run-train-scorer-finetune.py

#!/bin/bash
# source you pyenv or conda or whatever here
cd ~/code/ParlAI/
python examples/train_model.py \
-t convai2:self \
--max-train-time 82079 \
--model parlai_external.agents.seq2seq_on_steroids.seq2seq_on_steroids:SteroidSeq2seqAgent \
--init-model model # PUT INITIAL MODEL HERE, from model.tar.gz
--margin 1 \
--num-rank-cand 5 \
--cand-type current_labels \
--lmweight 0.1 \
-bi True \
-pt True \
--strict-load False \
--rankweight 1 \
--rankhiddensize 512 \
--ranknl 4 \
--ranklossreduce elementwise_mean \
--rank-activation Tanh \
-rc True \
-hs 512 \
-esz 300 \
-att general \
-nl 2 \
-rnn lstm \
-lr 0.001 \
-dr 0.5 \
-clip 0.1 \
-lt enc_dec \
-opt adam \
-emb glove \
-mom 0.9 \
-bs 64 \
-bsrt False \
--tensorboard-log True \
--tensorboard-tag margin,num_rank_cand,lmweight,rankweight,rankhiddensize,ranknl,ranklossreduce,rank_activation \
--tensorboard-comment withadam001 \
--tensorboard-metrics loss,rank_loss,ppl,f1,hits@1,hits@5,hits@10 \
-veps 0.5 \
--validation-metric hits@1 \
--validation-metric-mode max \
--validation-patience 24 \
--save-after-valid True \
--load-from-checkpoint True \
--dict-lower True \
--dict-maxexs -1 \
--dict-include-valid True \
--log-every-n-secs 25 \
--model-file #PATH TO SAVE NEWLY TRAINED FILES