Sample Model Config

model_config:
  mmbt:
    # Either pretraining or classification
    training_head_type: pretraining
    bert_model_name: bert-base-uncased
    direct_features_input: false
    freeze_text: false
    freeze_modal: false
    freeze_complete_base: false
    finetune_lr_multiplier: 1
    # Dimension of the embedding finally returned by the modal encoder
    modal_hidden_size: 2048
    # Dimension of the embedding finally returned by the text encoder
    text_hidden_size: 768
    # Used when classification head is activated
    num_labels: 2
    modal_encoder:
      type: resnet152
      params:
        pretrained: true
        pool_type: avg
        num_output_features: 1

    use_modal_start_token: true
    use_modal_end_token: true
    text_encoder:
      type: transformer
      params:
        bert_model_name: ${model_config.mmbt.bert_model_name}
        # Options below can be overridden to update the bert configuration used
        # to initialize the bert encoder. If some option is missing or
        # if you are using an encoder different then BERT, add extra parameters
        # to your projects configuration file under model_config.mmbt.
        # Those options will automatically override the options for your transformer
        # encoder's configuration. For e.g. vocab_size is missing here, just add
        # vocab_size: x to update the size of the vocabulary with which encoder is
        # initialized. If you update the default values, the transformer you
        # will get will be initialized from scratch.
        hidden_size: 768
        num_hidden_layers: 12
        num_attention_heads: 12
        output_attentions: false
        output_hidden_states: false