이게 ssd 인지, ssd + mobilenet v2 쪽인진 모르겠다.
[링크 : https://stackoverflow.com/questions/67868644/post-process-of-tf2-ssd-detection-models]
| softmax (0) | 2024.01.10 |
|---|---|
| 텐서플로우 학습 (0) | 2024.01.09 |
| 우분투에 jupyter notebook 설치 및 실행하기 (0) | 2024.01.02 |
| 주피터 노트북 프로젝트(?) 실행하기 (0) | 2024.01.02 |
| i.mx8mp gopoint 실행 경로 (0) | 2024.01.02 |
어쩌면 당연한데.. python 으로 짠녀석이니 pip로 설치하면 된다.
| $ pip install notebook $ jupyter notebook |
[링크 : https://jupyter.org/install]
주피터 노트북 실행하면 아래와 같이 웹으로 뜨고 ipynb를 더블클릭으로 열면 끝
아무튼 Run에 Run All Cells 하면 순차적으로 실행된다.
막상실행해보려니 2년전꺼라 패키지가 달라져서 안되는 듯.. 쩝
+
numpy 버전 문제인가.. 1.25.0 미만이어야 하는데 1.26.2라서 에러라..
1.17.3 ~ 1.24.x 면 될테니 적당하게 바꾸고 해봐야지
| pip show numpy pip uninstall numpy pip install numpy==1.16.4 |
[링크 : https://reyrei.tistory.com/m/28]
+
24.01.03
numpy 손 보니 tensorflow.keras 에서 막히는 마법이.. -_-
| 텐서플로우 학습 (0) | 2024.01.09 |
|---|---|
| ssd(single shot detect) 결과 처리 (0) | 2024.01.03 |
| 주피터 노트북 프로젝트(?) 실행하기 (0) | 2024.01.02 |
| i.mx8mp gopoint 실행 경로 (0) | 2024.01.02 |
| tensorflow keras dataset (0) | 2024.01.02 |
ipynb 라는 확장자가 보여서 확인 중
[링크 : https://github.com/saunack/MobileNetv2-SSD]
아나콘다 깔고
거기서 jupiter notebook을 설치하면 된다고
[링크 : https://mananacho.tistory.com/31]
[링크 : https://blog.naver.com/tamiel/221956194782]
커맨드 라인으로는 먼가 복잡한데, 쥬피터 없이 돌리는것도 아니고 무슨 의미가 있나 싶긴하다.
| $ jupyter nbconvert --execute --to notebook lda2.ipynb |
| ssd(single shot detect) 결과 처리 (0) | 2024.01.03 |
|---|---|
| 우분투에 jupyter notebook 설치 및 실행하기 (0) | 2024.01.02 |
| i.mx8mp gopoint 실행 경로 (0) | 2024.01.02 |
| tensorflow keras dataset (0) | 2024.01.02 |
| tensorflow lite / mnist 학습 (0) | 2024.01.02 |
문서를 찾다가 지쳐서 걍 실행하고 인자를 보는걸로..
| root 3019 925 72 06:38 ? 00:00:25 /usr/bin/python3 /home/root/.nxp-demo-experience/scripts/machine_learning/MLDemoLauncher.py detect |
| root@imx8mpevk:~/.nxp-demo-experience/scripts/machine_learning# cat MLDemoLauncher.py #!/usr/bin/env python3 """ Copyright 2021-2023 NXP SPDX-License-Identifier: BSD-2-Clause This script launches the NNStreamer ML Demos using a UI to pick settings. """ import gi import os import sys import glob from gi.repository import Gtk, GLib, Gio gi.require_version("Gtk", "3.0") sys.path.append("/home/root/.nxp-demo-experience/scripts/") import utils class MLLaunch(Gtk.Window): """The GUI window for the ML demo launcher""" def __init__(self, demo): """Creates the UI window""" # Initialization self.demo = demo super().__init__(title=demo) self.set_default_size(450, 200) self.set_resizable(False) # Get platform self.platform = os.uname().nodename # OpenVX graph caching is not available on i.MX 8QuadMax platform. if self.platform != "imx8qmmek" : os.environ["VIV_VX_CACHE_BINARY_GRAPH_DIR"] = "/home/root/.cache/gopoint" os.environ["VIV_VX_ENABLE_CACHE_GRAPH_BINARY"] = "1" # Get widget properties devices = [] if self.demo != "brand" and self.demo != "selfie_nn": if self.platform != "imx93evk": devices.append("Example Video") for device in glob.glob("/dev/video*"): devices.append(device) backends_available = ["CPU"] if ( os.path.exists("/usr/lib/libvx_delegate.so") and self.demo != "pose" and self.demo != "selfie_nn" ): backends_available.insert(1, "GPU") if os.path.exists("/usr/lib/libneuralnetworks.so") and self.demo != "brand" and self.platform != "imx8qmmek": backends_available.insert(0, "NPU") if os.path.exists("/usr/lib/libethosu_delegate.so"): backends_available.insert(0, "NPU") backends_available.pop() displays_available = ["Weston"] colors_available = ["Red", "Green", "Blue", "Black", "White"] demo_modes_available = ["Background Substitution", "Segmentation Mask"] # Create widgets main_grid = Gtk.Grid.new() device_label = Gtk.Label.new("Source") self.device_combo = Gtk.ComboBoxText() backend_label = Gtk.Label.new("Backend") self.backend_combo = Gtk.ComboBoxText() self.display_combo = Gtk.ComboBoxText() self.launch_button = Gtk.Button.new_with_label("Run") self.status_bar = Gtk.Label.new() header = Gtk.HeaderBar() quit_button = Gtk.Button() quit_icon = Gio.ThemedIcon(name="process-stop-symbolic") quit_image = Gtk.Image.new_from_gicon(quit_icon, Gtk.IconSize.BUTTON) separator = Gtk.Separator.new(0) time_title_label = Gtk.Label.new("Video Refresh") self.time_label = Gtk.Label.new("--.-- ms") self.fps_label = Gtk.Label.new("-- FPS") inference_title_label = Gtk.Label.new("Inference Time") self.inference_label = Gtk.Label.new("--.-- ms") self.ips_label = Gtk.Label.new("-- IPS") if self.demo != "selfie_nn": self.width_entry = self.r_scale = Gtk.Scale.new_with_range( Gtk.Orientation.HORIZONTAL, 300, 1920, 2 ) self.height_entry = self.r_scale = Gtk.Scale.new_with_range( Gtk.Orientation.HORIZONTAL, 300, 1080, 2 ) self.width_label = Gtk.Label.new("Height") self.height_label = Gtk.Label.new("Width") self.color_label = Gtk.Label.new("Label Color") else: self.color_label = Gtk.Label.new("Text Color") self.demo_mode = Gtk.Label.new("Demo Mode") self.mode_combo = Gtk.ComboBoxText() self.color_combo = Gtk.ComboBoxText() # Organize widgets self.add(main_grid) self.set_titlebar(header) quit_button.add(quit_image) header.pack_end(quit_button) main_grid.set_row_spacing(10) main_grid.set_border_width(10) main_grid.attach(device_label, 0, 1, 2, 1) device_label.set_hexpand(True) main_grid.attach(backend_label, 0, 2, 2, 1) # main_grid.attach(display_label, 0, 3, 2, 1) if self.demo != "selfie_nn": main_grid.attach(self.width_label, 0, 4, 2, 1) main_grid.attach(self.height_label, 0, 5, 2, 1) main_grid.attach(self.color_label, 0, 6, 2, 1) else: main_grid.attach(self.demo_mode, 0, 4, 2, 1) main_grid.attach(self.color_label, 0, 5, 2, 1) main_grid.attach(self.device_combo, 2, 1, 2, 1) self.device_combo.set_hexpand(True) main_grid.attach(self.backend_combo, 