There is no difference in execution speed between CPU and GPU.

[Nibosi0501]

Have I written custom code (as opposed to using a stock example script provided in MediaPipe)

Yes

OS Platform and Distribution

macOS Venture 13.7.3
Apple M1 16GB

MediaPipe Tasks SDK version

mediapipe==0.10.31

Task name (e.g. Image classification, Gesture recognition etc.)

Pose landmark detection

Programming Language and version (e.g. C++, Python, Java)

python 3.12.7

Describe the actual behavior

Memory leaks during GPU execution no longer occur with mediapipe==0.10.31, but the processing speed difference between CPU and GPU processing is no longer noticeable, as it was with mediapipe==0.10.20.

Describe the expected behaviour

I want to improve processing speed using a GPU.

Standalone code/steps you may have used to try to get what you need

import mediapipe as mp
from mediapipe.tasks import python
from mediapipe.tasks.python import vision

import cv2 as cv
import numpy as np

import time

BaseOptions = mp.tasks.BaseOptions
PoseLandmarker = mp.tasks.vision.PoseLandmarker
PoseLandmarkerOptions = mp.tasks.vision.PoseLandmarkerOptions
PoseLandmarkerResult = mp.tasks.vision.PoseLandmarkerResult
VisionRunningMode = mp.tasks.vision.RunningMode

base_options = python.BaseOptions(
    model_asset_path='./models/mp/pose_landmarker_heavy.task',
    # delegate=python.BaseOptions.Delegate.CPU,  # CPU
    delegate=python.BaseOptions.Delegate.GPU    # GPU
)

options = PoseLandmarkerOptions(
    #base_options=BaseOptions(model_asset_path='./models/mp/pose_landmarker_heavy.task'),   # CPU
    base_options=base_options,                             # GPU
    running_mode=VisionRunningMode.VIDEO,
    num_poses = 4,
    min_pose_detection_confidence=0.7
    )

def main():
    
    cap = cv.VideoCapture(0)

    if not cap.isOpened():
        print("Cannot open camera")
        return
    
    with PoseLandmarker.create_from_options(options) as landmarker:
        while True:
            ret, frame = cap.read()

            if not ret:
                print("Cannot receive frame")
                break

            mediapipe_start = time.time()

            # Convert BGR image to RGB
            # GPU
            rgba_frame = cv.cvtColor(frame, cv.COLOR_BGR2RGBA)
            mp_image = mp.Image(image_format=mp.ImageFormat.SRGBA, data=rgba_frame)

            # CPU
            #rgb_image = cv.cvtColor(frame, cv.COLOR_BGR2RGB)
            #mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=rgb_image)

            # landmark detection
            pose_landmarker_result = landmarker.detect_for_video(mp_image, int(time.time() * 1000))

            mediapipe_end = time.time()
            elapsed_time = mediapipe_end - mediapipe_start
            print(f"Mediapipe Pose Landmarker: {elapsed_time*1000:.2f} ms")

            cv.imshow('Pose Landmarks', frame)

            if cv.waitKey(1) & 0xFF == 27:
                break

if __name__ == "__main__":
    main()

Other info / Complete Logs

Additionally, the following warning appears during GPU execution. Is this related?

W0000 00:00:1768377893.862985  970510 landmark_projection_calculator.cc:81] Using NORM_RECT without IMAGE_DIMENSIONS is only supported for the square ROI. Provide IMAGE_DIMENSIONS or use PROJECTION_MATRIX.

[akashvverma1995]

Hi @Nibosi0501

Thank you for reporting the issue. We are currently discussing this with our internal team, and we will get back to you as soon as we have an update

[akashvverma1995]

Hi @Nibosi0501

Could you please share a screenshot of the logs?

[Nibosi0501]

We changed the format to load video files instead of using webcam footage.
We also added a feature to draw detected landmarks.
The program was set to loop 100 times.

URL of the video used (640 x 360px)
https://video-ac.com/video/6882#goog_rewarded
If you have any specific requests for the test video, please let me know so I can make the necessary changes.

In this version, enabling the GPU caused a memory leak bug.
However, processing is faster with the GPU enabled compared to the CPU.

