config_chatbot_livejupiter_transformer.yaml

## Where the samples will be written
save_data: data/run/example
## Where the vocab(s) will be written
src_vocab: data/run/vocab.src
tgt_vocab: data/run/vocab.tgt
# Prevent overwriting existing files in the folder
overwrite: False
share_vocab: True


# Corpus opts:
data:
    corpus_1:
        path_src: data/train.src
        path_tgt: data/train.tgt
    valid:
        path_src: data/valid.src
        path_tgt: data/valid.tgt

# Vocabulary files that were just created
src_vocab: data/run/vocab.src
tgt_vocab: data/run/vocab.tgt

# Train on a single GPU
world_size: 1
gpu_ranks: [0]

# Where to save the checkpoints
save_model: ./trained_model/model_chatbot_livejupiter
save_checkpoint_steps: 500
train_steps: 200000
valid_steps: 500
batch_size: 256
early_stopping: 4 
early_stopping_criteria: ppl
learning_rate: 0.1


# Model
encoder_type: transformer
decoder_type: transformer
enc_layers: 3
dec_layers: 3
heads: 8
rnn_size: 512
word_vec_size: 512
transformer_ff: 2048
dropout_steps: [0]
dropout: [0.1]
attention_dropout: [0.1]
share_decoder_embeddings: true
share_embeddings: true
position_encoding: true