2, 2, 2, 1) # main_grid.attach(self.display_combo, 2, 3, 2, 1) if self.demo != "selfie_nn": main_grid.attach(self.width_entry, 2, 4, 2, 1) main_grid.attach(self.height_entry, 2, 5, 2, 1) main_grid.attach(self.color_combo, 2, 6, 2, 1) else: main_grid.attach(self.mode_combo, 2, 4, 2, 1) main_grid.attach(self.color_combo, 2, 5, 2, 1) main_grid.attach(self.launch_button, 0, 7, 4, 1) main_grid.attach(self.status_bar, 0, 8, 4, 1) main_grid.attach(separator, 0, 9, 4, 1) main_grid.attach(time_title_label, 0, 10, 2, 1) main_grid.attach(self.time_label, 0, 11, 1, 1) main_grid.attach(self.fps_label, 1, 11, 1, 1) main_grid.attach(inference_title_label, 2, 10, 2, 1) main_grid.attach(self.inference_label, 2, 11, 1, 1) main_grid.attach(self.ips_label, 3, 11, 1, 1) # Configure widgets for device in devices: self.device_combo.append_text(device) for backend in backends_available: self.backend_combo.append_text(backend) for display in displays_available: self.display_combo.append_text(display) for color in colors_available: self.color_combo.append_text(color) if self.demo == "selfie_nn": for mode in demo_modes_available: self.mode_combo.append_text(mode) self.device_combo.set_active(0) self.backend_combo.set_active(0) self.display_combo.set_active(0) self.color_combo.set_active(0) if self.demo != "selfie_nn": self.width_entry.set_value(1920) self.height_entry.set_value(1080) self.width_entry.set_sensitive(False) self.height_entry.set_sensitive(False) else: self.mode_combo.set_active(0) self.device_combo.connect("changed", self.on_source_change) self.launch_button.connect("clicked", self.start) quit_button.connect("clicked", exit) if self.demo == "detect": header.set_title("Detection Demo") elif self.demo == "id": header.set_title("Classification Demo") elif self.demo == "pose": header.set_title("Pose Demo") elif self.demo == "brand": header.set_title("Brand Demo") elif self.demo == "selfie_nn": header.set_title("Selfie Segmenter Demo") else: header.set_title("NNStreamer Demo") header.set_subtitle("NNStreamer Examples") def start(self, button): """Starts the ML Demo with selected settings""" self.update_time = GLib.get_monotonic_time() self.launch_button.set_sensitive(False) if self.color_combo.get_active_text() == "Red": r = 1 g = 0 b = 0 elif self.color_combo.get_active_text() == "Blue": r = 0 g = 0 b = 1 elif self.color_combo.get_active_text() == "Green": r = 0 g = 1 b = 0 elif self.color_combo.get_active_text() == "Black": r = 0 g = 0 b = 0 elif self.color_combo.get_active_text() == "White": r = 1 g = 1 b = 1 else: r = 1 g = 0 b = 0 if self.demo == "detect": if self.platform == "imx93evk": model = utils.download_file("mobilenet_ssd_v2_coco_quant_postprocess_vela.tflite") else: model = utils.download_file("mobilenet_ssd_v2_coco_quant_postprocess.tflite") labels = utils.download_file("coco_labels.txt") if self.device_combo.get_active_text() == "Example Video": device = utils.download_file("detect_example.mov") else: device = self.device_combo.get_active_text() if model == -1 or model == -2 or model == -3: if self.platform == "imx93evk": error = "mobilenet_ssd_v2_coco_quant_postprocess_vela.tflite" else: error = "mobilenet_ssd_v2_coco_quant_postprocess.tflite" elif labels == -1 or labels == -2 or labels == -3: error = "coco_labels.txt" elif device == -1 or device == -2 or device == -3: error = "detect_example.mov" if self.demo == "id": if self.platform == "imx93evk": model = utils.download_file("mobilenet_v1_1.0_224_quant_vela.tflite") else: model = utils.download_file("mobilenet_v1_1.0_224_quant.tflite") labels = utils.download_file("1_1.0_224_labels.txt") if self.device_combo.get_active_text() == "Example Video": device = utils.download_file("id_example.mov") else: device = self.device_combo.get_active_text() if model == -1 or model == -2 or model == -3: if self.platform == "imx93evk": error = "mobilenet_v1_1.0_224_quant_vela.tflite" else: error = "mobilenet_v1_1.0_224_quant.tflite" elif labels == -1 or labels == -2 or labels == -3: error = "1_1.0_224_labels.txt" elif device == -1 or device == -2 or device == -3: error = "id_example.mov" if self.demo == "pose": model = utils.download_file("posenet_resnet50_uint8_float32_quant.tflite") labels = utils.download_file("key_point_labels.txt") if self.device_combo.get_active_text() == "Example Video": device = utils.download_file("pose_example.mov") else: device = self.device_combo.get_active_text() if model == -1 or model == -2 or model == -3: error = "posenet_resnet50_uint8_float32_quant.tflite" elif labels == -1 or labels == -2 or labels == -3: error = "key_point_labels.txt" elif device == -1 or device == -2 or device == -3: error = "pose_example.mov" if self.demo == "brand": model = utils.download_file("brand_model.tflite") labels = utils.download_file("brand_labels.txt") if self.device_combo.get_active_text() == "Example Video": device = utils.download_file("brand_example.mov") else: device = self.device_combo.get_active_text() if model == -1 or model == -2 or model == -3: error = "brand_model.tflite" elif labels == -1 or labels == -2 or labels == -3: error = "brand_labels.txt" elif device == -1 or device == -2 or device == -3: error = "brand_example.mov" if self.demo == "selfie_nn": if self.platform == "imx93evk": model = utils.download_file( "selfie_segmenter_landscape_int8_vela.tflite" ) else: model = utils.download_file("selfie_segmenter_int8.tflite") # Labels refer to background img if self.platform == "imx93evk": labels = utils.download_file("bg_image_landscape.jpg") else: labels = utils.download_file("bg_image.jpg") if self.device_combo.get_active_text() == "Example