Propheseeイベントヒストグラム

python -m pip install "libreyolo[rfdetr,onnx] @ git+https://github.com/LibreYOLO/libreyolo.git@event-histogram-input"

import numpy as np  def pack_histogram(frame, *, layout, polarity, count_divisor=1.0):    """Convert known producer output to HWC positive/negative counts."""    values = np.asarray(frame)    if values.dtype.kind not in "uif" or values.ndim != 3:        raise ValueError("Expected a numerical three-dimensional histogram")    if layout == "CHW":        values = values.transpose(1, 2, 0)    elif layout != "HWC":        raise ValueError("Declare layout as HWC or CHW")    if values.shape[2] != 2 or min(values.shape[:2]) == 0:        raise ValueError("Expected two nonempty polarity planes")    if polarity == "negative_positive":        values = values[..., [1, 0]]    elif polarity != "positive_negative":        raise ValueError("Declare the producer's polarity order")    if not np.isfinite(count_divisor) or count_divisor <= 0:        raise ValueError("count_divisor must be positive and finite")    with np.errstate(over="ignore", invalid="ignore"):        counts = values.astype(np.float32) * count_divisor    if not np.isfinite(counts).all() or (counts < 0).any():        raise ValueError("Counts must be finite, nonnegative float32 values")    return np.ascontiguousarray(counts)  # Replace this path with a numerical histogram from your producer.sdk_histogram = np.load("sdk-histogram.npy", allow_pickle=False)histogram = pack_histogram(    sdk_histogram, layout="HWC", polarity="negative_positive",    count_divisor=1.0,  # Counts already: no producer normalization to undo.)np.save("histogram.npy", histogram, allow_pickle=False)

import numpy as npfrom metavision_sdk_core import EventPreprocessor  def histogram_from_sdk_chunks(chunks, *, width, height, start_us,                              window_us=40_000, scale=16.0):    """chunks must cover one known window, using SDK EventCD arrays."""    processor = EventPreprocessor.create_HistoProcessor(        input_event_width=width,        input_event_height=height,        max_incr_per_pixel=1.0,  # 1増分を1カウントとして扱う        clip_value_after_normalization=scale,        use_CHW=False,          # SDK出力はHWC、負の極性が先        scale_width=1.0,        scale_height=1.0,    )    frame = processor.init_output_tensor()    frame.fill(0)    end_us = start_us + window_us    for events in chunks:        if len(events) == 0:            continue        if ((events["t"] < start_us) | (events["t"] >= end_us)).any():            raise ValueError("Split events into the declared window first")        if ((events["x"] < 0) | (events["x"] >= width)                | (events["y"] < 0) | (events["y"] >= height)).any():            raise ValueError("Event coordinates do not match the sensor canvas")        if not np.isin(events["p"], [0, 1]).all():            raise ValueError("Expected SDK polarity values 0 and 1")        processor.process_events(start_us, np.ascontiguousarray(events), frame)    return np.ascontiguousarray(frame[..., [1, 0]], dtype=np.float32)

path: /absolute/path/to/event-datasettrain: images/trainval: images/valnames:  0: personinput_profile:  format: event_histogram  layout: HWC  polarity: positive_negative  encoding: counts  scale: 16.0  window_us: 40000

0 0.5 0.5 0.25 0.6

import numpy as npfrom PIL import Imagefrom libreyolo.utils.event_histogram import load_histogram, visualize_histogram histogram = load_histogram("histogram.npy")print(histogram.shape, histogram.dtype)print(float(histogram.min()), float(histogram.max()))rgb = visualize_histogram(histogram, scale=16.0)Image.fromarray(rgb).save("histogram-view.png")

from libreyolo import LibreYOLO, LibreYOLO9 model = LibreYOLO9(None, size="t", device="cpu")run = model.train(    data="data.yaml", pretrained=False,    epochs=100, batch=4, imgsz=128, workers=0, device="cpu",)trained = LibreYOLO(run["best_checkpoint"], device="cpu")trained.save("event-detector.pt")

libreyolo train model=LibreRFDETRn.pt pretrained=False data=data.yaml epochs=100 batch=4 imgsz=128 workers=0 device=cpu project=runs name=event-rfdetr exist_ok=True

from libreyolo import LibreYOLO trained = LibreYOLO("runs/event-rfdetr/weights/best.pt", device="cpu")trained.save("event-detector.pt")

import numpy as npfrom libreyolo import LibreYOLO model = LibreYOLO("event-detector.pt", device="cpu")result = model.predict("histogram.npy", save=True)print(result.boxes.xyxy)print(result.boxes.conf) histogram = np.load("histogram.npy", allow_pickle=False)result = model.predict(histogram)results = model.predict([histogram, histogram.copy()])

libreyolo predict model=event-detector.pt source=histogram.npy save=True

from libreyolo import LibreYOLO model = LibreYOLO("event-detector.pt", device="cpu")metrics = model.val(data="data.yaml", split="val", batch=4, workers=0)print(metrics["metrics/mAP50-95"])

from libreyolo import LibreYOLO model = LibreYOLO("event-detector.pt", device="cpu")onnx_path = model.export(    format="onnx", imgsz=128, dynamic=False, simplify=False,    output_path="event-detector.onnx",)runtime = LibreYOLO(onnx_path, device="cpu")result = runtime.predict("histogram.npy", save=True)