config.yml

# model
initializer_gain: 1.0
shared_embedding: false
shallow_decoder: false

hidden_size: 512
filter_size: 4096
num_hidden_layers: 6
num_encoder_layers: 0
num_decoder_layers: 0
attention_key_channels: 0
attention_value_channels: 0
num_heads: 8

layer_preprocess_sequence: n
layer_postprocess_sequence: da
norm_epsilon: 0.000001
layer_prepostprocess_dropout: 0.1
attention_dropout: 0.1
relu_dropout: 0.1

bottleneck_kind: em # vq, mog
bottleneck_bits: 12
num_compress_steps: 3
beta: 0.25
gamma: 1.0
epsilon: 0.00001
decay: 0.999
num_samples: 10
mask_startup_steps: 50000


# training
max_input_length: 50
train_steps: 1000000
eval_steps: 100
save_summary_steps: 1000
save_checkpoints_steps: 2000
n_checkpoints: 20
batch_size: 2048

learning_rate: 0.2

lr_decay: noam
lr_warmup_steps: 4000

clip_gradients: null

optimizer: adam
adam_beta1: 0.9
adam_beta2: 0.997
adam_epsilon: 0.000000001

source_vocab_file: /path/to/wmt14_ende_distill/wmtende.vocab
target_vocab_file: /path/to/wmt14_ende_distill/wmtende.vocab
source_train_file: /path/to/wmt14_ende_distill/train.en
target_train_file: /path/to/wmt14_ende_distill/train.de
source_eval_file:  /path/to/wmt14_ende_distill/valid.en
target_eval_file:  /path/to/wmt14_ende_distill/valid.de
record_train_file: /path/to/wmt14_ende_distill/train.tfrecords
record_eval_file:  /path/to/wmt14_ende_distill/valid.tfrecords

# predict
predict_batch_size: 1
max_decode_length: 100