Video": device = utils.download_file("selfie_example.mov") else: device = self.device_combo.get_active_text() if model == -1 or model == -2 or model == -3: if self.platform == "imx93evk": error = "selfie_segmenter_landscape_int8_vela.tflite" else: error = "selfie_segmenter_int8.tflite" elif labels == -1 or labels == -2 or labels == -3: if self.platform == "imx93evk": error = "bg_image_landscape.jpg" else: error = "bg_image.jpg" elif device == -1 or device == -2 or device == -3: error = "selfie_example.mov" if self.mode_combo.get_active_text() == "Background Substitution": set_mode = 0 else: set_mode = 1 if model == -1 or labels == -1 or device == -1: """ dialog = Gtk.MessageDialog( transient_for=self, flags=0, message_type=Gtk.MessageType.ERROR, buttons=Gtk.ButtonsType.CANCEL, text="Cannot find files! The file that you requested" + " does not have any metadata that is related to it. " + "Please see /home/root/.nxp-demo-experience/downloads.txt" + " to see if the requested file exists! \n \n Cannot find:" + error) dialog.run() dialog.destroy() """ self.status_bar.set_text("Cannot find files!") self.launch_button.set_sensitive(True) return if model == -2 or labels == -2 or device == -2: """ dialog = Gtk.MessageDialog( transient_for=self, flags=0, message_type=Gtk.MessageType.ERROR, buttons=Gtk.ButtonsType.CANCEL, text="Cannot download files! The URL used to download the" + " file cannot be reached. If you are connected to the " + "internet, please check the /home/root/.nxp-demo-experience" + "/downloads.txt for the URL. For some regions, " + "these sites may be blocked. To install these manually," + " please go to the file listed above and provide the " + "path to the file in \"PATH\" \n \n Cannot download " + error) dialog.run() dialog.destroy() """ self.status_bar.set_text("Download failed!") self.launch_button.set_sensitive(True) return if model == -3 or labels == -3 or device == -4: """ dialog = Gtk.MessageDialog( transient_for=self, flags=0, message_type=Gtk.MessageType.ERROR, buttons=Gtk.ButtonsType.CANCEL, text="Invalid files! The files where not what we expected." + "If you are SURE that the files are correct, delete " + "the \"SHA\" value in /home/root/.nxp-demo-experience" + "/downloads.txt to bypass the SHA check. \n \n Bad SHA for " + error) dialog.run() dialog.destroy() """ self.status_bar.set_text("Downloaded bad file!") self.launch_button.set_sensitive(True) return if self.demo == "detect": import nndetection example = nndetection.ObjectDetection( self.platform, device, self.backend_combo.get_active_text(), model, labels, self.display_combo.get_active_text(), self.update_stats, self.width_entry.get_value(), self.height_entry.get_value(), r, g, b, ) example.run() if self.demo == "id": import nnclassification example = nnclassification.NNStreamerExample( self.platform, device, self.backend_combo.get_active_text(), model, labels, self.display_combo.get_active_text(), self.update_stats, self.width_entry.get_value(), self.height_entry.get_value(), r, g, b, ) example.run_example() if self.demo == "pose": import nnpose example = nnpose.NNStreamerExample( self.platform, device, self.backend_combo.get_active_text(), model, labels, self.display_combo.get_active_text(), self.update_stats, self.width_entry.get_value(), self.height_entry.get_value(), r, g, b, ) example.run_example() if self.demo == "brand": import nnbrand example = nnbrand.NNStreamerExample( self.platform, device, self.backend_combo.get_active_text(), model, labels, self.display_combo.get_active_text(), self.update_stats, self.width_entry.get_value(), self.height_entry.get_value(), r, g, b, ) example.run_example() if self.demo == "selfie_nn": import selfie_segmenter example = selfie_segmenter.SelfieSegmenter( self.platform, device, self.backend_combo.get_active_text(), model, labels, self.update_stats, set_mode, r, g, b, ) example.run() self.launch_button.set_sensitive(True) def update_stats(self, time): """Callback used the update stats in GUI""" interval_time = (GLib.get_monotonic_time() - self.update_time) / 1000000 if interval_time > 1: refresh_time = time.interval_time inference_time = time.tensor_filter.get_property("latency") if refresh_time != 0 and inference_time != 0: # Print pipeline information if self.demo == "selfie_nn" or self.demo == "id" or self.demo == "detect": self.time_label.set_text( "{:12.2f} ms".format(1.0 / time.current_framerate * 1000.0) ) self.fps_label.set_text( "{:12.2f} FPS".format(time.current_framerate) ) else: self.time_label.set_text("{:12.2f} ms".format(refresh_time / 1000)) self.fps_label.set_text( "{:12.2f} FPS".format(1 / (refresh_time / 1000000)) ) # Print inference information self.inference_label.set_text( "{:12.2f} ms".format(inference_time / 1000) ) self.ips_label.set_text( "{:12.2f} FPS".format(1 / (inference_time / 1000000)) ) self.update_time = GLib.get_monotonic_time() return True def on_source_change(self, widget): """Callback to lock sliders""" if self.demo != "selfie_nn": if self.device_combo.get_active_text() == "Example Video": self.width_entry.set_value(1920) self.height_entry.set_value(1080) self.width_entry.set_sensitive(False) self.height_entry.set_sensitive(False) else: self.width_entry.set_sensitive(True) self.height_entry.set_sensitive(True) if __name__ == "__main__": if ( len(sys.argv) != 2 and sys.argv[1] != "detect" and sys.argv[1] != "id" and sys.argv[1] != "pose" and sys.argv[1] != "selfie_nn" ): print("Demos available: detect, id, pose, selfie_nn") else: win = MLLaunch(sys.argv[1]) win.connect("destroy", Gtk.main_quit) win.show_all() Gtk.main() |
아래가 실행되는 녀석인데 nndetection을 import 하니까 그걸 따라가서 보는 중.