mediapipe==0.10.20

import mediapipe as mp
from mediapipe.tasks import python
from mediapipe.tasks.python import vision

from mediapipe import solutions
from mediapipe.framework.formats import landmark_pb2

import cv2 as cv
import numpy as np

import time

BaseOptions = mp.tasks.BaseOptions
PoseLandmarker = mp.tasks.vision.PoseLandmarker
PoseLandmarkerOptions = mp.tasks.vision.PoseLandmarkerOptions
PoseLandmarkerResult = mp.tasks.vision.PoseLandmarkerResult
VisionRunningMode = mp.tasks.vision.RunningMode

base_options = python.BaseOptions(
    model_asset_path='./pose_landmarker_heavy.task',
    # delegate=python.BaseOptions.Delegate.CPU,  # CPU
    delegate=python.BaseOptions.Delegate.GPU    # GPU
)

options = PoseLandmarkerOptions(
    #base_options=BaseOptions(model_asset_path='./pose_landmarker_heavy.task'),   # CPU
    base_options=base_options,                             # GPU
    running_mode=VisionRunningMode.VIDEO,
    num_poses = 4,
    min_pose_detection_confidence=0.7
    )

def main():

    frame_cnt = 0
    
    cap = cv.VideoCapture('./dance.mov')

    if not cap.isOpened():
        print("Cannot open camera")
        return
    
    with PoseLandmarker.create_from_options(options) as landmarker:
        while True:
            ret, frame = cap.read()

            if not ret:
                print("Cannot receive frame")
                break

            frame_cnt += 1
            if frame_cnt == 100:
                break

            mediapipe_start = time.time()

            # Convert BGR image to RGB
            # GPU
            rgba_frame = cv.cvtColor(frame, cv.COLOR_BGR2RGBA)
            mp_image = mp.Image(image_format=mp.ImageFormat.SRGBA, data=rgba_frame)

            # CPU
            #rgb_image = cv.cvtColor(frame, cv.COLOR_BGR2RGB)
            #mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=rgb_image)

            # landmark detection
            pose_landmarker_result = landmarker.detect_for_video(mp_image, int(time.time() * 1000))

            mediapipe_end = time.time()
            elapsed_time = mediapipe_end - mediapipe_start
            print(f"Mediapipe Pose Landmarker: {elapsed_time*1000:.2f} ms")

            annotated_image = frame.copy()
            if pose_landmarker_result.pose_landmarks is not None:
                for pose_landmarks in pose_landmarker_result.pose_landmarks:

                    pose_landmarks_proto = landmark_pb2.NormalizedLandmarkList()
                    pose_landmarks_proto.landmark.extend([
                        landmark_pb2.NormalizedLandmark(x=landmark.x, y=landmark.y, z=landmark.z, visibility=landmark.visibility) for landmark in pose_landmarks
                    ])

                    solutions.drawing_utils.draw_landmarks(
                        annotated_image,
                        pose_landmarks_proto,
                        solutions.pose_connections.POSE_CONNECTIONS,
                        solutions.drawing_styles.get_default_pose_landmarks_style()
                    )

            cv.imshow('Pose Landmarks', annotated_image)

            if cv.waitKey(1) & 0xFF == 27:
                break

if __name__ == "__main__":
    main()

