Configs #

Config parameters configures training process and model architectures, as well as routine i/o setups.

Full Parameters #

{
  "$schema": "../schemas/model_config_schema.json",
  "general": {...},
  "training_process": {...},
  "action_encoder": {...},
  "card_encoder": {...},
  "cross_attention": {...},
  "output_mlp": {...}
}

{
  "$schema": "../schemas/model_config_schema.json",
  "general": {
    "device": "cpu",
    "seed": 42069,
    "load_checkpoint_name": "default",
    "save_checkpoint_name": "default",
    "type_of_checkpoint": "final"
  },
  "training_process": {
    "batch_size": 32,
    "num_epochs": 300,
    "learning_rate": 0.0005,
    "weight_decay": 0.01,
    "optimizer": "AdamW",
    "momentum": 0,
    "dataset": "GTO",
    "warm_start": false,
    "p_train_test_split": 0.2,
    "min_epochs": 100,
    "patience": 50
  },
  "action_encoder": {
   "d_model": 32,
    "nhead": 4,
    "dim_feedforward": 128,
    "dropout": 0.1,
    "num_actions": 5,
    "num_positions": 6,
    "num_streets": 5,
    "action_embedding_dim": 4,
    "position_embedding_dim": 4,
    "street_embedding_dim": 4
  },
  "card_encoder": {
    "d_model": 32,
    "nhead": 4,
    "dim_feedforward": 128,
    "dropout": 0.1,
    "num_ranks": 13,
    "num_suits": 4,
    "num_streets": 4,
    "rank_embedding_dim": 8,
    "suit_embedding_dim": 4,
    "street_embedding_dim": 4
  },
  "cross_attention": {
    "num_heads": 4,
    "d_model": 32
  },
  "output_mlp": {
    "d_model": 32
  }
}

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "title": "ModelConfig",
  "type": "object",
  "properties": {
    "general": {
      "type": "object",
      "properties": {
        "device": { "type": "string", "enum": ["cpu", "cuda"] },
        "seed": { "type": "integer" },
        "load_checkpoint_name": { "type": "string" },
        "save_checkpoint_name": { "type": "string" },
        "type_of_checkpoint": { "type": "string", "enum": ["best", "final"] }
      }
    },
    "training_process": {
      "type": "object",
      "properties": {
        "batch_size": { "type": "integer" },
        "num_epochs": { "type": "integer" },
        "learning_rate": { "type": "number" },
        "weight_decay": { "type": "number" },
        "optimizer": { "type": "string", "enum": ["Adam", "SGD", "AdamW"] },
        "momentum": { "type": "number" },
        "dataset": { "type": "string", "enum": ["GTO", "Human"] },
        "warm_start": { "type": "boolean" },
        "p_train_test_split": { "type": "number" },
        "min_epochs": { "type": "integer" },
        "patience": { "type": "integer" }
      }
    },
    "action_encoder": {
      "type": "object",
      "properties": {
        "d_model": { "type": "integer" },
        "nhead": { "type": "integer" },
        "dim_feedforward": { "type": "integer" },
        "dropout": { "type": "number" },
        "num_actions": { "type": "integer" },
        "num_positions": { "type": "integer" },
        "num_streets": { "type": "integer" },
        "action_embedding_dim": { "type": "integer" },
        "position_embedding_dim": { "type": "integer" },
        "street_embedding_dim": { "type": "integer" }
      }
    },
    "card_encoder": {
      "type": "object",
      "properties": {
        "d_model": { "type": "integer" },
        "nhead": { "type": "integer" },
        "dim_feedforward": { "type": "integer" },
        "dropout": { "type": "number" },
        "num_ranks": { "type": "integer" },
        "num_suits": { "type": "integer" },
        "num_streets": { "type": "integer" },
        "rank_embedding_dim": { "type": "integer" },
        "suit_embedding_dim": { "type": "integer" },
        "street_embedding_dim": { "type": "integer" }
      }
    },
    "cross_attention": {
      "type": "object",
      "properties": {
        "num_heads": { "type": "integer" },
        "d_model": { "type": "integer" }
      }
    },
    "output_mlp": {
      "type": "object",
      "properties": {
        "d_model": { "type": "integer" }
      }
    }
  }
}

Detailed Descriptions #

At the top level, there are seven accepted keys:

• general: Configures general parameters for training process.
• training_process: Configures dataloaders, optimizers in the training process.
• action_encoder: Configures the action encoder layer of the model used in the training process.
• card_encoder: Configures the card encoder layer of the model used in the training process.
• cross_attention: Configures the cross attention layer of the model used in the training process.
• output_mlp: Configures the output MLP layer of the model used in the training process.
• $schema (optional): For users to utilize the schemas/model_config_schema.json file. This key will not affect the training process.

