<<<<<<< HEAD <<<<<<< HEAD

DDS

This is the implementation for the paper Server-Driven Video Streaming for Deep Learning Inference.

1. Related resources

Please check Kuntai Du's home page for more DDS-related resources.

2. Install Instructions

To run our code, please make sure that conda is installed. Then, under dds repo, run

conda env create -f conda_environment_configuration.yml

to install dds environment. Note that this installation assumes that you have GPU resources on your machine. If not, please edit tensorflow-gpu=1.14 to tensorflow=1.14 in conda_environment_configuration.yml.

Now run

conda activate dds

to activate dds environment, and

cd workspace

and run

wget people.cs.uchicago.edu/~kuntai/frozen_inference_graph.pb

to download the object detection model (FasterRCNN-ResNet101).

3. Run our code

Under DDSrepo/workspace, run

python entrance.py

to run DDS!

4. Get performance numbers

Under DDSrepo/workspace, run

python examine.py trafficcam_1 results stats

you should see something like

trafficcam_1_dds_0.8_0.8_36_26_0.0_twosides_batch_15_0.5_0.3_0.01 3474KB 0.901
trafficcam_1_mpeg_0.8_26 8652KB 0.904
trafficcam_1_mpeg_0.8_36 2369KB 0.876

The number might vary by platform.

5. Some details

If you are considering building your projects based on our codebase, here are some details.

5.1 Run in implementation mode

Implementation means we run DDS under real network environment through HTTP post. To do so, in DDSrepo, run

FLASK_APP=backend/backend.py flask run --port=5001

and copy the frozen_inference_graph.pb to DDSrepo to help the server find the model.

Then use another terminal, cd to DDSrepo/workspace, and edit the mode to implementation and edit the hname to ip:5001 (ip should be 127.0.0.1 if you run the server locally) to run DDS on implementation mode. You can also run other methods in implementation mode by changing the default value of mode to implementation.

5.2 Inside workspace folder

Inside workspace folder, we use a configuration file configuration.yml to control the configuration for both the client and the server. This file will be only loaded once inside the whole python entrance.py process. You can add new keys and values in this file. We even support caching, parameter sweeping, and some fancy functionalities in this file. Please read the comments inside this file to utilize it.

6. Dataset

6.1 Detection dataset

6.1.1 Traffic camera videos and dash camera videos.

We search some keywords through youtube in the anonymous mode of Chrome. The top-ranked search results, corresponding URLS are listed below. We filter out some of these videos.

After downloading these videos, we get 7 traffic camera videos and 9 dash camera videos.

6.1.2 Drone videos

We obtain drone videos through a public dataset called VisDrone. We only use 13 videos in this dataset (there are too many videos inside...).

6.1.3 Face videos

We use clips from TBBT and Friends.

streamDuet

29aba6a65ab73bc439831a0b6a63c522e1d2bddc ======= <<<<<<< HEAD

DDS

This is the implementation for the paper Server-Driven Video Streaming for Deep Learning Inference.

1. Related resources

Please check Kuntai Du's home page for more DDS-related resources.

2. Install Instructions

To run our code, please make sure that conda is installed. Then, under dds repo, run

conda env create -f conda_environment_configuration.yml

to install dds environment. Note that this installation assumes that you have GPU resources on your machine. If not, please edit tensorflow-gpu=1.14 to tensorflow=1.14 in conda_environment_configuration.yml.

Now run

conda activate dds

to activate dds environment, and

cd workspace

and run

wget people.cs.uchicago.edu/~kuntai/frozen_inference_graph.pb

to download the object detection model (FasterRCNN-ResNet101).

3. Run our code

Under DDSrepo/workspace, run

python entrance.py

to run DDS!

4. Get performance numbers

Under DDSrepo/workspace, run

python examine.py trafficcam_1 results stats

you should see something like

trafficcam_1_dds_0.8_0.8_36_26_0.0_twosides_batch_15_0.5_0.3_0.01 3474KB 0.901
trafficcam_1_mpeg_0.8_26 8652KB 0.904
trafficcam_1_mpeg_0.8_36 2369KB 0.876

The number might vary by platform.

5. Some details

If you are considering building your projects based on our codebase, here are some details.

5.1 Run in implementation mode

Implementation means we run DDS under real network environment through HTTP post. To do so, in DDSrepo, run

FLASK_APP=backend/backend.py flask run --port=5001

and copy the frozen_inference_graph.pb to DDSrepo to help the server find the model.

Then use another terminal, cd to DDSrepo/workspace, and edit the mode to implementation and edit the hname to ip:5001 (ip should be 127.0.0.1 if you run the server locally) to run DDS on implementation mode. You can also run other methods in implementation mode by changing the default value of mode to implementation.

5.2 Inside workspace folder

Inside workspace folder, we use a configuration file configuration.yml to control the configuration for both the client and the server. This file will be only loaded once inside the whole python entrance.py process. You can add new keys and values in this file. We even support caching, parameter sweeping, and some fancy functionalities in this file. Please read the comments inside this file to utilize it.

6. Dataset

6.1 Detection dataset

6.1.1 Traffic camera videos and dash camera videos.

We search some keywords through youtube in the anonymous mode of Chrome. The top-ranked search results, corresponding URLS are listed below. We filter out some of these videos.

After downloading these videos, we get 7 traffic camera videos and 9 dash camera videos.

6.1.2 Drone videos

We obtain drone videos through a public dataset called VisDrone. We only use 13 videos in this dataset (there are too many videos inside...).

6.1.3 Face videos

We use clips from TBBT and Friends.

streamDuet

29aba6a65ab73bc439831a0b6a63c522e1d2bddc dfa28daa30fd64fb2992255637bf717c4a1b05c2