GPU output

WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1769011147.776678 2963632 gl_context.cc:369] GL version: 2.1 (2.1 Metal - 83.1), renderer: Apple M1
INFO: Created TensorFlow Lite delegate for Metal.
W0000 00:00:1769011148.087704 2963698 landmark_projection_calculator.cc:186] Using NORM_RECT without IMAGE_DIMENSIONS is only supported for the square ROI. Provide IMAGE_DIMENSIONS or use PROJECTION_MATRIX.
Mediapipe Pose Landmarker: 34.66 ms
Mediapipe Pose Landmarker: 19.01 ms
Mediapipe Pose Landmarker: 19.88 ms
Mediapipe Pose Landmarker: 19.33 ms
Mediapipe Pose Landmarker: 19.98 ms
Mediapipe Pose Landmarker: 19.85 ms
Mediapipe Pose Landmarker: 14.11 ms
Mediapipe Pose Landmarker: 19.44 ms
Mediapipe Pose Landmarker: 21.18 ms
Mediapipe Pose Landmarker: 17.77 ms
Mediapipe Pose Landmarker: 20.20 ms
Mediapipe Pose Landmarker: 18.37 ms
Mediapipe Pose Landmarker: 21.73 ms
Mediapipe Pose Landmarker: 16.74 ms
Mediapipe Pose Landmarker: 16.42 ms
Mediapipe Pose Landmarker: 20.39 ms
Mediapipe Pose Landmarker: 21.27 ms
Mediapipe Pose Landmarker: 19.91 ms
Mediapipe Pose Landmarker: 14.93 ms
Mediapipe Pose Landmarker: 13.29 ms
Mediapipe Pose Landmarker: 15.47 ms
Mediapipe Pose Landmarker: 15.80 ms
Mediapipe Pose Landmarker: 16.28 ms
Mediapipe Pose Landmarker: 19.21 ms
Mediapipe Pose Landmarker: 19.14 ms
Mediapipe Pose Landmarker: 23.62 ms
Mediapipe Pose Landmarker: 22.38 ms
Mediapipe Pose Landmarker: 22.10 ms
Mediapipe Pose Landmarker: 17.62 ms
Mediapipe Pose Landmarker: 22.36 ms
Mediapipe Pose Landmarker: 17.36 ms
Mediapipe Pose Landmarker: 20.75 ms
Mediapipe Pose Landmarker: 20.07 ms
Mediapipe Pose Landmarker: 21.22 ms
Mediapipe Pose Landmarker: 20.50 ms
Mediapipe Pose Landmarker: 17.99 ms
Mediapipe Pose Landmarker: 19.96 ms
Mediapipe Pose Landmarker: 23.66 ms
Mediapipe Pose Landmarker: 22.37 ms
Mediapipe Pose Landmarker: 20.58 ms
Mediapipe Pose Landmarker: 19.73 ms
Mediapipe Pose Landmarker: 24.30 ms
Mediapipe Pose Landmarker: 20.49 ms
Mediapipe Pose Landmarker: 19.26 ms
Mediapipe Pose Landmarker: 21.79 ms
Mediapipe Pose Landmarker: 22.73 ms
Mediapipe Pose Landmarker: 21.44 ms
Mediapipe Pose Landmarker: 20.23 ms
Mediapipe Pose Landmarker: 22.27 ms
Mediapipe Pose Landmarker: 25.18 ms
Mediapipe Pose Landmarker: 19.22 ms
Mediapipe Pose Landmarker: 23.15 ms
Mediapipe Pose Landmarker: 20.89 ms
Mediapipe Pose Landmarker: 17.79 ms
Mediapipe Pose Landmarker: 21.54 ms
Mediapipe Pose Landmarker: 21.07 ms
Mediapipe Pose Landmarker: 24.47 ms
Mediapipe Pose Landmarker: 16.95 ms
Mediapipe Pose Landmarker: 13.85 ms
Mediapipe Pose Landmarker: 13.99 ms
Mediapipe Pose Landmarker: 15.90 ms
Mediapipe Pose Landmarker: 17.01 ms
Mediapipe Pose Landmarker: 17.64 ms
Mediapipe Pose Landmarker: 16.29 ms
Mediapipe Pose Landmarker: 14.82 ms
Mediapipe Pose Landmarker: 14.69 ms
Mediapipe Pose Landmarker: 17.23 ms
Mediapipe Pose Landmarker: 21.65 ms
Mediapipe Pose Landmarker: 22.46 ms
Mediapipe Pose Landmarker: 20.97 ms
Mediapipe Pose Landmarker: 17.01 ms
Mediapipe Pose Landmarker: 16.47 ms
Mediapipe Pose Landmarker: 18.41 ms
Mediapipe Pose Landmarker: 23.32 ms
Mediapipe Pose Landmarker: 21.80 ms
Mediapipe Pose Landmarker: 18.29 ms
Mediapipe Pose Landmarker: 19.36 ms
Mediapipe Pose Landmarker: 15.29 ms
Mediapipe Pose Landmarker: 17.53 ms
Mediapipe Pose Landmarker: 19.66 ms
Mediapipe Pose Landmarker: 23.82 ms
Mediapipe Pose Landmarker: 24.40 ms
Mediapipe Pose Landmarker: 16.57 ms
Mediapipe Pose Landmarker: 14.39 ms
Mediapipe Pose Landmarker: 13.65 ms
Mediapipe Pose Landmarker: 16.46 ms
Mediapipe Pose Landmarker: 17.11 ms
Mediapipe Pose Landmarker: 22.09 ms
Mediapipe Pose Landmarker: 19.93 ms
Mediapipe Pose Landmarker: 18.80 ms
Mediapipe Pose Landmarker: 24.98 ms
Mediapipe Pose Landmarker: 23.45 ms
Mediapipe Pose Landmarker: 16.63 ms
Mediapipe Pose Landmarker: 13.99 ms
Mediapipe Pose Landmarker: 15.11 ms
Mediapipe Pose Landmarker: 14.56 ms
Mediapipe Pose Landmarker: 16.21 ms
Mediapipe Pose Landmarker: 22.49 ms
Mediapipe Pose Landmarker: 21.78 ms