그나저나 LGPL 이면 그냥 공개해도 되려나?
| root@imx8mpevk:~/.nxp-demo-experience/scripts/machine_learning# find / -name nndetection.py /run/media/root-mmcblk2p2/home/root/.nxp-demo-experience/scripts/machine_learning/nndetection.py /home/root/.nxp-demo-experience/scripts/machine_learning/nndetection.py root@imx8mpevk:~/.nxp-demo-experience/scripts/machine_learning# cat /home/root/.nxp-demo-experience/scripts/machine_learning/nndetection.py #!/usr/bin/env python3 """ Copyright SSAFY Team 1 <jangjongha.sw@gmail.com> Copyright 2021-2023 NXP SPDX-License-Identifier: LGPL-2.1-only Original Source: https://github.com/nnstreamer/nnstreamer-example This demo shows how you can use the NNStreamer to detect objects. From the original source, this was modified to better work with the a GUI and to get better performance on the i.MX 8M Plus and i.MX93. """ import os import sys import gi import re import logging import numpy as np import cairo gi.require_version("Gst", "1.0") gi.require_foreign("cairo") from gi.repository import Gst, GObject, GLib DEBUG = False class ObjectDetection: """The class that manages the demo""" def __init__( self, platform, device, backend, model, labels, display="Weston", callback=None, width=1920, height=1080, r=1, g=0, b=0, ): """Creates an instance of the demo Arguments: device -- What camera or video file to use backend -- Whether to use NPU or CPU model -- the path to the model labels -- the path to the labels display -- Whether to use X11 or Weston callback -- Callback to pass stats to width -- Width of output height -- Height of output r -- Red value for labels g -- Green value for labels b -- Blue value for labels """ self.loop = None self.pipeline = None self.running = False self.video_caps = None self.first_frame = True self.BOX_SIZE = 4 self.LABEL_SIZE = 91 self.DETECTION_MAX = 20 self.MAX_OBJECT_DETECTION = 20 self.Y_SCALE = 10.0 self.X_SCALE = 10.0 self.H_SCALE = 5.0 self.W_SCALE = 5.0 self.VIDEO_WIDTH = width self.VIDEO_HEIGHT = height self.MODEL_WIDTH = 300 self.MODEL_HEIGHT = 300 self.tflite_model = model self.label_path = labels self.device = device self.backend = backend self.display = display self.tflite_labels = [] self.detected_objects = [] self.callback = callback self.r = r self.b = b self.g = g self.platform = platform self.current_framerate = 1000 # Define PXP or GPU2D converter if self.platform == "imx93evk": self.nxp_converter = "imxvideoconvert_pxp " else: self.nxp_converter = "imxvideoconvert_g2d " if not self.tflite_init(): raise Exception Gst.init(None) def run(self): """Starts pipeline and run demo""" if self.backend == "CPU": if self.platform == "imx93evk": backend = "true:cpu custom=NumThreads:2" else: backend = "true:cpu custom=NumThreads:4" elif self.backend == "GPU": os.environ["USE_GPU_INFERENCE"] = "1" backend = ( "true:gpu custom=Delegate:External," "ExtDelegateLib:libvx_delegate.so" ) else: if self.platform == "imx93evk": backend = ( "true:npu custom=Delegate:External," "ExtDelegateLib:libethosu_delegate.so" ) else: os.environ["USE_GPU_INFERENCE"] = "0" backend = ( "true:npu custom=Delegate:External," "ExtDelegateLib:libvx_delegate.so" ) if self.display == "X11": display = "ximagesink name=img_tensor " elif self.display == "None": self.print_time = GLib.get_monotonic_time() display = "fakesink " else: display = "fpsdisplaysink name=img_tensor text-overlay=false video-sink=waylandsink sync=false" # main loop self.loop = GLib.MainLoop() self.old_time = GLib.get_monotonic_time() self.update_time = GLib.get_monotonic_time() self.reload_time = -1 self.interval_time = 999999 # Create decoder for video file if self.platform == "imx8qmmek": decoder = "h264parse ! v4l2h264dec " else: decoder = "vpudec " if "/dev/video" in self.device: gst_launch_cmdline = "v4l2src name=cam_src device=" + self.device gst_launch_cmdline += " ! " + self.nxp_converter + "! video/x-raw,width=" gst_launch_cmdline += str(int(self.VIDEO_WIDTH)) + ",height=" gst_launch_cmdline += str(int(self.VIDEO_HEIGHT)) gst_launch_cmdline += ",framerate=30/1,format=BGRx ! tee name=t" else: gst_launch_cmdline = "filesrc location=" + self.device gst_launch_cmdline += " ! qtdemux ! " + decoder + "! tee name=t" gst_launch_cmdline += " t. ! " + self.nxp_converter + "! video/x-raw," gst_launch_cmdline += "width={:d},".format(self.MODEL_WIDTH) gst_launch_cmdline += "height={:d},".format(self.MODEL_HEIGHT) gst_launch_cmdline += " ! queue max-size-buffers=2 leaky=2 ! " gst_launch_cmdline += "videoconvert ! video/x-raw,format=RGB !" gst_launch_cmdline += " tensor_converter ! tensor_filter" gst_launch_cmdline += " framework=tensorflow-lite model=" gst_launch_cmdline += self.tflite_model + " accelerator=" + backend gst_launch_cmdline += " silent=FALSE name=tensor_filter latency=1 ! " gst_launch_cmdline += "tensor_sink name=tensor_sink t. ! " gst_launch_cmdline += self.nxp_converter + "! " gst_launch_cmdline += "cairooverlay name=tensor_res ! " gst_launch_cmdline += "queue max-size-buffers=2 leaky=2 ! " gst_launch_cmdline += display self.pipeline = Gst.parse_launch(gst_launch_cmdline) # bus and message callback bus = self.pipeline.get_bus() bus.add_signal_watch() bus.connect("message", self.on_bus_message) self.tensor_filter = self.pipeline.get_by_name("tensor_filter") self.wayland_sink = self.pipeline.get_by_name("img_tensor") # tensor sink signal : new data callback tensor_sink = self.pipeline.get_by_name("tensor_sink") tensor_sink.connect("new-data", self.new_data_cb) tensor_res = self.pipeline.get_by_name("tensor_res") tensor_res.connect("draw", self.draw_overlay_cb) tensor_res.connect("caps-changed", self.prepare_overlay_cb) if