I merge the pytorch implementation version of Tensor fusion network TFN and CMU-MOSI dataset preprocessing CMU-MultimodalSDK-Tutorials, such that tensor fusion network can run on the latest version of CMU-MOSI dataset.
The following part is adapted from the README.txt in CMU-MultimodalSDK-Tutorials):
This is a short tutorial for using the CMU-MultimodalSDK.
First of all you need to download the SDK from here.
After that you will need to clone this repo and set up some file paths in ./constants/paths.py. The comments there should be self-explanatory.
If you have downloaded pretrained embeddings such as GloVe on your hard drive, you can setup the path to them in ./constants/paths.py for improved performance.
Note that in the ./data/ directory there is a ./CMU_MOSI_ModifiedTimestampedWords.csd. This is an updated copy of transcripts compared to what the SDK is currently offering by default and improves the performance by a observable margin. If you are changing the DATA_PATH in ./constants/paths.py, please also copy this file to your specified location.
We modify the codings in ./tutorial_interactive.ipynb in CMU-MultimodalSDK. and replace it by pipeline.py.