@nina-vilela
Despite following the tutorial closely, I’m encountering a segmentation fault when the OCR inference runs after a detection is made. The script runs fine if there are no detections, but as soon as the OCR model processes its first input, it crashes with a segmentation fault.
I’ve tried several approaches, including ensuring that I’m using a single VDevice instance with the model scheduler enabled and the same group_id. I also attempted to use threading locks to synchronize access to the Hailo SDK, but the issue persists.
Here’s a simplified version of my code focusing on the Hailo SDK and threading:
import threading
from queue import Queue
import numpy as np
from functools import partial
from hailo_platform import VDevice, HailoSchedulingAlgorithm, FormatType
frame_queue = Queue(maxsize=100)
ocr_queue = Queue(maxsize=100)
def grab_frames():
while True:
# Simulate grabbing a frame
frame = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
frame_queue.put(frame)
def detection_thread(vdevice, detection_model_path):
try:
infer_model = vdevice.create_infer_model(detection_model_path)
infer_model.set_batch_size(1)
with infer_model.configure() as configured_infer_model:
while True:
frame = frame_queue.get()
image_data = np.expand_dims(frame, axis=0)
bindings = configured_infer_model.create_bindings()
bindings.input().set_buffer(image_data)
output_shape = infer_model.output().shape
output_buffer = np.empty(output_shape, dtype=np.float32)
bindings.output().set_buffer(output_buffer)
configured_infer_model.wait_for_async_ready(timeout_ms=10000)
job = configured_infer_model.run_async(
[bindings],
partial(detection_callback, bindings=bindings)
)
job.wait(timeout_ms=10000)
frame_queue.task_done()
except Exception as e:
print(f"Exception in detection_thread: {e}")
def detection_callback(completion_info, bindings):
if completion_info.exception:
print(f"Inference error: {completion_info.exception}")
else:
# Process detection results
# If a detection is found, put the frame into ocr_queue
ocr_queue.put(bindings.input().get_buffer()[0])
def ocr_thread(vdevice, ocr_model_path):
try:
infer_model = vdevice.create_infer_model(ocr_model_path)
infer_model.set_batch_size(1)
with infer_model.configure() as configured_infer_model:
while True:
frame = ocr_queue.get()
image_data = np.expand_dims(frame, axis=0)
bindings = configured_infer_model.create_bindings()
bindings.input().set_buffer(image_data)
output_shape = infer_model.output().shape
output_buffer = np.empty(output_shape, dtype=np.float32)
bindings.output().set_buffer(output_buffer)
configured_infer_model.wait_for_async_ready(timeout_ms=10000)
job = configured_infer_model.run_async(
[bindings],
partial(ocr_callback, bindings=bindings)
)
job.wait(timeout_ms=10000)
ocr_queue.task_done()
except Exception as e:
print(f"Exception in ocr_thread: {e}")
def ocr_callback(completion_info, bindings):
if completion_info.exception:
print(f"OCR Inference error: {completion_info.exception}")
else:
# Process OCR results
pass
def main():
vdevice_params = VDevice.create_params()
vdevice_params.scheduling_algorithm = HailoSchedulingAlgorithm.ROUND_ROBIN
vdevice_params.group_id = "SHARED"
with VDevice(vdevice_params) as vdevice:
# Start the frame grabbing thread
threading.Thread(target=grab_frames, daemon=True).start()
# Start the detection thread
threading.Thread(
target=detection_thread,
args=(vdevice, "./license_plate_model.hef"),
daemon=True
).start()
# Start the OCR thread
threading.Thread(
target=ocr_thread,
args=(vdevice, "./license_plate_ocr.hef"),
daemon=True
).start()
# Keep the main thread alive
threading.Event().wait()
if __name__ == '__main__':
main()
Notes:
- I’m using a single
VDevice instance shared across threads.
- The scheduling algorithm is set to
ROUND_ROBIN, and the group_id is set to "SHARED".
- The
grab_frames function simulates frame grabbing and puts frames into frame_queue.
- The
detection_thread processes frames from frame_queue and, upon detecting an object, puts frames into ocr_queue.
- The
ocr_thread processes frames from ocr_queue.
Issue:
- When the OCR thread starts processing after a detection, the script crashes with a segmentation fault (it still processes the OCR and all fine there but after processgin OCR in 30sec the error appears).
- If there are no detections, the script runs indefinitely without crashing.
- Adding threading locks or using multiprocessing did not resolve the issue.
Request:
Could someone please help me identify what might be causing this segmentation fault when the OCR inference runs? Is there a specific way I should manage the VDevice and models in a multithreaded environment with the Hailo SDK?
Any guidance or suggestions would be greatly appreciated!