IA-for-AI : Intelligence Augmentation for Artificial Intelligence

Emotions form an integral part of human interactions. The Intelligence Augmentation for AI Hackathon 2021 paves the way toward more empathy AI systems by aiming to build systems to recognize emotions from audio. The best entry into the competition from our team - Prompt Engineers - is a system that leverages not only the audio features but also the semantics of the spoken words, fusing the two intertwined modalities to achieve a runner-up position on the leaderboard with 61.38% Accuracy. We further improve the latency of the approach by more than 42% via feature reuse, weight sharing and multi-task learning at the cost of only 0.2% Accuracy drop.

Our best performing model is a phono-linguistic model, leveraging both the semantics of the spoken works and the speech features. We obtain speech features from Hubert - a speech pretrained transformer model and language features from Bert - a language model running over the output of the transcribed speech. The features from the two modalities are fused together to achieve 61.38% Accuracy. We observe that Bert features over the transcribed speech alone achieves 55.77% Accuracy. Whereas classifying on only the speech features from Hubert yields 58.98% Accuracy. Together the two modalities achieves the best performance.

We improve latency by multi-task learning the HuBert for Audio features as well as speech transcribing (ASR). This leads to 42% less model parameters with only 0.2% performance drop.

Dependencies:

We exported our conda environments for training the models and running the app.

• train_env.yml: Our environment for training. Create using conda env create --name prompt --file=train_env.yml & conda activate prompt
• app_env.yml: Our environment for app. Create using conda env create --name prompt_app --file=app_env.yml & conda activate prompt_app

Please note that our app_env was ran on a MacOS 11.2 Machine with Intel processor, whereas our training (train_env) was done on a linux machine with Nvidia GPUs. The same conda environment may not work on other machines. Instead you may download the packages individually.

You may also download the following dependencies individually as an alternate means to create the environment:

• Pytorch
• Huggingface's Transformers
• Huggingface's Datasets
• torchaudio
• soundfile
• sounddevice
• scikit-learn
• scipy
• numpy

Additional dependencies for running the webapp: streamlit, plotly

• Create a fresh conda environment
• pip install streamlit
• pip install soundfile
• pip install sounddevice
• pip install pydub
• Install PyTorch-1.8 and torchaudio
• pip install transformers==4.10

If you are training the AST model, then also download the following dependencies: matplotlib, numba, timm, zipp, wget, llvmlite

Running the webapp

• Download the required dependencies or replicate & activate our conda environment, as detailed above.
• Our webapp is in the app folder: cd app.
• Download the pre-trained model weights and save it in the webapp folder with the name cpu_model.pt
• Run webapp: streamlit run app.py
• Note that it may take some time to run first time as models, tokenizers, feature-extractors and config are downloaded.
• The webapp will be hosted locally (port and address printed on command line).

Training model

• Download the required dependencies or replicate our conda environment, as detailed above.
• Download the set of audio files in TrainAudioFiles & TestAudioFiles and place it inside the dataset folder.
• Get pseudo ASR ground truth labels by running python3 run_asr.py from inside the linguistic folder.
• If you want to train the linguistic (text-only) model, then you can run python3 train.py from inside linguistic folder with the following optional command line arguments .
  • Batch Size: --batch_size=16
  • Learning Rate: --lr=1e-5
  • Number of Epochs: --n_epochs=10
  • Do dummy run for debugging purposes: --dummy_run
  • Device to training the model on: --device=cuda
  • Randomizing seed: --seed=1
  • Test the model: --test_model
  • Bert Model name or path: --bert_type=bert-base-uncased
• For training AST: Refer to the instructions in ast/README.md, the dataset needs to downloaded and kept inside ast/egs/hack/data/.
• If you want to train a model only on audio features, then from inside phono folder, run bash run.sh. You may edit the run.sh to change the following arguments:
  • Pooling method to extract model features: --pooling_mode Options: ["mean",'max', 'sum']
  • Name or path of audio only model's pretrained file: --model_name_or_path
  • Type of model to use: --model_mode Example arguments: hubert or wav2vec2
  • Training Batch size per device: --per_device_train_batch_size [Type: Integer]
  • Eval Batch Size per device: --per_device_eval_batch_size [Type: Integer]
  • Learning rate: --learning_rate
  • Number of epochs: --num_train_epochs
  • Gradient Accumulation steps: --gradient_accumulation_steps (set 1 for no accumulation)
• Save, eval and logger steps: --save_steps, eval_steps, logging_steps
  • Maximum number of models to save: --save_total_limit
  • Required Arguments (do not change): --freeze_feature_extractor, --input_column=filename, --target_column=emotion, output_dir="output_dir" delimiter="comma", --evaluation_strategy="steps", --fp16, --train_file="./../dataset/train_set.csv", --validation_file="./../dataset/valid_set.csv", --test_file="./../dataset/test_set.csv"
  • Whether to do train, eval and predict on test: --do_eval, --do_train, --do_predict
• If you want to train phono_linguistic transformer model:
  • Extract phonetic features for train and test set from phono_feat_extractor:
    • After obtaining ASR predictions from linguistic, put the test.json and train.json in the phono_feat_extractor folder.
    • Merge the above two files using python3 merge_text.py inside the phono_feat_extractor folder.
    • Then do bash run.sh from inside the same folder. You may change the same arguments as above mentioned for only on audio features. Additional argument of the Bert model: --bert_name='bert-base-uncased'.
  • After the above step, you will obtain the train.pkl and test.pkl files inside phono_feat_extractor. Put these files in phono_linguistic/data folder.
  • Run python3 bertloader.py from phono-linguistic folder to cache the dataloader for training.
  • And then train model using python3 train.py.
  • You may include the following arguments for python3 bertloader.py and python3 train.py. Make sure the same arguments are passed to the two commands
    • Seed: --seed=1 (type=int)
    • Batch_Size: --batch_size=16 (type=int)
    • Learning rate: --lr=1e-5 (type=float)
    • Number epochs: --n_epochs=5 (type=int)
    • Dummy run (for debugging purposes): --dummy_run
    • Device to train model on: --device
    • Whether to log on wanbd: --wandb
    • Bert model name or path: --bert_type=bert-base-uncased
    • Audio model name or path: --model_name_or_path='facebook/hubert-large-ll60k'

Demo Video

• Instruction to use webapp

Miscellanous

• You may contact us by opening an issue on this repo. Please allow 2-3 days of time to address the issue.
• License: MIT

Made with ❤️