nwojke / deep_sort

Simple Online Realtime Tracking with a Deep Association Metric
GNU General Public License v3.0
5.19k stars 1.46k forks source link

TensorFlow 2.x with deep_sort #240

Open fdominik opened 3 years ago

fdominik commented 3 years ago

In case somebody needs to use TF 2.x and cant get deep_sort working, we migrated one script to be working with TF 2.x (there are still other parts of code we havent changed yet). https://github.com/mk-michal/deep_sort

The old code is using tf.Session() and Graph data, so we needed to change the structure. However as PR are not merged in this repo, we dont intend to do a PR until somebody will start maintaing this repo.

Olkaaaaaa12 commented 2 years ago

Hello, I also needed to use tf2 and I used your repo but you didn't update freeze_model.py for tf2. Without it I'm not able to run generate_detection.py with tf2. I used tf_upgrade_v2 to convert freeze_model.py to tf2. After that and and few changes in generate_detection.py (like delate 'net/' from 97 and 99 line) I was able to run your repo and it works ok but it takes a lot of memory on my gpu (about 24GB). With tf1 model it was 160MB. Can you tell me how you converted model from original repo to run it with your 'generate_detection.py' script? Thanks