OpenDWM/src/dwm/streaming.py at main · SenseTime-FVG/OpenDWM · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
import argparse
import av
import dwm.common
import einops
import json
import numpy as np
from PIL import Image
import queue
import time
import torch


def create_parser():
    parser = argparse.ArgumentParser(
        description="The script to finetune a stable diffusion model to the "
        "driving dataset.")
    parser.add_argument(
        "-c", "--config-path", type=str, required=True,
        help="The config to load the train model and dataset.")
    parser.add_argument(
        "-l", "--log-path", type=str, required=True,
        help="The path to save log files.")
    parser.add_argument(
        "-s", "--streaming-path", type=str, required=True,
        help="The path to upload the video stream.")
    parser.add_argument(
        "-f", "--format", default="rtsp", type=str,
        help="The streaming format.")
    parser.add_argument(
        "--fps", default=2, type=int,
        help="The streaming FPS.")
    parser.add_argument(
        "-vcodec", "--video-encoder", default="libx264", type=str,
        help="The video encoder type.")
    parser.add_argument(
        "--pix-fmt", default="yuv420p", type=str,
        help="The pixel format.")
    return parser


def merge_multiview_images(pipeline_frame, data_condition=None):
    image_data = np.concatenate([np.asarray(i) for i in pipeline_frame], 1)
    if data_condition is not None:
        _3dbox_data = torch.nn.functional.interpolate(
            einops.rearrange(
                data_condition["3dbox_images"],
                "b t v c h w -> b c h (t v w)"),
            image_data.shape[:2]
        )[0].permute(1, 2, 0).numpy()
        hdmap_data = torch.nn.functional.interpolate(
            einops.rearrange(
                data_condition["hdmap_images"],
                "b t v c h w -> b c h (t v w)"),
            image_data.shape[:2]
        )[0].permute(1, 2, 0).numpy()
        condition_data = np.maximum(_3dbox_data, hdmap_data)
        condition_ahpla = np.max(condition_data, -1, keepdims=True) * 0.6
        image_data = (
            condition_data * 255 * condition_ahpla +
            image_data * (1 - condition_ahpla)
        ).astype(np.uint8)

    return Image.fromarray(image_data)


if __name__ == "__main__":
    parser = create_parser()
    args = parser.parse_args()

    with open(args.config_path, "r", encoding="utf-8") as f:
        config = json.load(f)

    # setup the global state
    if "global_state" in config:
        for key, value in config["global_state"].items():
            dwm.common.global_state[key] = \
                dwm.common.create_instance_from_config(value)

    # load the pipeline including the models
    pipeline = dwm.common.create_instance_from_config(
        config["pipeline"], output_path=args.log_path, config=config,
        device=torch.device(config["device"]))
    print("The pipeline is loaded.")

    data_adapter = dwm.common.create_instance_from_config(
        config["data_adapter"])

    size = pipeline.inference_config["preview_image_size"]
    latent_shape = (
        1, pipeline.inference_config["sequence_length_per_iteration"],
        len(data_adapter.sensor_channels), pipeline.vae.config.latent_channels,
        config["latent_size"][0], config["latent_size"][1]
    )
    pipeline.reset_streaming(latent_shape, "pil")

    streaming_state = {}
    data_queue = queue.Queue()
    with av.open(
        args.streaming_path, mode="w", format=args.format,
        container_options=config.get("container_options", {})
    ) as container:
        stream = container.add_stream(args.video_encoder, args.fps)
        stream.pix_fmt = args.pix_fmt
        stream.options = config.get("stream_options", {})
        while True:
            data = data_adapter.query_data()
            data_queue.put_nowait(data)
            pipeline.send_frame_condition(data)
            pipeline_frame = pipeline.receive_frame()
            if pipeline_frame is None:
                continue

            matched_data = data_queue.get_nowait()
            image = merge_multiview_images(
                pipeline_frame,
                (
                    matched_data
                    if config.get("preview_condition", False)
                    else None
                ))
            if not streaming_state.get("is_frame_size_set", False):
                stream.width = image.width
                stream.height = image.height
                streaming_state["is_frame_size_set"] = True

            while (
                "expected_time" in streaming_state and
                time.time() < streaming_state["expected_time"]
            ):
                time.sleep(0.01)

            frame = av.VideoFrame.from_image(image)
            for p in stream.encode(frame):
                container.mux(p)

            streaming_state["expected_time"] = (
                time.time()
                if "expected_time" not in streaming_state
                else streaming_state["expected_time"]
            ) + 1 / args.fps
            print("{:.1f}".format(streaming_state["expected_time"]))