Multimodal Adaptation Gate (MAG)

Open source code for ACL 2020 Paper: Integrating Multimodal Information in Large Pretrained Transformers

If you use the model or results, please consider citing the research paper:

@inproceedings{rahman-etal-2020-integrating,
    title = "Integrating Multimodal Information in Large Pretrained Transformers",
    author = "Rahman, Wasifur  and
      Hasan, Md Kamrul  and
      Lee, Sangwu  and
      Bagher Zadeh, AmirAli  and
      Mao, Chengfeng  and
      Morency, Louis-Philippe  and
      Hoque, Ehsan",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.214",
    doi = "10.18653/v1/2020.acl-main.214",
    pages = "2359--2369",
    abstract = "",
}

Getting started

1. Configure global_configs.py

   global_configs.py defines global constants for runnning experiments. Dimensions of data modality (text, acoustic, visual), cpu/gpu settings, and MAG's injection position. Default configuration is set to MOSI. For running experiments on MOSEI or on custom dataset, make sure that ACOUSTIC_DIM and VISUAL_DIM are set approperiately.

   os.environ["CUDA_VISIBLE_DEVICES"] = "0"
   os.environ["WANDB_PROGRAM"] = "multimodal_driver.py"
   
   DEVICE = torch.device("cuda:0")
   
    # MOSI SETTING
    ACOUSTIC_DIM = 74
    VISUAL_DIM = 47
    TEXT_DIM = 768
   
    # MOSEI SETTING
    # ACOUSTIC_DIM = 74
    # VISUAL_DIM = 35
    # TEXT_DIM = 768
   
    # CUSTOM DATASET
    # ACOUSTIC_DIM = ??
    # VISUAL_DIM = ??
    # TEXT_DIM = ??
   
   XLNET_INJECTION_INDEX = 1

2. Download datasets Inside ./datasets folder, run ./download_datasets.sh to download MOSI and MOSEI datasets

3. Training MAG-BERT / MAG-XLNet on MOSI

   First, install python dependancies using pip install -r requirements.txt

   Training scripts:

   • MAG-BERT python multimodal_driver.py --model bert-base-uncased
   • MAG-XLNet python multimodal_driver.py --model xlnet-base-cased

   By default, multimodal_driver.py will attempt to create a Weights and Biases (W&B) project to log your runs and results. If you wish to disable W&B logging, set environment variable to WANDB_MODE=dryrun.

4. Model usage

   We would like to thank huggingface for providing and open-sourcing BERT / XLNet code for developing our models. Note that bert.py / xlnet.py are based on huggingface's implmentation.

   MAG

   from modeling import MAG
   
   hidden_size, beta_shift, dropout_prob = 768, 1e-3, 0.5
   multimodal_gate = MAG(hidden_size, beta_shift, dropout_prob)
   
   fused_embedding = multimodal_gate(text_embedding, visual_embedding, acoustic_embedding)

   MAG-BERT

   from bert import MAG_BertForSequenceClassification
   
   class MultimodalConfig(object):
       def __init__(self, beta_shift, dropout_prob):
           self.beta_shift = beta_shift
           self.dropout_prob = dropout_prob
   
   multimodal_config = MultimodalConfig(beta_shift=1e-3, dropout_prob=0.5)
   model = MAG_BertForSequenceClassification.from_pretrained(
           'bert-base-uncased', multimodal_config=multimodal_config, num_labels=1,
       )
   
   outputs = model(input_ids, visual, acoustic, attention_mask, position_ids)
   logits = outputs[0]

   MAG-XLNet

   from xlnet import MAG_XLNetForSequenceClassification
   
   class MultimodalConfig(object):
       def __init__(self, beta_shift, dropout_prob):
           self.beta_shift = beta_shift
           self.dropout_prob = dropout_prob
   
   multimodal_config = MultimodalConfig(beta_shift=1e-3, dropout_prob=0.5)
   model = MAG_XLNet_ForSequenceClassification.from_pretrained(
           'xlnet-base-cased', multimodal_config=multimodal_config, num_labels=1,
       )
   
   outputs = model(input_ids, visual, acoustic, attention_mask, position_ids)
   logits = outputs[0]

   For MAG-BERT / MAG-XLNet usage, visual, acoustic are torch.FloatTensor of shape (batch_size, sequence_length, modality_dim).

   input_ids, attention_mask, position_ids are torch.LongTensor of shape (batch_size, sequence_length). For more details on how these tensors should be formatted / generated, please refer to multimodal_driver.py's convert_to_features method and huggingface's documentation

Dataset Format

All datasets are saved under ./datasets/ folder and is encoded as .pkl file. Format of dataset is as follows:

{
    "train": [
        (words, visual, acoustic), label_id, segment,
        ...
    ],
    "dev": [ ... ],
    "test": [ ... ]
}

• words (List[str]): List of words
• visual (np.array): Numpy array of shape (sequence_len, VISUAL_DIM)
• acoustic (np.array): Numpy array of shape (seqeunce_len, ACOUSTIC_DIM)
• label_id (float): Label for data point
• segment (Any): Unique identifier for each data point

Dataset is encoded as python dictionary and saved as .pkl file

import pickle as pkl

# NOTE: Use 'wb' mode
with open('data.pkl', 'wb') as f:
    pkl.dump(data, f)

Contacts

• Wasifur Rahman: rahmanwasifur@gmail.com
• Sangwu Lee: sangwulee2@gmail.com
• Kamrul Hasan: mhasan8@cs.rochester.edu