Deep Feature Transmission Simulator (DFTS) version 2.
Originally released in 2018 [1], DFTS was developed to be compatible with Tensorflow version 1 (more specifically, version 1.12) and Keras 2.2.2. The demo paper [2] gave a brief overview of the simulator. Various changes in Tensorflow 2 [3] break the operation of DFTS.
We have modified DFTS to be now fully Tensorflow version 2-compatible in this repository. Previously we edited the original DFTS to run (with minimal modification) in Tensorflow 2 by disabling the v2 behavior in [4].
DFTS2 is a sophisticated simulation framework. It has new features:
- TensorFlow version 2 compatibility.
- Additional communication channel models and simulation modes.
- Missing feature recovery methods from the recent literature.
The following figure gives a system overview of Collaborative Intelligence strategies implemented in DFTS2.
Two peer reviewed conference papers were published on work done with DFTS2.
- A. Dhondea, R. A. Cohen, and I. V.Bajić, CALTeC: Content-adaptive linear tensor completion for collaborative intelligence, Proc. IEEE ICIP, 2021.
- A. Dhondea, R. A. Cohen, and I.V.Bajić, DFTS2: Deep feature transmission simulator for collaborative intelligence. For benchmarking purposes and to assist future users, we provide our packet traces, example simulation scripts and Monte Carlo experiment result files in a Dropbox directory. The full-scale test set used in our experiments is the same subset of the Imagenet validation set from the original DFTS demo paper [1]. Our extended paper, available on ArXiv, provides Monte Carlo results on the image classification task on ResNet-18, ResNet-34, DenseNet-121 and EfficientNet-B0.
In this YouTube video, we present our simulator, demonstrate how to set up a Python virtual environment for DFTS2, and show how to use DFTS2.
The latest version of the user documentation manual can be found [here]
[1] Unnibhavi, H. (2018) DFTS (Version 1.0) [repo]
[2] H. Unnibhavi, H. Choi, S. R. Alvar, and I. V. Bajić, "DFTS: Deep Feature Transmission Simulator," demo paper at IEEE MMSP'18, Vancouver, BC, Aug. 2018. [pdf]
[3] Effective TensorFlow 2 [guide]
[4] Dhondea, A. (2020) DFTS_compat_v1 (Version 1.0) [repo]
This project is licensed under the MIT License - see the LICENSE.md file for details.