CPU output

WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1769011318.866541 2965176 gl_context.cc:369] GL version: 2.1 (2.1 Metal - 83.1), renderer: Apple M1
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
W0000 00:00:1769011318.926003 2965225 inference_feedback_manager.cc:114] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1769011318.960662 2965227 inference_feedback_manager.cc:114] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1769011319.043692 2965229 landmark_projection_calculator.cc:186] Using NORM_RECT without IMAGE_DIMENSIONS is only supported for the square ROI. Provide IMAGE_DIMENSIONS or use PROJECTION_MATRIX.
Mediapipe Pose Landmarker: 80.31 ms
Mediapipe Pose Landmarker: 77.68 ms
Mediapipe Pose Landmarker: 77.65 ms
Mediapipe Pose Landmarker: 77.81 ms
Mediapipe Pose Landmarker: 77.23 ms
Mediapipe Pose Landmarker: 77.55 ms
Mediapipe Pose Landmarker: 77.25 ms
Mediapipe Pose Landmarker: 78.19 ms
Mediapipe Pose Landmarker: 76.77 ms
Mediapipe Pose Landmarker: 77.19 ms
Mediapipe Pose Landmarker: 77.21 ms
Mediapipe Pose Landmarker: 77.13 ms
Mediapipe Pose Landmarker: 77.07 ms
Mediapipe Pose Landmarker: 77.50 ms
Mediapipe Pose Landmarker: 77.17 ms
Mediapipe Pose Landmarker: 77.18 ms
Mediapipe Pose Landmarker: 77.02 ms
Mediapipe Pose Landmarker: 77.34 ms
Mediapipe Pose Landmarker: 77.22 ms
Mediapipe Pose Landmarker: 77.06 ms
Mediapipe Pose Landmarker: 77.17 ms
Mediapipe Pose Landmarker: 77.34 ms
Mediapipe Pose Landmarker: 76.83 ms
Mediapipe Pose Landmarker: 77.12 ms
Mediapipe Pose Landmarker: 77.20 ms
Mediapipe Pose Landmarker: 77.02 ms
Mediapipe Pose Landmarker: 77.22 ms
Mediapipe Pose Landmarker: 77.55 ms
Mediapipe Pose Landmarker: 77.01 ms
Mediapipe Pose Landmarker: 77.50 ms
Mediapipe Pose Landmarker: 77.23 ms
Mediapipe Pose Landmarker: 77.25 ms
Mediapipe Pose Landmarker: 76.85 ms
Mediapipe Pose Landmarker: 77.39 ms
Mediapipe Pose Landmarker: 77.31 ms
Mediapipe Pose Landmarker: 77.09 ms
Mediapipe Pose Landmarker: 77.08 ms
Mediapipe Pose Landmarker: 77.06 ms
Mediapipe Pose Landmarker: 77.22 ms
Mediapipe Pose Landmarker: 76.97 ms
Mediapipe Pose Landmarker: 77.62 ms
Mediapipe Pose Landmarker: 77.40 ms
Mediapipe Pose Landmarker: 76.90 ms
Mediapipe Pose Landmarker: 77.11 ms
Mediapipe Pose Landmarker: 77.16 ms
Mediapipe Pose Landmarker: 77.08 ms
Mediapipe Pose Landmarker: 77.21 ms
Mediapipe Pose Landmarker: 77.18 ms
Mediapipe Pose Landmarker: 77.68 ms
Mediapipe Pose Landmarker: 77.42 ms
Mediapipe Pose Landmarker: 77.27 ms
Mediapipe Pose Landmarker: 77.03 ms
Mediapipe Pose Landmarker: 77.08 ms
Mediapipe Pose Landmarker: 77.40 ms
Mediapipe Pose Landmarker: 77.19 ms
Mediapipe Pose Landmarker: 77.03 ms
Mediapipe Pose Landmarker: 77.40 ms
Mediapipe Pose Landmarker: 77.01 ms
Mediapipe Pose Landmarker: 77.35 ms
Mediapipe Pose Landmarker: 77.09 ms
Mediapipe Pose Landmarker: 77.47 ms
Mediapipe Pose Landmarker: 77.53 ms
Mediapipe Pose Landmarker: 77.08 ms
Mediapipe Pose Landmarker: 77.30 ms
Mediapipe Pose Landmarker: 76.91 ms
Mediapipe Pose Landmarker: 77.56 ms
Mediapipe Pose Landmarker: 77.29 ms
Mediapipe Pose Landmarker: 77.36 ms
Mediapipe Pose Landmarker: 77.28 ms
Mediapipe Pose Landmarker: 76.97 ms
Mediapipe Pose Landmarker: 77.37 ms
Mediapipe Pose Landmarker: 77.44 ms
Mediapipe Pose Landmarker: 77.33 ms
Mediapipe Pose Landmarker: 77.42 ms
Mediapipe Pose Landmarker: 77.43 ms
Mediapipe Pose Landmarker: 76.98 ms
Mediapipe Pose Landmarker: 77.23 ms
Mediapipe Pose Landmarker: 77.09 ms
Mediapipe Pose Landmarker: 77.09 ms
Mediapipe Pose Landmarker: 76.98 ms
Mediapipe Pose Landmarker: 77.19 ms
Mediapipe Pose Landmarker: 77.53 ms
Mediapipe Pose Landmarker: 78.32 ms
Mediapipe Pose Landmarker: 77.20 ms
Mediapipe Pose Landmarker: 77.01 ms
Mediapipe Pose Landmarker: 78.25 ms
Mediapipe Pose Landmarker: 77.43 ms
Mediapipe Pose Landmarker: 77.70 ms
Mediapipe Pose Landmarker: 77.87 ms
Mediapipe Pose Landmarker: 77.60 ms
Mediapipe Pose Landmarker: 77.37 ms
Mediapipe Pose Landmarker: 77.04 ms
Mediapipe Pose Landmarker: 77.05 ms
Mediapipe Pose Landmarker: 77.00 ms
Mediapipe Pose Landmarker: 77.22 ms
Mediapipe Pose Landmarker: 77.16 ms
Mediapipe Pose Landmarker: 77.45 ms
Mediapipe Pose Landmarker: 77.40 ms
Mediapipe Pose Landmarker: 77.11 ms

