config.json

{
  "model": {

    "name" : "VQA historical baseline (small)  with classic attention",

    "image":
    {
      "image_input": "raw",
      "dim": [224, 224, 3],
      "normalize": true,

      "resnet_version": 50,
      "finetune" : [],

      "attention" : {
        "mode": "classic",
        "no_attention_mlp": 256
      },

      "cbn": {
        "use_cbn": true,
        "cbn_embedding_size": 512,
        "excluded_scope_names": ["*"]
      }

    },

    "glove": true,

    "word_embedding_dim": 300,
    "no_hidden_LSTM": 1024,
    "no_LSTM_cell": 1,

    "no_question_mlp": 1024,
    "no_image_mlp": 1024,

    "no_hidden_final_mlp": 1024,
    "dropout_keep_prob": 0.5,

    "loss": "soft",
    "activation": "tanh"
  },

  "optimizer": {
    "max_iters": 100000,
    "learning_rate": 3e-4,
    "batch_size": 20,
    "step_size": 20000,
    "clip_val": 5
  },

  "dico_name" : "dict.json",
  "glove_name" : "glove_dict.pkl",
  "checkpoint_path" : "checkpoint10.pth",

  "merge_dataset": false,
  "seed": -1
}