def get_file_content_as_string(path):
    url = 'https://raw.githubusercontent.com/streamlit/demo-self-driving/master/' + path
    response = urllib.request.urlopen(url)
    return response.read().decode("utf-8")

# This function loads an image from Streamlit public repo on S3. We use st.cache on this
# function as well, so we can reuse the images across runs. 

def get_batch(device):
    # gets a random batch using cached load
    @st.cache
    def load_batch():
        return pickle.load(open(DATAPATH + 'batch.pkl', 'rb'))
    # todo remove randomness
    return [i.to(device) for i in random.choice(load_batch())] 

def run_the_app():
    # To make Streamlit fast, st.cache allows us to reuse computation across runs.
    # In this common pattern, we download data from an endpoint only once.
    @st.cache
    def load_metadata(url):
        return pd.read_csv(url)

    # This function uses some Pandas magic to summarize the metadata Dataframe.
    @st.cache
    def create_summary(metadata):
        one_hot_encoded = pd.get_dummies(metadata[["frame", "label"]], columns=["label"])
        summary = one_hot_encoded.groupby(["frame"]).sum().rename(columns={
            "label_biker": "biker",
            "label_car": "car",
            "label_pedestrian": "pedestrian",
            "label_trafficLight": "traffic light",
            "label_truck": "truck"
        })
        return summary

    # An amazing property of st.cached functions is that you can pipe them into
    # one another to form a computation DAG (directed acyclic graph). Streamlit
    # recomputes only whatever subset is required to get the right answer!
    metadata = load_metadata(os.path.join(DATA_URL_ROOT, "labels.csv.gz"))
    summary = create_summary(metadata)

    # Uncomment these lines to peek at these DataFrames.
    # st.write('## Metadata', metadata[:1000], '## Summary', summary[:1000])

    # Draw the UI elements to search for objects (pedestrians, cars, etc.)
    selected_frame_index, selected_frame = frame_selector_ui(summary)
    if selected_frame_index == None:
        st.error("No frames fit the criteria. Please select different label or number.")
        return

    # Draw the UI element to select parameters for the YOLO object detector.
    confidence_threshold, overlap_threshold = object_detector_ui()

    # Load the image from S3.
    image_url = os.path.join(DATA_URL_ROOT, selected_frame)
    image = load_image(image_url)

    # Add boxes for objects on the image. These are the boxes for the ground image.
    boxes = metadata[metadata.frame == selected_frame].drop(columns=["frame"])
    draw_image_with_boxes(image, boxes, "Ground Truth",
        "**Human-annotated data** (frame `%i`)" % selected_frame_index)

    # Get the boxes for the objects detected by YOLO by running the YOLO model.
    yolo_boxes = yolo_v3(image, confidence_threshold, overlap_threshold)
    draw_image_with_boxes(image, yolo_boxes, "Real-time Computer Vision",
        "**YOLO v3 Model** (overlap `%3.1f`) (confidence `%3.1f`)" % (overlap_threshold, confidence_threshold))

# This sidebar UI is a little search engine to find certain object types.