2 Stage prediction with pre-trained models

Hi All,

I am trying to do a 2 stage prediction with pretrained models face-detection-retail-0004 and saved_model_openvino_2021.4_6shave for face and mask detection. Where only if a face is detected them submit the whole frame for mask detection. I get each individual model to predict properly when run individually in separate scripts, but when I try to mat it a 2 stage prediction in one script, the face detection works, but mask detection never happens. What am I doing wrong in the script below? Any Thoughts?

`import sys
sys.path.insert(0,'../..')
import cv2 # opencv - display the videos stream
import depthai # depthai - access the camera and its data packets
from uielements import DisplayValueLabel
import pyttsx3
import threading
import queue

#color for label text and confidence in BGR order

text_color = (255, 0, 0)

#Prevents multiple speech to talk
tts_queue = queue.Queue()

#Define type to speech

def tts_worker():
engine = pyttsx3.init()
while True:
message = tts_queue.get()
if message is None:
break
engine.say(message)
engine.runAndWait()
tts_queue.task_done()

#Start TTS worker thread

tts_thread = threading.Thread(target=tts_worker)
tts_thread.daemon = True
tts_thread.start()

#Main
def main():
pipeline = depthai.Pipeline()

# Create ColorCamera Node
cam_rgb = pipeline.create(depthai.node.ColorCamera)
cam_rgb.setPreviewSize(400, 400)
cam_rgb.setInterleaved(False)

# ImageManip that will crop the frame before sending it to the Face detection NN node
manip_mask = pipeline.create(depthai.node.ImageManip)
manip_mask.initialConfig.setResize(224, 224)

# ImageManip that will crop the frame before sending it to the Face detection NN node
manip_face = pipeline.create(depthai.node.ImageManip)
manip_face.initialConfig.setResize(300, 300)



# connect the 600x600 preview output to the input of the ImageManip node
cam_rgb.preview.link(manip_face.inputImage)

#Create face dectection NuralNetwork 
face_nn = pipeline.create(depthai.node.NeuralNetwork)
face_nn.setBlobPath("./face-detection-retail-0004.blob")
manip_face.out.link(face_nn.input)

# Create NeuralNetwork Node that we will use to load our saved customer model
custom_nn = pipeline.create(depthai.node.NeuralNetwork)
custom_nn.setBlobPath("./saved_model_openvino_2021.4_6shave.blob")
manip_mask.out.link(custom_nn.input)

# Create XLinkOut nodes to send data to host for the camera
xout_rgb = pipeline.create(depthai.node.XLinkOut)
xout_rgb.setStreamName("rgb")
cam_rgb.preview.link(xout_rgb.input)

# Create XLinkOut nodes to send data to host for the neural network-ORIGINAL
xout_face_nn = pipeline.create(depthai.node.XLinkOut)
xout_face_nn.setStreamName("face_nn")
face_nn.out.link(xout_face_nn.input)

# Create XLinkOut nodes to send data to host for the neural network-ORIGINAL
xout_custom_nn = pipeline.create(depthai.node.XLinkOut)
xout_custom_nn.setStreamName("custom_nn")
custom_nn.out.link(xout_custom_nn.input)


display_label = DisplayValueLabel(10, 10, 275, 40, 'Mask?: ')


# get the virtual device, and loop forever reading messages
# from the internal queue
with depthai.Device(pipeline) as device:
    # get a reference to the rgb queue, which contains the 600x600 frames from OAK camera
    # and a reference to the NeuralNetwork output with the model outputs of the mask
    # predictor
    q_rgb = device.getOutputQueue("rgb")
    q_fnn = device.getOutputQueue("face_nn")
    q_nn = device.getOutputQueue("custom_nn")
    det_fac = ""
    det_msk =""

    frame = None
    while True:
        # read a message from each of the queues
        in_rgb = q_rgb.tryGet()
        in_face_nn = q_fnn.tryGet()
        in_custom_nn = q_nn.tryGet()

        if in_rgb is not None:
            # then we have a frame from the OAK
            frame = in_rgb.getCvFrame()

        if in_face_nn is not None:
            detections = in_face_nn.getFirstLayerFp16()
            print ("Face Detection output:", detections) 
            for i in range(0, len(detections), 7):
                confidence = detections[i + 2]
                if confidence > 0.6:
                    print(f"Face detected with confidence {confidence}")
                    det_fac = "YES"

        # Check for mask only if face detected
        if in_custom_nn is not None:
            print("CUSTOM_NOT_NONE")
            mask, no_mask = in_custom_nn.getLayerFp16('StatefulPartitionedCall/model/dense_1/Softmax')
            
            # print the results of the prediction
            print(f"Mask[{round(mask,1)}], No Mask[{round(no_mask,1)}]")
            if round(mask,1) > round(no_mask,1):
                display_label.set_value("MASK")
            else:
                display_label.set_value("NO MASK")
                det_msk = "NO"

        if det_fac == "YES" and det_msk == "NO":
            print("HIT")
            if tts_queue.empty():
                tts_queue.put("Please wear a mask")


        if frame is not None:
            display_label.draw(frame)
            # Show the frame from the OAK device
            cv2.imshow("TF Face Mask", frame)

        if cv2.waitKey(1) == ord('q'):
            break


#SIgnal the TTS worker thread to exit
tts_queue.put(None)
tts_thread.join()

ifname== "main":
main()`