This is one of the model that we proposed for Advanced Projects at the Quality and Usability Lab at TU Berlin. We refer to this paper and follow this approach from Github/DMD.
Data files consist of MOSI and MOSEI datasets can be found here.
The datasets are placed into the folder ./dataset
.
By default, the trained model will be saved in folder ./pt
. Our trained model can be download from pt folder
If there is no difference in validation loss during, the last model contains layers and weights will be additionally saved in ./dmd.pth
.
Before testing the model we set the path of trained model in run.py (line 174). The results are saved into folder ./result
We run the codes on google colab, implemented under the .ipynb file Unimse_Submission.ipynb
.
Acc_2: 0.8430 F1_score: 0.8437 Acc_7: 0.4548 MAE: 0.7334 Loss: 0.7334
- Create IEMOCAP or EMODB datasets and train these datasets
- Fine tune the model