=0, le=3, title="Number of NPU cores to use.")
purge_model_cache: bool = Field(default=True)
class Rknn(DetectionApi):
type_key = DETECTOR_KEY
def __init__(self, config: RknnDetectorConfig):
# find out SoC
try:
with open("/proc/device-tree/compatible") as file:
soc = file.read().split(",")[-1].strip("\x00")
except FileNotFoundError:
logger.error("Make sure to run docker in privileged mode.")
raise Exception("Make sure to run docker in privileged mode.")
if soc not in supported_socs:
logger.error(
"Your SoC is not supported. Your SoC is: {}. Currently these SoCs are supported: {}.".format(
soc, supported_socs
)
)
raise Exception(
"Your SoC is not supported. Your SoC is: {}. Currently these SoCs are supported: {}.".format(
soc, supported_socs
)
)
if not os.path.isfile("/usr/lib/librknnrt.so"):
if "rk356" in soc:
os.rename("/usr/lib/librknnrt_rk356x.so", "/usr/lib/librknnrt.so")
elif "rk3588" in soc:
os.rename("/usr/lib/librknnrt_rk3588.so", "/usr/lib/librknnrt.so")
self.core_mask = config.core_mask
super().__init__(config)
self.height = config.model.height
self.width = config.model.width
core_mask = 2**config.num_cores - 1
soc = self.get_soc()
if True:
os.makedirs("/config/model_cache/rknn", exist_ok=True)
model_props = self.parse_model_input(config.model.path, soc)
if (config.model.width != 320) or (config.model.height != 320):
logger.error(
"Make sure to set the model width and height to 320 in your config.yml."
)
raise Exception(
"Make sure to set the model width and height to 320 in your config.yml."
)
if model_props["preset"]:
config.model.model_type = model_props["model_type"]
if config.model.input_pixel_format != "bgr":
logger.error(
'Make sure to set the model input_pixel_format to "bgr" in your config.yml.'
)
raise Exception(
'Make sure to set the model input_pixel_format to "bgr" in your config.yml.'
)
if config.model.input_tensor != "nhwc":
logger.error(
'Make sure to set the model input_tensor to "nhwc" in your config.yml.'
)
raise Exception(
'Make sure to set the model input_tensor to "nhwc" in your config.yml.'
)
if model_props["model_type"] == ModelTypeEnum.yolonas:
logger.info("""
You are using yolo-nas with weights from DeciAI.
These weights are subject to their license and can't be used commercially.
For more information, see: https://docs.deci.ai/super-gradients/latest/LICENSE.YOLONAS.html
""")
from rknnlite.api import RKNNLite
self.rknn = RKNNLite(verbose=False)
if self.rknn.load_rknn(self.model_path) != 0:
if self.rknn.load_rknn(model_props["path"]) != 0:
logger.error("Error initializing rknn model.")
if self.rknn.init_runtime(core_mask=self.core_mask) != 0:
if self.rknn.init_runtime(core_mask=core_mask) != 0:
logger.error(
"Error initializing rknn runtime. Do you run docker in privileged mode?"
)
raise Exception(
"RKNN does not currently support any models. Please see the docs for more info."
)
def __del__(self):
self.rknn.release()
@hide_warnings
def inference(self, tensor_input):
return self.rknn.inference(inputs=tensor_input)
def get_soc(self):
try:
with open("/proc/device-tree/compatible") as file:
soc = file.read().split(",")[-1].strip("\x00")
except FileNotFoundError:
raise Exception("Make sure to run docker in privileged mode.")
if soc not in supported_socs:
raise Exception(
f"Your SoC is not supported. Your SoC is: {soc}. Currently these SoCs are supported: {supported_socs}."
)
return soc
def parse_model_input(self, model_path, soc):
model_props = {}
# find out if user provides his own model
# user provided models should be a path and contain a "/"
if "/" in model_path:
model_props["preset"] = False
model_props["path"] = model_path
else:
model_props["preset"] = True
"""
Filenames follow this pattern:
origin-quant-basename-soc-tk_version-rev.rknn
origin: From where comes the model? default: upstream repo; rknn: modifications from airockchip
quant: i8 or fp16
basename: e.g. yolonas_s
soc: e.g. rk3588
tk_version: e.g. v2.0.0
rev: e.g. 1
Full name could be: default-fp16-yolonas_s-rk3588-v2.0.0-1.rknn
"""
model_matched = False
for model_type, pattern in supported_models.items():
if re.match(pattern, model_path):
model_matched = True
model_props["model_type"] = model_type
if model_matched:
model_props["filename"] = model_path + f"-{soc}-v2.0.0-1.rknn"
model_props["path"] = model_chache_dir + model_props["filename"]
if not os.path.isfile(model_props["path"]):
self.download_model(model_props["filename"])
else:
supported_models_str = ", ".join(
model[1:-1] for model in supported_models
)
raise Exception(
f"Model {model_path} is unsupported. Provide your own model or choose one of the following: {supported_models_str}"
)
return model_props
def download_model(self, filename):
if not os.path.isdir(model_chache_dir):
os.mkdir(model_chache_dir)
urllib.request.urlretrieve(
f"https://github.com/MarcA711/rknn-models/releases/download/v2.0.0/{filename}",
model_chache_dir + filename,
)
def check_config(self, config):
if (config.model.width != 320) or (config.model.height != 320):
raise Exception(
"Make sure to set the model width and height to 320 in your config.yml."
)
if config.model.input_pixel_format != "bgr":
raise Exception(
'Make sure to set the model input_pixel_format to "bgr" in your config.yml.'
)
if config.model.input_tensor != "nhwc":
raise Exception(
'Make sure to set the model input_tensor to "nhwc" in your config.yml.'
)
def detect_raw(self, tensor_input):
output = self.inference(
output = self.rknn.inference(
[
tensor_input,
]
)
return self.postprocess(output[0])
return self.post_process(output)