YOLOv8 to TFLite conversion results

Community Help
1

Hello,

I have trained YOLOv8m on a custom dataset with 5 classes obtaining quite good results. Afterwards I have tried to convert this model to TFLite. I have converted it and created my detection script. I have implemented the preprocessing in the following manner:

def preprocess(img):
    # Letterbox
    img = letterbox(img, (640, 640))

    # BGR to RGB
    img = img[:, :, ::-1]

    img = img / 255.

    # Expand dims for batch
    img = np.expand_dims(img, 0)

    return img

def letterbox(img, target_size):
    ih, iw = img.shape[:2]
    w, h = target_size

    scale = min(w / iw, h / ih)
    nw = int(iw * scale)
    nh = int(ih * scale)

    image_resized = cv2.resize(img, (nw, nh))

    image_padded = np.full(shape=[h, w, 3], fill_value=128.0)
    dw = int((w - nw) / 2)
    dh = int((h - nh) / 2)
    image_padded[dh:nh + dh, dw:nw + dw, :] = image_resized

    return image_padded

The main function loads the model, reads an image, process the image, feeds it to the model, and runs the inference:

interpreter = tf.lite.Interpreter(model_path="yolov8_last_float32.tflite")
    interpreter.allocate_tensors()

    input_details = interpreter.get_input_details()

    # Load image
    img = cv2.imread("WIN_20230719_13_10_23_Pro.jpg")

    # Preprocess
    img_input = preprocess(img)

    # Set input tensor
    img_input = img_input.astype(np.float32)
    
    interpreter.set_tensor(input_details[0]['index'], img_input)

    # Run inference
    interpreter.invoke()

    output_details = interpreter.get_output_details()
    # Get predictions
    output = interpreter.get_tensor(output_details[0]['index'])

    output = output[0].transpose()

    boxes = output[:, :4]
    scores = output[:, 4:]

    format_boxes = np.zeros(shape=boxes.shape)

    format_boxes[:, 0] = boxes[:, 0] - (boxes[:, 2] / 2)
    format_boxes[:, 1] = boxes[:, 1] - (boxes[:, 3] / 2)
    format_boxes[:, 2] = boxes[:, 0] + (boxes[:, 2] / 2)
    format_boxes[:, 3] = boxes[:, 1] + (boxes[:, 3] / 2)

    format_boxes = format_boxes.astype(np.int32)

    nms_boxes, scores = non_max_suppression(format_boxes, scores, 0.5)

I also process the output in order to transform it from the bounding box center point, width and height format into bounding box top left and bottom right point format. I also apply non-maximum suppression.

The problem is the score values are very low:

image
image830×224 24.1 KB

These are the values for each prediction after applying NMS.

Any ideas what am I doing wrong?

2

Hi @Theodor_Nanu

Sorry to hear of your trouble with TFLite. What did you use to convert to your .tflite file?

Did you use our notebook or was this something you did yourself?

3

I used the export function provided in the ultralytics package.

4

Hi @Theodor_Nanu

Generally, we don’t support other products here on Roboflow. That’s something you should reach out to Ultralytics for.

If you’d like to use Roboflow Train or one of our many notebooks, we’d be more than happy to support you with that.