[Nibosi0501]

mediapipe==0.10.31

Since solutions were no longer available, we changed to implementing Landmark's rendering process independently.

In this version, the memory leak issue when enabling the GPU had been resolved.
However, even with the GPU enabled, there seems to be no difference in execution time compared to the CPU.

This is something I recently noticed: with mediapipe==0.10.20, the log when enabling the GPU is:

`I0000 00:00:1769011147.776678 2963632 gl_context.cc:369] GL version: 2.1 (2.1 Metal - 83.1), renderer: Apple M1
INFO: Created TensorFlow Lite delegate for Metal.

However, in mediapipe=0.10.31, the log when GPU is enabled shows:

`I0000 00:00:1769012003.034233 2972780 gl_context.cc:407] GL version: 2.1 (2.1 Metal - 83.1), renderer: Apple M1
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.

import mediapipe as mp
from mediapipe.tasks import python
from mediapipe.tasks.python import vision

import cv2 as cv
import numpy as np

import time

BaseOptions = mp.tasks.BaseOptions
PoseLandmarker = mp.tasks.vision.PoseLandmarker
PoseLandmarkerOptions = mp.tasks.vision.PoseLandmarkerOptions
PoseLandmarkerResult = mp.tasks.vision.PoseLandmarkerResult
VisionRunningMode = mp.tasks.vision.RunningMode

POSE_CONNECTIONS = frozenset([(0, 1), (1, 2), (2, 3), (3, 7), (0, 4), (4, 5),
                              (5, 6), (6, 8), (9, 10), (11, 12), (11, 13),
                              (13, 15), (15, 17), (15, 19), (15, 21), (17, 19),
                              (12, 14), (14, 16), (16, 18), (16, 20), (16, 22),
                              (18, 20), (11, 23), (12, 24), (23, 24), (23, 25),
                              (24, 26), (25, 27), (26, 28), (27, 29), (28, 30),
                              (29, 31), (30, 32), (27, 31), (28, 32)])

base_options = python.BaseOptions(
    model_asset_path='./models/mp/pose_landmarker_heavy.task',
    # delegate=python.BaseOptions.Delegate.CPU,  # CPU
    delegate=python.BaseOptions.Delegate.GPU    # GPU
)

options = PoseLandmarkerOptions(
    #base_options=BaseOptions(model_asset_path='./models/mp/pose_landmarker_heavy.task'),   # CPU
    base_options=base_options,                             # GPU
    running_mode=VisionRunningMode.VIDEO,
    num_poses = 4,
    min_pose_detection_confidence=0.7
    )

def main():

    frame_cnt = 0
    
    cap = cv.VideoCapture('./dance.mov')

    if not cap.isOpened():
        print("Cannot open camera")
        return
    
    with PoseLandmarker.create_from_options(options) as landmarker:
        while True:
            ret, frame = cap.read()

            if not ret:
                print("Cannot receive frame")
                break

            frame_cnt += 1
            if frame_cnt == 100:
                break

            mediapipe_start = time.time()

            # Convert BGR image to RGB
            # GPU
            rgba_frame = cv.cvtColor(frame, cv.COLOR_BGR2RGBA)
            mp_image = mp.Image(image_format=mp.ImageFormat.SRGBA, data=rgba_frame)

            # CPU
            #rgb_image = cv.cvtColor(frame, cv.COLOR_BGR2RGB)
            #mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=rgb_image)

            # landmark detection
            pose_landmarker_result = landmarker.detect_for_video(mp_image, int(time.time() * 1000))

            mediapipe_end = time.time()
            elapsed_time = mediapipe_end - mediapipe_start
            print(f"Mediapipe Pose Landmarker: {elapsed_time*1000:.2f} ms")

            annotated_image = frame.copy()
            if pose_landmarker_result.pose_landmarks is not None:
                for pose_landmarks in pose_landmarker_result.pose_landmarks:
                    for connection in POSE_CONNECTIONS:
                        start_idx = connection[0]
                        end_idx = connection[1]