self.callback is not None: GObject.timeout_add(500, self.callback, self) # start pipeline self.pipeline.set_state(Gst.State.PLAYING) self.running = True self.set_window_title("img_tensor", "NNStreamer Object Detection Example") # run main loop self.loop.run() # quit when received eos or error message self.running = False self.pipeline.set_state(Gst.State.NULL) bus.remove_signal_watch() def tflite_init(self): """ :return: True if successfully initialized """ if not os.path.exists(self.tflite_model): logging.error("cannot find tflite model [%s]", self.tflite_model) return False label_path = self.label_path try: with open(label_path, "r") as label_file: for line in label_file.readlines(): if line[0].isdigit(): while str(len(self.tflite_labels)) not in line: self.tflite_labels.append("Invalid") self.tflite_labels.append(line[line.find(" ") + 1 :]) else: self.tflite_labels.append(line) except FileNotFoundError: logging.error("cannot find tflite label [%s]", label_path) return False logging.info("finished to load labels, total [%d]", len(self.tflite_labels)) return True # @brief Callback for tensor sink signal. def new_data_cb(self, sink, buffer): """Callback for tensor sink signal. :param sink: tensor sink element :param buffer: buffer from element :return: None """ if self.running: new_time = GLib.get_monotonic_time() self.interval_time = new_time - self.old_time self.old_time = new_time if buffer.n_memory() != 4: return False # tensor type is float32. # LOCATIONS_IDX:CLASSES_IDX:SCORES_IDX:NUM_DETECTION_IDX # 4:20:1:1\,20:1:1:1\,20:1:1:1\,1:1:1:1 # [0] detection_boxes (default 4th tensor). BOX_SIZE : # #MaxDetection, ANY-TYPE # [1] detection_classes (default 2nd tensor). # #MaxDetection, ANY-TYPE # [2] detection_scores (default 3rd tensor) # #MaxDetection, ANY-TYPE # [3] num_detection (default 1st tensor). 1, ANY-TYPE # bytestrings that are based on float32 must be # decoded into float list. # boxes mem_boxes = buffer.peek_memory(0) ret, info_boxes = mem_boxes.map(Gst.MapFlags.READ) if ret: assert info_boxes.size == ( self.BOX_SIZE * self.DETECTION_MAX * 4 ), "Invalid info_box size" decoded_boxes = list( np.frombuffer(info_boxes.data, dtype=np.float32) ) # decode bytestrings to float list # detections mem_detections = buffer.peek_memory(1) ret, info_detections = mem_detections.map(Gst.MapFlags.READ) if ret: assert info_detections.size == ( self.DETECTION_MAX * 4 ), "Invalid info_detection size" decoded_detections = list( np.frombuffer(info_detections.data, dtype=np.float32) ) # decode bytestrings to float list # scores mem_scores = buffer.peek_memory(2) ret, info_scores = mem_scores.map(Gst.MapFlags.READ) if ret: assert info_scores.size == ( self.DETECTION_MAX * 4 ), "Invalid info_score size" decoded_scores = list( np.frombuffer(info_scores.data, dtype=np.float32) ) # decode bytestrings to float list # num detection mem_num = buffer.peek_memory(3) ret, info_num = mem_num.map(Gst.MapFlags.READ) if ret: assert info_num.size == 4, "Invalid info_num size" decoded_num = list( np.frombuffer(info_num.data, dtype=np.float32) ) # decode bytestrings to float list self.get_detected_objects( decoded_boxes, decoded_detections, decoded_scores, int(decoded_num[0]) ) mem_boxes.unmap(info_boxes) mem_detections.unmap(info_detections) mem_scores.unmap(info_scores) mem_num.unmap(info_num) if self.display == "None": if (GLib.get_monotonic_time() - self.print_time) > 1000000: inference = self.tensor_filter.get_property("latency") print( "Inference time: " + str(inference / 1000) + " ms (" + "{:5.2f}".format(1 / (inference / 1000000)) + " IPS)" ) self.print_time = GLib.get_monotonic_time() def get_detected_objects(self, boxes, detections, scores, num): """Pairs boxes with dectected objects""" threshold_score = 0.5 detected = list() for i in range(num): score = scores[i] if score < threshold_score: continue c = detections[i] box_offset = self.BOX_SIZE * i ymin = boxes[box_offset + 0] xmin = boxes[box_offset + 1] ymax = boxes[box_offset + 2] xmax = boxes[box_offset + 3] x = xmin * self.MODEL_WIDTH y = ymin * self.MODEL_HEIGHT width = (xmax - xmin) * self.MODEL_WIDTH height = (ymax - ymin) * self.MODEL_HEIGHT obj = { "class_id": int(c), "x": x, "y": y, "width": width, "height": height, "prob": score, } detected.append(obj) # update result self.detected_objects.clear() for d in detected: self.detected_objects.append(d) if DEBUG: print("==============================") print("LABEL : {}".format(self.tflite_labels[d["class_id"]])) print("x : {}".format(d["x"])) print("y : {}".format(d["y"])) print("width : {}".format(d["width"])) print("height : {}".format(d["height"])) print("Confidence Score: {}".format(d["prob"])) def prepare_overlay_cb(self, overlay, caps): """Store the information from the caps that we are interested in.""" self.video_caps = caps def draw_overlay_cb(self, overlay, context, timestamp, duration): """Callback to draw the overlay.""" if self.video_caps is None or not self.running: return scale_height = self.VIDEO_HEIGHT / 1080 scale_width = self.VIDEO_WIDTH / 1920 scale_text = max(scale_height, scale_width) # mutex_lock alternative required detected = self.detected_objects # mutex_unlock alternative needed drawed = 0 context.select_font_face( "Sans", cairo.FONT_SLANT_NORMAL, cairo.FONT_WEIGHT_BOLD ) context.set_font_size(int(50.0 * scale_text)) context.set_source_rgb(self.r, self.g, self.b) for obj in detected: label = self.tflite_labels[obj["class_id"]][:-1] x = obj["x"] * self.VIDEO_WIDTH // self.MODEL_WIDTH y = obj["y"] * self.VIDEO_HEIGHT // self.MODEL_HEIGHT width = obj["width"] * self.VIDEO_WIDTH // self.MODEL_WIDTH height = obj["height"] * self.VIDEO_HEIGHT // self.MODEL_HEIGHT # draw rectangle context.rectangle(x, y, width, height) context.set_line_width(3) context.stroke() # draw title context.move_to(x + 5, y + int(50.0 * scale_text)) context.show_text(label) drawed += 1 if drawed >= self.MAX_OBJECT_DETECTION: break inference = self.tensor_filter.get_property("latency") # Get current framerate and avg. framerate output_wayland = self.wayland_sink.get_property("last-message") if output_wayland: current_text = re.findall(r"current:\s[\d]+[.