Kmbin93/split_computing

Split Computing Testbed

Demo

Requirements

• OpenCV
• Tensorflow >=2.0
• pynvjpeg

Server

usage: server.py [-h] [--split_layer_name SPLIT_LAYER_NAME] [--model_path MODEL_PATH] [--decoder DECODER]

optional arguments:
  -h, --help            show this help message and exit
  --split_layer_name SPLIT_LAYER_NAME
                        Target splitting layer name
  --model_path MODEL_PATH
                        DNN model path
  --decoder DECODER     Type of encoder (JPEG or AE). JPEG - Use JPEG for encoding input frame. Note that JPEG does not work for intermediate output tensor of DNN. AE - Use AutoEncoder for encoding       
                        input frame or intermediate output tensor of DNN.

Client

usage: client.py [-h] --server_ip SERVER_IP [--port PORT] --file_name FILE_NAME [--model_path MODEL_PATH] [--split_layer_name SPLIT_LAYER_NAME] [--resize_factor RESIZE_FACTOR] [--encoder ENCODER]
                 [--jpeg_qp JPEG_QP]

optional arguments:
  -h, --help            show this help message and exit
  --server_ip SERVER_IP
                        Server IP
  --port PORT           Port number
  --file_name FILE_NAME
                        Input video source file name.
  --model_path MODEL_PATH
                        DNN model path
  --split_layer_name SPLIT_LAYER_NAME
                        Target layer name for splitting the model.
  --resize_factor RESIZE_FACTOR
                        Resize rate [0, 1]. Resizes input frame with the resize rate. e.g., For rf 0.5: 3840 x 2160 (4K) -> 1920 x 1080 (FHD)
  --encoder ENCODER     Type of encoder (JPEG or AE). JPEG - Use JPEG for encoding input frame. Note that JPEG does not work for intermediate output tensor of DNN. AE - Use AutoEncoder for encoding       
                        input frame or intermediate output tensor of DNN.
  --jpeg_qp JPEG_QP     JPEG quality factor. Default=90

Example

Download ResNet50 model and AutoEncoder for running example

Download Example 4K Video here

• Server

 python server.py --model_path resnet50_classification.h5 --split_layer_name=conv3_block1_out --decoder=AE

• Client

 python client.py --server_ip=147.46.130.213 --file_name=video_4k.mp4 --model_path=resnet50_classification.h5 --split_layer_name=conv3_block1_out --encoder=AE --resize_factor=0.5