GPU memory not get released after using CNN for face location in web_service_example.py (flask example)

• face_recognition version: 1.2.3
• Python version: 3.5.2
• Operating System: Ubuntu 16.04 LTS
• GPU: NVIDIA GeForce GTX 1050TI SSE2

Description

GPU memory not released after adding "CNN for face location" line in web_service_example.py

What I Did

By adding **line of finding face location using CNN** in web_service_example.py like this ::
    **# Load the uploaded image file
    img = face_recognition.load_image_file(file_stream)

    _# Get face locations for any face in the uploaded image
    unknown_face_location = face_recognition.face_locations(img, number_of_times_to_upsample=1, model="cnn")_

    # Get face encodings for any faces in the uploaded image
    unknown_face_encodings = face_recognition.face_encodings(img, unknown_face_location)**

i found that my GPU does not release memory after CNN,
So when i try to upload image 1st time, it uses 77% of GPU's memory and when i upload image 2nd time it starts using GPU' memory from 77% ( that means GPU is not releasing its memory after complition of request),
after multiple attempts GPU's memory reach to 100% and then it shows ::
  **return cnn_face_detector(img, number_of_times_to_upsample)
RuntimeError: Error while calling cudaMalloc(&data, n) in file /home/dlib-19.13/dlib/cuda/cuda_data_ptr.cpp:28. code: 2, reason: out of memory**


i tried ::
**del unknown_face_location
gc_collect()**
below unknown_face_location
but it didn't work, 
when i restart flask only then GPU releases memory (which is not good option to use)

i used following version::
Nvidia Driver : nvidia-396
CUDA Version 9.0.176
Cudnn Version 7.4.2
Dlib Version 19.13

I need to know how to overcome from this memory issue by using another drivers and version or by changing the code.
NOTE: i need CNN and when i use CNN in a code where i don't use Flask and other web service applications it works fine.(I am new to this So spare me for any small mistakes done by me)

Start with web_service_example.py and reproduce repeated Flask uploads using face_recognition.face_locations(..., model="cnn"). Inspect the request lifecycle and the CNN detector's GPU allocation behavior with the listed Python, dlib, CUDA, and cuDNN versions. Done means repeated requests no longer grow GPU usage until cudaMalloc fails.