\d]*", output_wayland)[0] self.current_framerate = float(re.findall(r"[\d]+[.\d]*", current_text)[0]) context.set_font_size(int(25.0 * scale_text)) context.move_to( int(50 * scale_width), int(self.VIDEO_HEIGHT - (100 * scale_height)) ) context.show_text("i.MX NNStreamer Detection Demo") if inference == 0: context.move_to( int(50 * scale_width), int(self.VIDEO_HEIGHT - (75 * scale_height)) ) context.show_text("FPS: ") context.move_to( int(50 * scale_width), int(self.VIDEO_HEIGHT - (50 * scale_height)) ) context.show_text("IPS: ") elif ( GLib.get_monotonic_time() - self.reload_time ) < 100000 and self.refresh_time != -1: context.move_to( int(50 * scale_width), int(self.VIDEO_HEIGHT - (75 * scale_height)) ) context.show_text( "FPS: {:6.2f} ({:6.2f} ms)".format( self.current_framerate, 1.0 / self.current_framerate * 1000 ) ) context.move_to( int(50 * scale_width), int(self.VIDEO_HEIGHT - (50 * scale_height)) ) context.show_text( "IPS: {:6.2f} ({:6.2f} ms)".format( 1 / (inference / 1000000), inference / 1000 ) ) else: self.reload_time = GLib.get_monotonic_time() self.refresh_time = self.interval_time self.inference = self.tensor_filter.get_property("latency") context.move_to( int(50 * scale_width), int(self.VIDEO_HEIGHT - (75 * scale_height)) ) context.show_text( "FPS: {:6.2f} ({:6.2f} ms)".format( self.current_framerate, 1.0 / self.current_framerate * 1000 ) ) context.move_to( int(50 * scale_width), int(self.VIDEO_HEIGHT - (50 * scale_height)) ) context.show_text( "IPS: {:6.2f} ({:6.2f} ms)".format( 1 / (inference / 1000000), inference / 1000 ) ) if self.first_frame: context.move_to(int(400 * scale_width), int(600 * scale_height)) context.set_font_size(int(200.0 * min(scale_width, scale_height))) context.show_text("Loading...") self.first_frame = False context.fill() def on_bus_message(self, bus, message): """Callback for message. :param bus: pipeline bus :param message: message from pipeline :return: None """ if message.type == Gst.MessageType.EOS: logging.info("received eos message") self.loop.quit() elif message.type == Gst.MessageType.ERROR: error, debug = message.parse_error() logging.warning("[error] %s : %s", error.message, debug) self.loop.quit() elif message.type == Gst.MessageType.WARNING: error, debug = message.parse_warning() logging.warning("[warning] %s : %s", error.message, debug) elif message.type == Gst.MessageType.STREAM_START: logging.info("received start message") elif message.type == Gst.MessageType.QOS: data_format, processed, dropped = message.parse_qos_stats() format_str = Gst.Format.get_name(data_format) logging.debug( "[qos] format[%s] processed[%d] dropped[%d]", format_str, processed, dropped, ) def set_window_title(self, name, title): """Set window title for X11. :param name: GstXImageasink element name :param title: window title :return: None """ element = self.pipeline.get_by_name(name) if element is not None: pad = element.get_static_pad("sink") if pad is not None: tags = Gst.TagList.new_empty() tags.add_value(Gst.TagMergeMode.APPEND, "title", title) pad.send_event(Gst.Event.new_tag(tags)) if __name__ == "__main__": if ( len(sys.argv) != 7 and len(sys.argv) != 5 and len(sys.argv) != 9 and len(sys.argv) != 12 and len(sys.argv) != 6 ): print( "Usage: python3 nndetection.py <dev/video*/video file>" + " <NPU/CPU> <model file> <label file>" ) exit() # Get platform platform = os.uname().nodename if len(sys.argv) == 7: example = ObjectDetection( platform, sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4], sys.argv[5], sys.argv[6], ) if len(sys.argv) == 5: example = ObjectDetection( platform, sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4] ) if len(sys.argv) == 6: example = ObjectDetection( platform, sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4], sys.argv[5] ) if len(sys.argv) == 9: example = ObjectDetection( platform, sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4], sys.argv[5], sys.argv[6], int(sys.argv[7]), int(sys.argv[8]), ) if len(sys.argv) == 12: example = ObjectDetection( platform, sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4], sys.argv[5], sys.argv[6], int(sys.argv[7]), int(sys.argv[8]), int(sys.argv[9]), int(sys.argv[10]), int(sys.argv[11]), ) example.run() |
| self.pipeline = Gst.parse_launch( 'v4l2src name=cam_src ! videoconvert ! videoscale ! ' 'video/x-raw,width=640,height=480,format=RGB ! tee name=t_raw ' 't_raw. ! queue leaky=2 max-size-buffers=2 ! videoscale ! video/x-raw,width=300,height=300 ! tensor_converter ! ' 'tensor_transform mode=arithmetic option=typecast:float32,add:-127.5,div:127.5 ! ' 'tensor_filter framework=tensorflow-lite model=' + self.tflite_model + ' ! ' 'tensor_decoder mode=bounding_boxes option1=mobilenet-ssd option2=' + self.tflite_label + ' option3=' + self.tflite_box_prior + ' option4=640:480 option5=300:300 !' 'compositor name=mix sink_0::zorder=2 sink_1::zorder=1 ! videoconvert ! ximagesink ' 't_raw. ! queue leaky=2 max-size-buffers=10 ! mix. ' ) |
gst_launch_cmdline 를 출력해보니 아래와 같이 gstreamer 파이프라인이 나온다.