general #

• device: "cpu" (default) or "cuda", specify whether to use GPU for training.
• seed: type int (default 42069), for setting random seed for reproducible results.
• load_checkpoint_name, save_checkpoint_name: type str (default "default"), specifies which subdirectory to save the best and final models, training histories, configs from the training process; e.g. if save_checkpoint_name="my_checkpoint", it would save the models to checkpoints/my_checkpoint. If warm_start is set to true, and load_checkpoint_name="my_checkpoint", it would attempt to load previous training loss histories and models from checkpoints/my_checkpoint/final.pth.
• type_of_checkpoint: "final" (default) or "best", for setting which model to load within the subdirectory specified by load_checkpoint_name.

training_process #

• batch_size: type int (default 32), size of batch for dataloaders.
• num_epochs: type int (default 50), number of epochs for training process.
• learning_rate: type number (default 0.001), learning rate for optimizer.
• weight_decay: type number (default 0.01), weight decay for optimizer.
• optimizer: "AdamW" (default) or "Adam", "SGD", type of optimizer to use.
• momentum: type number (default 0), momentum for optimizer; ignored for certain types of optimizers, check torch docs.
• dataset: "GTO" (default) or "Human", training dataset - either on game-theoretic optimal, or on real human data.
• warm_start: type bool (default false), continue training model based on previous trainings (which will load checkpoint from supplied general.load_checkpoint_name key) or fresh new.
• p_train_test_split: type number (default 0.2), test set size ratio, e.g. with 0.2, testing set has size 20% of the whole training data.
• min_epoch: type int (default 100); see patience.
• patience: type int (default 50), when validation loss does not decrease for a period of time over this parameter, and current epoch number is not less than min_epoch, then early-stop the training process.

action_encoder #

The action encoder is a sequential combination of:

1. an initial encoder combining different encoding utilities for action sequences (direct encoding, one-hot encoding, n-way embeddings),
2. a fully-connected layer,
3. a transformer encoder layer using torch.nn.TransformerEncoderLayer for self-attention.

• d_model: type int (default 32), intermediate dimension for the fully-connected layer.
• nhead: type int (default 4), number of heads for the transformer encoder layer.
• dim_feedforward: type int (default 128), the dimension of the feedforward layer in the transformer encoder layer.
• dropout: type number (default 0.1), dropout rate for the drpoout layer in the transformer encoder layer.
• num_actions: type int (default 5), the number of possible actions to encode in the initial encoding layer.
• num_positions: type int (default 6), the number of possible positions to encode in the initial encoding layer.
• num_streets: type int (default 5), the number of possible streets to encode in the initial encoding layer.
• action_embedding_dim: type int (default 4), the dimension for the action encoding in the initial encoding layer.
• position_embedding_dim: type int (default 4), the dimension for the position encoding in the initial encoding layer.
• street_embedding_dim: type int (default 4) the dimension for the street encoding in the initial encoding layer.

card_encoder #

The card sequence encoder is a sequential combination of:

1. an initial encoder combining different encoding utilities for action sequences (direct encoding, one-hot encoding, n-way embeddings),
2. a fully-connected layer,
3. a transformer encoder layer using torch.nn.TransformerEncoderLayer for self-attention.

• d_model: type int (default 32), intermediate dimension for the fully-connected layer.
• nhead: type int (default 4), number of heads for the transformer encoder layer.
• dim_feedforward: type int (default 128), the dimension of the feedforward layer in the transformer encoder layer.
• dropout: type number (default 0.1), dropout rate for the drpoout layer in the transformer encoder layer.
• num_ranks: type int (default 13), the number of possible ranks to encode in the initial encoding layer.
• num_suits: type int (default 4), the number of possible suits to encode in the initial encoding layer.
• num_streets: type int (default 4), the number of possible streets to encode in the initial encoding layer.
• rank_embedding_dim: type int (default 8), the dimension for the rank encoding in the initial encoding layer.
• suit_embedding_dim: type int (default 4), the dimension for the suit encoding in the initial encoding layer.
• street_embedding_dim: type int (default 4) the dimension for the street encoding in the initial encoding layer.

cross_attention #

The cross attention layer uses torch.nn.MultiheadAttention, with bidirectional attention mechanisms (actions attending to cards, as well as cards attending to actions).

• num_heads: type int (default 4), number of heads to use in multihead attention.
• d_model: type int (default 32) intermediate dimension for multihead attention.

output_mlp #

The final output MLP is a sequential combinations of three layers of fully-connected (FC) layers, with ReLU and dropout (in order) applied to the 2nd and 3rd FC layers.

• d_model: type int (default 3), designates hidden dimensions of these FC layers (the middle one has 2 * d_model, the final one has d_model dim).

$schema (Optional) #

Path to a schema json file for writing the schema.