blob: 0ebe7790b9896b3aa36dbb12756051c2b31af628 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
|
#!/bin/bash -x
#####################################################################
# SUMMARY: Test early stopping after stalling on the first validation metric
# AUTHOR: snukky
# TAGS: valid valid-script
#####################################################################
# Exit on error
set -e
# Test code goes here
rm -rf stop_on_1st stop_on_1st.log stop_on_script.temp
mkdir -p stop_on_1st
# Start training with a fully trained model so that validation metrics do not improve easily
cp $MRT_MODELS/wngt19/model.small.npz stop_on_1st/model.npz
test -s devset.en || head -n 50 $MRT_MODELS/wngt19/newstest2014.en | sed -r 's/@@ //g' > devset.en
test -s devset.de || head -n 50 $MRT_MODELS/wngt19/newstest2014.de | sed -r 's/@@ //g' > devset.de
# Training sides are intentionaly reversed to test early stopping
$MRT_MARIAN/marian \
--seed 2222 --no-shuffle --clip-norm 1 --maxi-batch 1 --mini-batch 32 -w 2500 \
-m stop_on_1st/model.npz -t $MRT_DATA/europarl.de-en/corpus.small.{de,en}.gz \
-v $MRT_MODELS/wngt19/en-de.{spm,spm} \
--disp-freq 5 --valid-freq 10 --after-batches 200 \
--valid-metrics valid-script ce-mean-words \
--valid-script-path ./stop_on_script.sh \
--valid-sets devset.{en,de} \
--valid-log stop_on_1st.log \
--early-stopping 4 --early-stopping-on first
test -e stop_on_1st/model.npz
test -e stop_on_1st/model.npz.yml
test -e stop_on_1st.log
$MRT_TOOLS/strip-timestamps.sh < stop_on_1st.log | grep '\[valid\]' > stop_on_1st.out
$MRT_TOOLS/diff-nums.py stop_on_1st.out stop_on_1st.expected -p 0.2 -o stop_on_1st.diff
# Exit with success code
exit 0
|