| v4l2src name=cam_src device=/dev/video3 ! imxvideoconvert_g2d ! video/x-raw,width=1920,height=1080,framerate=30/1,format=BGRx ! tee name=t t. ! imxvideoconvert_g2d ! video/x-raw,width=300,height=300, ! queue max-size-buffers=2 leaky=2 ! videoconvert ! video/x-raw,format=RGB ! tensor_converter ! tensor_filter framework=tensorflow-lite model=/home/root/.cache/gopoint/mobilenet_ssd_v2_coco_quant_postprocess.tflite accelerator=true:npu custom=Delegate:External,ExtDelegateLib:libvx_delegate.so silent=FALSE name=tensor_filter latency=1 ! tensor_sink name=tensor_sink t. ! imxvideoconvert_g2d ! cairooverlay name=tensor_res ! queue max-size-buffers=2 leaky=2 ! fpsdisplaysink name=img_tensor text-overlay=false video-sink=waylandsink sync=false |
보기어려우니 엔터로 구분
| v4l2src name=cam_src device=/dev/video3 ! imxvideoconvert_g2d ! video/x-raw,width=1920,height=1080,framerate=30/1,format=BGRx ! tee name=t t. ! imxvideoconvert_g2d ! video/x-raw,width=300,height=300, ! queue max-size-buffers=2 leaky=2 ! videoconvert ! video/x-raw,format=RGB ! tensor_converter ! tensor_filter framework=tensorflow-lite model=/home/root/.cache/gopoint/mobilenet_ssd_v2_coco_quant_postprocess.tflite accelerator=true:npu custom=Delegate:External,ExtDelegateLib:libvx_delegate.so silent=FALSE name=tensor_filter latency=1 ! tensor_sink name=tensor_sink t. ! imxvideoconvert_g2d ! cairooverlay name=tensor_res ! queue max-size-buffers=2 leaky=2 ! fpsdisplaysink name=img_tensor text-overlay=false video-sink=waylandsink sync=false |
+
2024.01.03
| # cd /home/root/.nxp-demo-experience/scripts/machine_learning # python3 nndetection.py /dev/video3 NPU /home/root/.cache/gopoint/mobilenet_ssd_v2_coco_quant_postprocess.tflite /home/root/.cache/gopoint/coco_labels.txt |
gst-launch 로도 실행은 되는데 callback 처리가 안되서 overlay가 출력이 안되어 동일한 화면을 보여주진 않는다.
| gst-launch-1.0 v4l2src name=cam_src device=/dev/video3 ! imxvideoconvert_g2d ! video/x-raw,width=1920,height=1080,framerate=30/1,format=BGRx ! tee name=t t. ! imxvideoconvert_g2d ! video/x-raw,width=300,height=300, ! queue max-size-buffers=2 leaky=2 ! videoconvert ! video/x-raw,format=RGB ! tensor_converter ! tensor_filter framework=tensorflow-lite model=/home/root/.cache/gopoint/mobilenet_ssd_v2_coco_quant_postprocess.tflite accelerator=true:npu custom=Delegate:External,ExtDelegateLib:libvx_delegate.so silent=FALSE name=tensor_filter latency=1 ! tensor_sink name=tensor_sink t. ! imxvideoconvert_g2d ! cairooverlay name=tensor_res ! queue max-size-buffers=2 leaky=2 ! fpsdisplaysink name=img_tensor text-overlay=false video-sink=waylandsink sync=false |
| 우분투에 jupyter notebook 설치 및 실행하기 (0) | 2024.01.02 |
|---|---|
| 주피터 노트북 프로젝트(?) 실행하기 (0) | 2024.01.02 |
| tensorflow keras dataset (0) | 2024.01.02 |
| tensorflow lite / mnist 학습 (0) | 2024.01.02 |
| yolo-label (0) | 2022.03.22 |
자동완성으로 해보니 몇가지 나오는데 mnist 말고는 몰라서 찾아보는 중
| >>> tf.keras.datasets. tf.keras.datasets.boston_housing tf.keras.datasets.cifar100 tf.keras.datasets.imdb tf.keras.datasets.reuters tf.keras.datasets.cifar10 tf.keras.datasets.fashion_mnist tf.keras.datasets.mnist |
imdb는 영화 db
boston_housing은 statlib 사이트에서 정의된 보스톤 주택가격
reuter는 46 주제에 따른 11228 뉴스(로이터 뉴스) 인 듯.
| This is a dataset of 11,228 newswires from Reuters, labeled over 46 topics. |
[링크 : https://www.tensorflow.org/api_docs/python/tf/keras/datasets/boston_housing/load_data]
[링크 : https://www.tensorflow.org/api_docs/python/tf/keras/datasets]
cifar10은 10개 클래스니까.. 결과도 MNIST 처럼 10개로 나올 것 같고..
| The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images. |
[링크 : https://www.tensorflow.org/datasets/catalog/cifar10?hl=en]
| This dataset is just like the CIFAR-10, except it has 100 classes containing 600 images each. There are 500 training images and 100 testing images per class. The 100 classes in the CIFAR-100 are grouped into 20 superclasses. Each image comes with a "fine" label (the class to which it belongs) and a "coarse" label (the superclass to which it belongs). |
[링크 : https://www.tensorflow.org/datasets/catalog/cifar100?hl=en]
[링크 : https://www.tensorflow.org/datasets/catalog/fashion_mnist?hl=en]
[링크 : https://www.tensorflow.org/datasets/catalog/emnist?hl=en]
| 주피터 노트북 프로젝트(?) 실행하기 (0) | 2024.01.02 |
|---|---|
| i.mx8mp gopoint 실행 경로 (0) | 2024.01.02 |
| tensorflow lite / mnist 학습 (0) | 2024.01.02 |
| yolo-label (0) | 2022.03.22 |
| tflite bazel rpi3b+ (0) | 2022.01.27 |
신기하네.. 그냥 알아서 받네?