                        start_landmark = pose_landmarks[start_idx]
                        end_landmark = pose_landmarks[end_idx]

                        start_x, start_y = normalized2pixel(start_landmark.x, start_landmark.y, frame.shape[1], frame.shape[0])
                        end_x, end_y = normalized2pixel(end_landmark.x, end_landmark.y, frame.shape[1], frame.shape[0])

                        cv.line(annotated_image, (start_x, start_y), (end_x, end_y), (0, 255, 0), 2)
                    
                    for idx, landmark in enumerate(pose_landmarks):
                        x_px, y_px = normalized2pixel(landmark.x, landmark.y, frame.shape[1], frame.shape[0])
                        cv.circle(annotated_image, (x_px, y_px), 3, (0, 0, 255), -1)

            cv.imshow('Pose Landmarks', annotated_image)

            if cv.waitKey(1) & 0xFF == 27:
                break

def normalized2pixel(normalized_x, normalized_y, image_width, image_height):
    x_px = min(int(normalized_x * image_width), image_width - 1)
    y_px = min(int(normalized_y * image_height), image_height - 1)
    return x_px, y_px

if __name__ == "__main__":
    main()

GPU output

I0000 00:00:1769012003.034233 2972780 gl_context.cc:407] GL version: 2.1 (2.1 Metal - 83.1), renderer: Apple M1
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
W0000 00:00:1769012003.093855 2972782 inference_feedback_manager.cc:121] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1769012003.192838 2972785 inference_feedback_manager.cc:121] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1769012003.286996 2972782 landmark_projection_calculator.cc:81] Using NORM_RECT without IMAGE_DIMENSIONS is only supported for the square ROI. Provide IMAGE_DIMENSIONS or use PROJECTION_MATRIX.
Mediapipe Pose Landmarker: 89.86 ms
Mediapipe Pose Landmarker: 78.14 ms
Mediapipe Pose Landmarker: 76.77 ms
Mediapipe Pose Landmarker: 77.34 ms
Mediapipe Pose Landmarker: 76.60 ms
Mediapipe Pose Landmarker: 77.01 ms
Mediapipe Pose Landmarker: 76.43 ms
Mediapipe Pose Landmarker: 76.57 ms
Mediapipe Pose Landmarker: 76.87 ms
Mediapipe Pose Landmarker: 76.56 ms
Mediapipe Pose Landmarker: 76.81 ms
Mediapipe Pose Landmarker: 76.81 ms
Mediapipe Pose Landmarker: 77.25 ms
Mediapipe Pose Landmarker: 76.76 ms
Mediapipe Pose Landmarker: 76.44 ms
Mediapipe Pose Landmarker: 76.59 ms
Mediapipe Pose Landmarker: 76.48 ms
Mediapipe Pose Landmarker: 76.49 ms
Mediapipe Pose Landmarker: 76.43 ms
Mediapipe Pose Landmarker: 76.66 ms
Mediapipe Pose Landmarker: 76.53 ms
Mediapipe Pose Landmarker: 76.42 ms
Mediapipe Pose Landmarker: 76.35 ms
Mediapipe Pose Landmarker: 76.29 ms
Mediapipe Pose Landmarker: 78.62 ms
