how to run the model

[vivald99]

Hi @beckdaniel ,

I would like to try to run the model. However, I can't do that using the README.md. I think I need the complete steps ( 2 Training models, 3 Decoding, postprocessing and evaluation) to try to run it.

Do you plan to update this information?

[vivald99]

hi @beckdaniel ,

I'm trying to replicate the values for BLEU scores using LDC2017T10 dataset.

I did the preprocessing as you suggested in README. After, I ran the following command to training the model: python -m sockeye.train --source ../data/amr/train/nodes.pp.txt --source-graphs ../data/amr/train/triples.pp.txt --validation-source ../data/amr/dev/nodes.pp.txt --val-source-graphs ../data/amr/dev/triples.pp.txt --edge-vocab ../data/amr/edge_vocab.json --use-grn --skip-rnn --grn-type gated --grn-activation relu --grn-num-layers 8 --grn-num-networks 1 --grn-num-hidden 576 --grn-norm --grn-positional --grn-pos-embed 12 --batch-size 16 --target ../data/amr/train/anon.surface.pp.txt --validation-target ../data/amr/dev/anon.surface.pp.txt --output results/ --max-seq-len 200 --word-min-count 2 --bucket-width 10 --initial-learning-rate 0.0003 --learning-rate-reduce-num-not-improved 3 --learning-rate-reduce-factor 0.5 --max-num-checkpoint-not-improved 8 --num-layers 2:2 --rnn-attention-type bilinear --embed-dropout 0.5

Then, I ran the translation: python -u -m sockeye.translate -m results/ --edge-vocab ../data/amr/edge_vocab.json --max-input-len 200 --beam-size 5 < ../data/amr/test/nodes_triples.pp.txt

I tried to de-anonymize the result using map.pp.txt. I calculate the BLEU score for the result and I got 20.93. In the paper, you reported 23.3.

Could you tell me if my setting is correct? Do you have the script to de-anonymize the results?

Thank you for sharing your work!