| >>> mnist = tf.keras.datasets.mnist >>> (train_images, train_labels), (test_images, test_labels) = mnist.load_data() Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz 11490434/11490434 [==============================] - 1s 0us/step |
학습하고 tflite 파일로 저장하기
| import tensorflow as tf import numpy as np mnist = tf.keras.datasets.mnist (train_images, train_labels), (test_images, test_labels) = mnist.load_data() train_images = train_images.astype(np.float32) / 255.0 test_images = test_images.astype(np.float32) / 255.0 model = tf.keras.Sequential([ tf.keras.layers.InputLayer(input_shape=(28, 28)), tf.keras.layers.Reshape(target_shape=(28, 28, 1)), tf.keras.layers.Conv2D(filters=12, kernel_size=(3, 3), activation='relu'), tf.keras.layers.MaxPooling2D(pool_size=(2, 2)), tf.keras.layers.Flatten(), tf.keras.layers.Dense(10) ]) model.compile(optimizer='adam', loss=tf.keras.losses.SparseCategoricalCrossentropy( from_logits=True), metrics=['accuracy']) model.fit( train_images, train_labels, epochs=5, validation_data=(test_images, test_labels) ) # 일반 모델로 변환 converter = tf.lite.TFLiteConverter.from_keras_model(model) tflite_model = converter.convert() # 요건 차이 없음 # converter = tf.lite.TFLiteConverter.from_keras_model(model) # converter.optimizations = [tf.lite.Optimize.DEFAULT] # tflite_model_quant = converter.convert() # quant 를 하려면 아래 코드 실행해야 함 def representative_data_gen(): for input_value in tf.data.Dataset.from_tensor_slices(train_images).batch(1).take(100): yield [input_value] converter = tf.lite.TFLiteConverter.from_keras_model(model) converter.optimizations = [tf.lite.Optimize.DEFAULT] converter.representative_dataset = representative_data_gen converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8] converter.inference_input_type = tf.uint8 converter.inference_output_type = tf.uint8 tflite_model_quant = converter.convert() # 파일로 저장하기 import pathlib tflite_models_dir = pathlib.Path("/tmp/mnist_tflite_models/") tflite_models_dir.mkdir(exist_ok=True, parents=True) # Save the unquantized/float model: tflite_model_file = tflite_models_dir/"mnist_model.tflite" tflite_model_file.write_bytes(tflite_model) # Save the quantized model: tflite_model_quant_file = tflite_models_dir/"mnist_model_quant.tflite" tflite_model_quant_file.write_bytes(tflite_model_quant) |
[링크 : https://www.tensorflow.org/lite/performance/post_training_integer_quant?hl=ko]
netron을 통해 생성한걸 보는데 quant나 그냥이나 어째 차이가 없냐?
+
quantization 하면 uint8로 변경된다.
그나저나, MNIST에 대해서 오해가 있었다.
출력이 [1,10] 인데 0~9 까지의 숫자에 대한 필기 데이터베이스지 알파벳이 아니란 것 -_-!
그래서 출력이 딱 10개인 건 당연하다는 것..
[링크 : https://en.wikipedia.org/wiki/MNIST_database]
+
EMNIST 라고 알파벳 손글씨가 따로 있다.
[링크 : https://www.nist.gov/itl/products-and-services/emnist-dataset]
[링크 : https://www.tensorflow.org/datasets/catalog/emnist?hl=ko]
| i.mx8mp gopoint 실행 경로 (0) | 2024.01.02 |
|---|---|
| tensorflow keras dataset (0) | 2024.01.02 |
| yolo-label (0) | 2022.03.22 |
| tflite bazel rpi3b+ (0) | 2022.01.27 |
| bazel cross compile (0) | 2022.01.27 |
윈도우용 라벨링 프로그램 찾다가 발견
| tensorflow keras dataset (0) | 2024.01.02 |
|---|---|
| tensorflow lite / mnist 학습 (0) | 2024.01.02 |
| tflite bazel rpi3b+ (0) | 2022.01.27 |
| bazel cross compile (0) | 2022.01.27 |
| bazel clean (0) | 2021.10.19 |
요약하면.. 메모리 부족하니 빌드 속도를 포기하고 swap을 이용해서 커버치기!
[링크 : https://gist.github.com/EKami/9869ae6347f68c592c5b5cd181a3b205]
| tensorflow lite / mnist 학습 (0) | 2024.01.02 |
|---|---|
| yolo-label (0) | 2022.03.22 |
| bazel cross compile (0) | 2022.01.27 |
| bazel clean (0) | 2021.10.19 |
| 2.7.0-rc with opencl (0) | 2021.10.13 |
[링크 : https://docs.bazel.build/versions/main/platforms-intro.html]
[링크 : https://docs.bazel.build/versions/main/tutorial/cc-toolchain-config.html]
| yolo-label (0) | 2022.03.22 |
|---|---|
| tflite bazel rpi3b+ (0) | 2022.01.27 |
| bazel clean (0) | 2021.10.19 |
| 2.7.0-rc with opencl (0) | 2021.10.13 |
| tf release 2.7.0-rc (0) | 2021.10.12 |
리눅스 홈 디렉토리 용량 검사하다 보니 이상하게 많이 먹어 추적해보니
~/.cache/bazel 이 7기가 정도?
bazel clean 을 통해서 용량을 감소시킬수 있다고 하는데 문제는 워크스페이스 날렸으면 무리
그냥 쿨(?) 하게 ~/.cache/bazel 을 날리니 용량이 훅 줄어든다.
| $ bazel clean Extracting Bazel installation... ERROR: The 'clean' command is only supported from within a workspace (below a directory having a WORKSPACE file). See documentation at https://docs.bazel.build/versions/master/build-ref.html#workspace |
| tflite bazel rpi3b+ (0) | 2022.01.27 |
|---|---|
| bazel cross compile (0) | 2022.01.27 |
| 2.7.0-rc with opencl (0) | 2021.10.13 |
| tf release 2.7.0-rc (0) | 2021.10.12 |
| tflite delegate (0) | 2021.10.11 |
|
