main_noweb.py

# Pose inferencing
import mmpose
from mmpose.apis import MMPoseInferencer

# Ultralytics
from ultralytics import YOLO
import torch

# Gradio
import gradio as gr
import moviepy.editor as moviepy


# System and files
import os
import glob
import uuid

# Image manipulation
import numpy as np
import cv2

print(torch.__version__)
# Use GPU if available
if torch.cuda.is_available():
    device = torch.device("cuda")
else:
    device = torch.device("cpu")

os.system("nvidia-smi")

print("[INFO]: Imported modules!")
human = MMPoseInferencer(pose2d="human")
hand = MMPoseInferencer("hand")
human3d = MMPoseInferencer(pose3d="human3d")
track_model = YOLO('yolov8n.pt')  # Load an official Detect model

print("[INFO]: Downloaded models!")

def check_extension(video):
    split_tup = os.path.splitext(video)

    # extract the file name and extension
    file_name = split_tup[0]
    file_extension = split_tup[1]

    if file_extension != ".mp4":
        print("Converting to mp4")
        clip = moviepy.VideoFileClip(video)

        video = file_name+".mp4"
        clip.write_videofile(video)
    
    return video


def tracking(video, model, boxes=True):
    print("[INFO] Is cuda available? ", torch.cuda.is_available())
    print(device)

    print("[INFO] Loading model...")
    # Load an official or custom model

    # Perform tracking with the model
    print("[INFO] Starting tracking!")
    # https://docs.ultralytics.com/modes/predict/
    annotated_frame = model(video, boxes=boxes, device=device)

    return annotated_frame

def show_tracking(video_content):

        # https://docs.ultralytics.com/datasets/detect/coco/
        video = cv2.VideoCapture(video_content)

        # Track
        video_track = tracking(video_content, track_model.track)

        # Prepare to save video
        #out_file = os.path.join(vis_out_dir, "track.mp4")
        out_file = "track.mp4"
        print("[INFO]: TRACK", out_file)

        fourcc = cv2.VideoWriter_fourcc(*"mp4v")  # Codec for MP4 video
        fps = video.get(cv2.CAP_PROP_FPS)
        height, width, _ = video_track[0][0].orig_img.shape
        size = (width,height)

        out_track = cv2.VideoWriter(out_file, fourcc, fps, size)

        # Go through frames and write them 
        for frame_track in video_track:
            result_track = frame_track[0].plot()  # plot a BGR numpy array of predictions
            out_track.write(result_track)

        print("[INFO] Done with frames")
        #print(type(result_pose)) numpy ndarray
    
        out_track.release()

        video.release()
        cv2.destroyAllWindows() # Closing window

        return out_file


def pose3d(video):
    video = check_extension(video)
    print(device)
    temp_3d = human3d


    # Define new unique folder
    add_dir = str(uuid.uuid4())
    vis_out_dir = os.path.join("/".join(video.split("/")[:-1]), add_dir)
    os.makedirs(vis_out_dir)

    result_generator = temp_3d(video, 
                                vis_out_dir = vis_out_dir,
                                thickness=4,
                                radius = 5,
                                return_vis=True,
                                kpt_thr=0.3,
                                rebase_keypoint_height=True,
                                device=device)    
    
    result = [result for result in result_generator] #next(result_generator)        

    out_file = glob.glob(os.path.join(vis_out_dir, "*.mp4")) #+ glob.glob(os.path.join(vis_out_dir, "*.webm")) 
    
    # Reinitialize
    return "".join(out_file)


def pose2d(video, kpt_threshold):
    video = check_extension(video)
    print(device)

    # Define new unique folder
    add_dir = str(uuid.uuid4())
    vis_out_dir = os.path.join("/".join(video.split("/")[:-1]), add_dir)

    os.makedirs(add_dir)

    result_generator = human(video, 
                            vis_out_dir = add_dir,
                            #return_vis=True,
                            radius = 5,
                            thickness=4,
                            rebase_keypoint_height=True,
                            kpt_thr=kpt_threshold,
                            device=device,
                            pred_out_dir = add_dir
                            )    
    
    result = [result for result in result_generator] #next(result_generator)        

    out_file = glob.glob(os.path.join(add_dir, "*.mp4")) #+ glob.glob(os.path.join(vis_out_dir, "*.webm")) 
    kpoints = glob.glob(os.path.join(add_dir, "*.json"))
    print(os.listdir(glob.glob(os.path.join(add_dir, "*.mp4"))))
    print(glob.glob(os.path.join(add_dir, "*.json")))
    print(os.listdir(add_dir))

    return "".join(out_file), "".join(kpoints)


def pose2dhand(video, kpt_threshold):
    video = check_extension(video)
    print(device)
    # ultraltics