Mediapipe Pose Landmarker: 76.55 ms
Mediapipe Pose Landmarker: 76.53 ms
Mediapipe Pose Landmarker: 76.53 ms
Mediapipe Pose Landmarker: 76.44 ms
Mediapipe Pose Landmarker: 76.84 ms
Mediapipe Pose Landmarker: 76.86 ms
Mediapipe Pose Landmarker: 76.64 ms
Mediapipe Pose Landmarker: 76.61 ms
Mediapipe Pose Landmarker: 76.82 ms
Mediapipe Pose Landmarker: 76.28 ms
Mediapipe Pose Landmarker: 76.63 ms
Mediapipe Pose Landmarker: 76.61 ms
Mediapipe Pose Landmarker: 76.58 ms
Mediapipe Pose Landmarker: 76.68 ms
Mediapipe Pose Landmarker: 76.26 ms
Mediapipe Pose Landmarker: 76.41 ms
Mediapipe Pose Landmarker: 76.41 ms
Mediapipe Pose Landmarker: 76.68 ms
Mediapipe Pose Landmarker: 76.52 ms
Mediapipe Pose Landmarker: 76.61 ms
Mediapipe Pose Landmarker: 76.60 ms
Mediapipe Pose Landmarker: 77.05 ms
Mediapipe Pose Landmarker: 76.41 ms
Mediapipe Pose Landmarker: 76.44 ms
Mediapipe Pose Landmarker: 76.66 ms
Mediapipe Pose Landmarker: 76.55 ms
Mediapipe Pose Landmarker: 76.35 ms
Mediapipe Pose Landmarker: 76.60 ms
Mediapipe Pose Landmarker: 76.78 ms
Mediapipe Pose Landmarker: 76.62 ms
Mediapipe Pose Landmarker: 76.65 ms
Mediapipe Pose Landmarker: 76.48 ms
Mediapipe Pose Landmarker: 77.39 ms
Mediapipe Pose Landmarker: 76.99 ms
Mediapipe Pose Landmarker: 76.43 ms
Mediapipe Pose Landmarker: 76.43 ms
Mediapipe Pose Landmarker: 76.38 ms
Mediapipe Pose Landmarker: 76.36 ms
Mediapipe Pose Landmarker: 76.51 ms
Mediapipe Pose Landmarker: 76.44 ms
Mediapipe Pose Landmarker: 76.42 ms
Mediapipe Pose Landmarker: 76.31 ms
Mediapipe Pose Landmarker: 76.68 ms
Mediapipe Pose Landmarker: 77.08 ms
Mediapipe Pose Landmarker: 76.56 ms
Mediapipe Pose Landmarker: 76.58 ms
Mediapipe Pose Landmarker: 76.46 ms
Mediapipe Pose Landmarker: 76.37 ms
Mediapipe Pose Landmarker: 76.46 ms
Mediapipe Pose Landmarker: 76.66 ms
Mediapipe Pose Landmarker: 76.34 ms
Mediapipe Pose Landmarker: 76.63 ms
Mediapipe Pose Landmarker: 77.03 ms
Mediapipe Pose Landmarker: 76.30 ms
Mediapipe Pose Landmarker: 76.37 ms
Mediapipe Pose Landmarker: 76.54 ms
Mediapipe Pose Landmarker: 76.53 ms
Mediapipe Pose Landmarker: 76.38 ms
Mediapipe Pose Landmarker: 76.43 ms
Mediapipe Pose Landmarker: 76.36 ms
Mediapipe Pose Landmarker: 76.36 ms
Mediapipe Pose Landmarker: 76.49 ms
Mediapipe Pose Landmarker: 76.72 ms
Mediapipe Pose Landmarker: 77.81 ms
Mediapipe Pose Landmarker: 76.63 ms
Mediapipe Pose Landmarker: 76.97 ms
Mediapipe Pose Landmarker: 79.14 ms
Mediapipe Pose Landmarker: 76.53 ms
Mediapipe Pose Landmarker: 76.41 ms
Mediapipe Pose Landmarker: 76.28 ms
Mediapipe Pose Landmarker: 75.99 ms
Mediapipe Pose Landmarker: 76.56 ms
Mediapipe Pose Landmarker: 76.50 ms
Mediapipe Pose Landmarker: 76.84 ms

