hred_small.yml

type: hred
learning_rate: 0.0002
max_gradient_norm: 5.0
vocab_size: 40000
num_layers: 2
residual: True
embedding_type: 'glove840B'
cell_type: 'gru'
encoder_type: 'uni'
context_type: 'uni'
hidden_units: 512
optimizer: 'adam'
batch_size: 64

num_buckets: 5
src_max_len: 30
tgt_max_len: 30

num_turns: 3

num_train_epochs: 15
steps_per_stats: 10
steps_per_eval: 3000
# Early Stopping parameters, inspired from VHRED impl
patience: 50
degrade_threshold: 1.003

encoder_dropout_rate: 0.2
context_dropout_rate: 0.2
decoder_dropout_rate: 0.2

decoding_length_factor: 2.0

learning_rate_decay_scheme: luong234
#start_decay_step: 0
#decay_steps: 10000
#decay_factor: 0.5

scheduled_sampling_prob: 0.0
scheduled_sampling_decay_scheme: luong234

beam_width: 5
length_penalty_weight: 1.0
sampling_temperature: 0.0
infer_batch_size: 128

log_device: False