    # Define new unique folder
    add_dir = str(uuid.uuid4())
    vis_out_dir = os.path.join("/".join(video.split("/")[:-1]), add_dir)
    os.makedirs(vis_out_dir)

    result_generator = hand(video, 
                                 vis_out_dir = vis_out_dir,
                                 return_vis=True,
                                 thickness = 4,
                                 radius = 5,
                                 rebase_keypoint_height=True,
                                 kpt_thr=kpt_threshold,
                                 device=device)    
    
    result = [result for result in result_generator] #next(result_generator)        

    out_file = glob.glob(os.path.join(vis_out_dir, "*.mp4")) #+ glob.glob(os.path.join(vis_out_dir, "*.webm")) 
   
    return "".join(out_file)

block = gr.Blocks()
with block:
    with gr.Column():       
        with gr.Tab("Upload video"):
            with gr.Column():
                with gr.Row():
                    with gr.Column():
                        video_input = gr.Video(source="upload", type="filepath", height=612)
                        # Insert slider with kpt_thr
                        file_kpthr = gr.Slider(0, 1, value=0.3, label='Keypoint threshold')
                        with gr.Row():
                            submit_pose_file = gr.Button("Make 2d pose estimation", variant="primary")
                            submit_pose3d_file = gr.Button("Make 3d pose estimation", variant="primary")
                            submit_hand_file = gr.Button("Make 2d hand estimation", variant="primary")
                            submit_detect_file = gr.Button("Detect and track objects", variant="primary")

                with gr.Row():
                    video_output1 = gr.PlayableVideo(height=512,  label = "Estimate human 2d poses", show_label=True)
                    video_output2 = gr.PlayableVideo(height=512,  label = "Estimate human 3d poses", show_label=True)
                    video_output3 = gr.PlayableVideo(height=512,  label = "Estimate human hand poses", show_label=True)
                    video_output4 = gr.Video(height=512, label = "Detection and tracking", show_label=True, format="mp4")
                jsonoutput = gr.Code()


        with gr.Tab("General information"):
            gr.Markdown(""" 
                        \n # Information about the models 

                        \n ## Pose models: 
                        
                        \n All the pose estimation models come from the library [MMpose](https://github.com/open-mmlab/mmpose). It is a library for human pose estimation that provides pre-trained models for 2D and 3D pose estimation. 

                        \n The 2D pose model is used for estimating the 2D coordinates of human body joints from an image or a video frame. The model uses a convolutional neural network (CNN) to predict the joint locations and their confidence scores. 
                        
                        \n The 2D hand model is a specialized version of the 2D pose model that is designed for hand pose estimation. It uses a similar CNN architecture to the 2D pose model but is trained specifically for detecting the joints in the hand. 
                        
                        \n The 3D pose model is used for estimating the 3D coordinates of human body joints from an image or a video frame. The model uses a combination of 2D pose estimation and depth estimation to infer the 3D joint locations. 
                        
                        \n 

                        \n ## Detection and tracking:                            
                        
                        \n The tracking method in the Ultralight's YOLOv8 model is used for object tracking in videos. It takes a video file or a camera stream as input and returns the tracked objects in each frame. The method uses the COCO dataset classes for object detection and tracking. 
                        
                        \n The COCO dataset contains 80 classes of objects such as person, car, bicycle, etc. See https://docs.ultralytics.com/datasets/detect/coco/ for all available classes. The tracking method uses the COCO classes to detect and track the objects in the video frames. The tracked objects are represented as bounding boxes with labels indicating the class of the object.""")
            gr.Markdown("You can load the keypoints in python in the following way: ")
            gr.Code(
                    value="""def hello_world():
                                    return "Hello, world!"
                
                            print(hello_world())""",
                    language="python",
                    interactive=False,
                    show_label=False,
                )
            

        # From file
        submit_pose_file.click(fn=pose2d, 
                            inputs=  [video_input, file_kpthr], 
                            outputs = [video_output1, jsonoutput],
                            queue=True)
        
        submit_pose3d_file.click(fn=pose3d, 
                                inputs= video_input, 
                                outputs = video_output2,
                                queue=True)
        
        submit_hand_file.click(fn=pose2dhand, 
                            inputs= [video_input, file_kpthr], 
                            outputs = video_output3,
                            queue=True)
        
        submit_detect_file.click(fn=show_tracking, 
                                inputs= video_input, 
                                outputs = video_output4,
                                queue=True)
        
if __name__ == "__main__":
    block.queue(
                concurrency_count=40, # When you increase the concurrency_count parameter in queue(), max_threads() in launch() is automatically increased as well.
                max_size=25, # Maximum number of requests that the queue processes
                api_open = False # When creating a Gradio demo, you may want to restrict all traffic to happen through the user interface as opposed to the programmatic API that is automatically created for your Gradio demo.  
                    )  # https://www.gradio.app/guides/setting-up-a-demo-for-maximum-performance
    block.launch(
                server_name="0.0.0.0", 
                 server_port=7860,
                 auth=("novouser", "bstad2023")
                 )

    # The total concurrency = number of processors * 10.
    # 4vCPU 15 GB ram 40GV VRAM = 40?