CPU output

I0000 00:00:1769012084.270656 2973500 gl_context.cc:407] GL version: 2.1 (2.1 Metal - 83.1), renderer: Apple M1
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
W0000 00:00:1769012084.326727 2973502 inference_feedback_manager.cc:121] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1769012084.405965 2973504 inference_feedback_manager.cc:121] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1769012084.501666 2973508 landmark_projection_calculator.cc:81] Using NORM_RECT without IMAGE_DIMENSIONS is only supported for the square ROI. Provide IMAGE_DIMENSIONS or use PROJECTION_MATRIX.
Mediapipe Pose Landmarker: 90.59 ms
Mediapipe Pose Landmarker: 76.86 ms
Mediapipe Pose Landmarker: 76.84 ms
Mediapipe Pose Landmarker: 76.61 ms
Mediapipe Pose Landmarker: 76.23 ms
Mediapipe Pose Landmarker: 76.78 ms
Mediapipe Pose Landmarker: 76.76 ms
Mediapipe Pose Landmarker: 76.53 ms
Mediapipe Pose Landmarker: 75.65 ms
Mediapipe Pose Landmarker: 76.73 ms
Mediapipe Pose Landmarker: 79.54 ms
Mediapipe Pose Landmarker: 76.80 ms
Mediapipe Pose Landmarker: 76.83 ms
Mediapipe Pose Landmarker: 76.65 ms
Mediapipe Pose Landmarker: 76.42 ms
Mediapipe Pose Landmarker: 78.00 ms
Mediapipe Pose Landmarker: 76.86 ms
Mediapipe Pose Landmarker: 76.86 ms
Mediapipe Pose Landmarker: 76.65 ms
Mediapipe Pose Landmarker: 76.56 ms
Mediapipe Pose Landmarker: 76.13 ms
Mediapipe Pose Landmarker: 77.49 ms
Mediapipe Pose Landmarker: 76.59 ms
Mediapipe Pose Landmarker: 76.47 ms
Mediapipe Pose Landmarker: 76.79 ms
Mediapipe Pose Landmarker: 76.56 ms
Mediapipe Pose Landmarker: 76.32 ms
Mediapipe Pose Landmarker: 76.54 ms
Mediapipe Pose Landmarker: 76.44 ms
Mediapipe Pose Landmarker: 76.39 ms
Mediapipe Pose Landmarker: 76.34 ms
Mediapipe Pose Landmarker: 76.30 ms
Mediapipe Pose Landmarker: 77.03 ms
Mediapipe Pose Landmarker: 76.70 ms
Mediapipe Pose Landmarker: 76.54 ms
Mediapipe Pose Landmarker: 76.72 ms
Mediapipe Pose Landmarker: 76.57 ms
Mediapipe Pose Landmarker: 76.50 ms
Mediapipe Pose Landmarker: 76.52 ms
Mediapipe Pose Landmarker: 76.36 ms
Mediapipe Pose Landmarker: 76.33 ms
Mediapipe Pose Landmarker: 76.55 ms
Mediapipe Pose Landmarker: 76.51 ms
Mediapipe Pose Landmarker: 78.04 ms
Mediapipe Pose Landmarker: 78.34 ms
Mediapipe Pose Landmarker: 76.38 ms
Mediapipe Pose Landmarker: 76.47 ms
Mediapipe Pose Landmarker: 76.45 ms
Mediapipe Pose Landmarker: 76.38 ms
Mediapipe Pose Landmarker: 76.62 ms
Mediapipe Pose Landmarker: 76.52 ms
Mediapipe Pose Landmarker: 76.91 ms
Mediapipe Pose Landmarker: 76.54 ms
Mediapipe Pose Landmarker: 76.43 ms
Mediapipe Pose Landmarker: 76.81 ms
Mediapipe Pose Landmarker: 76.42 ms
Mediapipe Pose Landmarker: 76.55 ms
Mediapipe Pose Landmarker: 77.31 ms
Mediapipe Pose Landmarker: 76.45 ms
Mediapipe Pose Landmarker: 76.64 ms
Mediapipe Pose Landmarker: 76.42 ms
Mediapipe Pose Landmarker: 76.57 ms
Mediapipe Pose Landmarker: 76.28 ms
Mediapipe Pose Landmarker: 76.42 ms
Mediapipe Pose Landmarker: 76.52 ms
Mediapipe Pose Landmarker: 76.27 ms
Mediapipe Pose Landmarker: 76.42 ms
Mediapipe Pose Landmarker: 76.51 ms
Mediapipe Pose Landmarker: 76.35 ms
Mediapipe Pose Landmarker: 76.61 ms
Mediapipe Pose Landmarker: 76.52 ms
Mediapipe Pose Landmarker: 77.71 ms
Mediapipe Pose Landmarker: 76.89 ms
Mediapipe Pose Landmarker: 76.68 ms
Mediapipe Pose Landmarker: 76.56 ms
Mediapipe Pose Landmarker: 76.82 ms
Mediapipe Pose Landmarker: 76.27 ms
Mediapipe Pose Landmarker: 77.66 ms
Mediapipe Pose Landmarker: 76.77 ms
Mediapipe Pose Landmarker: 76.45 ms
Mediapipe Pose Landmarker: 76.64 ms
Mediapipe Pose Landmarker: 76.65 ms
Mediapipe Pose Landmarker: 76.59 ms
Mediapipe Pose Landmarker: 76.46 ms
Mediapipe Pose Landmarker: 76.66 ms
Mediapipe Pose Landmarker: 76.58 ms
Mediapipe Pose Landmarker: 76.62 ms
Mediapipe Pose Landmarker: 76.39 ms
Mediapipe Pose Landmarker: 76.74 ms
Mediapipe Pose Landmarker: 76.83 ms
Mediapipe Pose Landmarker: 76.38 ms
Mediapipe Pose Landmarker: 76.37 ms
Mediapipe Pose Landmarker: 76.46 ms
Mediapipe Pose Landmarker: 76.49 ms
Mediapipe Pose Landmarker: 76.64 ms
Mediapipe Pose Landmarker: 77.11 ms
Mediapipe Pose Landmarker: 76.30 ms
Mediapipe Pose Landmarker: 76.52 ms
Mediapipe Pose Landmarker: 76.88 ms

[Nibosi0501]

Hi @akashvverma1995 , Thanks for the response.
Looking at the log output, I realized that in my code using MediaPipe 0.10.31, the GPU might not be enabled.
Is there something wrong with my code?
If there are any other logs you need besides these, please let me know.

[akashvverma1995]

Hi @schmidt-sebastian

Could you please look into this? I tried from my end as well, and I’m facing the same issue—the GPU is not getting enabled.

[schmidt-sebastian]

Sorry, we had an issue with the GPU delegate in 0.10.31. We just published 0.10.32 that sets it correctly again.