axotzero/avsd

2D-MapFormer

Source Code for my master thesis "2D-MapFormer: 2D-Map Transformer for Audio-Visual Scene-Aware Dialogue and Reasoning" (Currently not published).

The Source Code is derived from

• AVSD-DSTC10 Baseline: Link
• 2D-Tan module: Link

Usage

1. Requirments
   • conda
   • wandb
2. Environments Setting

   . ./setup.sh
   

3. Download I3D and VGGish pretrained features

   . ./download_data.sh
   python3 utils/combine_files.py # combine feature files into ./data/features/train.pkl and ./data/features/test.pkl
   

4. Train model
   1. Specify the exp_name in the run.sh. The trained model and model outputs will stored in ./log/{exp_name}/. It will also be the experiment name of wandb
   2. Specify the procedure='train_test'
   3. Specify other hyperparameters. Please see run.sh and main.py for more details.
   4. run . ./run.sh.
      1. It will run training and testing automatically
      2. You will see the following procedure in the command line

         train 15, tan:0.125, dig:2.272: 100%|█████| 4787/4787 [21:15<00:00,  3.75it/s]
         train 15, tan:0.112, dig:2.153
         val   15, tan:0.087, dig:1.985: 100%|█████| 1117/1117 [06:12<00:00,  3.00it/s]
         val   15, tan:0.109, dig:2.295
         The best metric was  for 0 epochs.
         Expected early stop @ 19
         train 16, tan:0.094, dig:2.097: 100%|█████| 4787/4787 [21:10<00:00,  3.77it/s]
         train 16, tan:0.112, dig:2.136
         val   16, tan:0.088, dig:2.005: 100%|█████| 1117/1117 [06:11<00:00,  3.01it/s]
         val   16, tan:0.109, dig:2.298
         

      3. You will see the following test result in the command line

         DSTC10_beam_search result:
         | Bleu_1: 68.7000
         | Bleu_2: 55.5832
         | Bleu_3: 45.4938
         | Bleu_4: 37.5887
         | METEOR: 24.3038
         | ROUGE_L: 53.4955
         | CIDEr: 86.9928
         | IoU-1: 54.7007
         | IoU-2: 57.6148
         

Model Architecture

Model Overview

Audio Visual Encoder

Sentence Cross Attention

Update Gate