# For full desciption to setup and use this file, refer to
# https://opennmt.net/OpenNMT-py/examples/GGNN.html

# save_data is where the necessary objects will be written
save_data: ../OpenNMT-py-ggnn-example/run/example

# Filter long examples
src_seq_length: 1000
tgt_seq_length: 30

# Data definition
data:
    cnndm:
        path_src: ../OpenNMT-py-ggnn-example/src-train.txt
        path_tgt: ../OpenNMT-py-ggnn-example/tgt-train.txt
        transforms: [filtertoolong]
        weight: 1
    valid:
        path_src: ../OpenNMT-py-ggnn-example/src-val.txt
        path_tgt: ../OpenNMT-py-ggnn-example/tgt-val.txt

src_vocab: ../OpenNMT-py-ggnn-example/srcvocab.txt
tgt_vocab: ../OpenNMT-py-ggnn-example/tgtvocab.txt

save_model: ../OpenNMT-py-ggnn-example/run/model

# Model options
train_steps: 10000
save_checkpoint_steps: 5000
encoder_type: ggnn
layers: 2
decoder_type: rnn
learning_rate: 0.1
start_decay_steps: 5000
learning_rate_decay: 0.8
global_attention: general
batch_size: 32
# src_ggnn_size is larger than vocab plus features to allow one-hot settings
src_ggnn_size: 100
# src_word_vec_size less than rnn_size allows rnn learning during GGNN steps
src_word_vec_size: 16
# Increase tgt_word_vec_size, rnn_size, and state_dim together
# to provide larger GGNN embeddings and larger decoder RNN
tgt_word_vec_size: 64
rnn_size: 64
state_dim: 64
bridge: true
gpu_ranks: 0
n_edge_types: 9
# Increasing n_steps slows model computation but allows information
# to be aggregated over more node hops
n_steps: 5
n_node: 70