# Model options
type: transformer
dim-emb: 512
enc-depth: 6
dec-depth: 6
tied-embeddings-all: false
transformer-heads: 8
transformer-dim-ffn: 2048
transformer-ffn-activation: relu
transformer-preprocess: ""
transformer-postprocess: dan
transformer-dropout: 0.1

# Training options
cost-type: ce-mean-words
max-length: 100
mini-batch: 1000
mini-batch-fit: true
maxi-batch: 1000
optimizer-params:
  - 0.9
  - 0.98
  - 1e-09
sync-sgd: true
learn-rate: 0.0003
lr-decay-inv-sqrt:
  - 16000
lr-warmup: 16000
label-smoothing: 0.1
clip-norm: 0
exponential-smoothing: 0.0001
disp-freq: 50
disp-first: 10
save-freq: 10ku
early-stopping: 10

# Validation set options
keep-best: true
beam-size: 8
normalize: 1
valid-freq: 20ku
valid-metrics:
  - ce-mean-words
  - bleu
  - perplexity
valid-mini-batch: 64
train-sets:
  - /mnt/d/model2train/data/src-train.txt
  - /mnt/d/model2train/data/tgt-train.txt
data-weighting-type: sentence
data-weighting: /mnt/d/model2train/data/weight-train.txt
valid-sets:
  - /mnt/d/model2train/data/src-val.txt
  - /mnt/d/model2train/data/tgt-val.txt
model: /home/ogaith/model2train/model.npz
vocabs:
  - /mnt/d/model2train/data/src-vocab.yml
  - /mnt/d/model2train/data/tgt-vocab.yml
log: /home/ubuntu/model2train/train.log
valid-log: /home/ubuntu/model2train/valid.log
after: 10e
workspace: 8000