forked from anmoisio/morphogen-dbca
-
Notifications
You must be signed in to change notification settings - Fork 0
/
run-genbench2023.sh
227 lines (195 loc) · 7.89 KB
/
run-genbench2023.sh
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
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
#!/bin/bash
# Experiment name
exp="deprel-europarl-4way-len30"
# The confguration file is placed in exp/$exp/config.sh
############################################
# Europarl sentence alignments for English, German, French, Greek, and Finnish
python utils/europarl_sent_alignments.py
############################################
# Filter data with OPUSFilter
# use only sentences with max length of 30 words
sbatch 02-opusfilter.sh exp/$exp/opusfilter.yaml --single 1
for lang in "de" "fr" "el" "fi" "en"
do
python utils/opusfilter_decisions_select.py \
text_common_sents_${lang}.txt
exp/deprel-europarl-4way-len30/data/text_common_sents_en.filtered_len30.decisions \
text_common_sents_${lang}_len30.txt
done
# remove duplicates
sbatch 02-opusfilter.sh exp/$exp/opusfilter.yaml --single 2
############################################
# Prepare the data matrices for the data split algorithm
# compound weight threshold
comw="0.5"
# ranges of the lemmas in the frequency-ranked list of lemma types
ranges="200-1000000-100-100.10"
feats_type="deprel"
data_file_name="filtered_en_len30_dedupl.300000_"
parsed_data="exp/$exp/data/$data_file_name"
previous_job_id=$(sbatch 05-prepare-divide-data.sh \
--exp-name $exp --stage 1 --stop-after-stage 1 \
--feats-type $feats_type \
--parsed-data $parsed_data \
| awk '{print $NF}')
previous_job_id=$(sbatch --time=00:30:00 --dependency=afterok:$previous_job_id \
05-prepare-divide-data.sh \
--exp-name $exp --stage 2 --stop-after-stage 3 \
--feats-type $feats_type \
--com-weight-threshold $comw \
--parsed-data $parsed_data \
--ranges $ranges | awk '{print $NF}')
# split the data into parts to create the matrice parts in parallel (to make it faster)
num_parts=10
previous_job_id=$(sbatch --time=00:10:00 --mem=800M --dependency=afterok:$previous_job_id \
05-prepare-divide-data.sh \
--parsed-data $parsed_data \
--exp-name $exp --stage 4 --stop-after-stage 4 \
--feats-type $feats_type \
--com-weight-threshold $comw \
--ranges $ranges --num-parts $num_parts \
| awk '{print $NF}')
# make matrices
previous_job_id=$(sbatch --dependency=afterok:$previous_job_id \
--array=1-$num_parts --time=00:40:00 --mem=600M \
05-prepare-divide-data.sh \
--parsed-data $parsed_data \
--exp-name $exp --stage 5 --stop-after-stage 5 \
--com-weight-threshold $comw \
--ranges $ranges --feats-type $feats_type | awk '{print $NF}')
# combine the matrix parts
previous_job_id=$(sbatch --dependency=afterok:$previous_job_id \
05-prepare-divide-data.sh \
--parsed-data $parsed_data \
--exp-name $exp --stage 6 --stop-after-stage 6 \
--com-weight-threshold $comw --feats-type $feats_type \
--ranges $ranges --num-parts $num_parts | awk '{print $NF}')
############################################
# Split data
atomdiv=0.0
for seed in 11 22 33
do
for comdiv in 0.0 1.0
do
echo $exp exp/$exp/splits/$prep_data_dir $comdiv $seed
sbatch 06-divide.sh $exp exp/$exp/splits/$prep_data_dir $atomdiv $comdiv $seed
done
done
############################################
# Prepare data for NMT training
# The iteration of the data split algorithm that we want to use
iter="200000"
src_lang="en"
filtered_src_data="exp/$exp/data/text_common_sents_${src_lang}_len30_fidedupl.txt"
for seed in 11 22 33
do
for comdiv in 0.0 1.0
do
for tgt_lang in "de" "fr" "el" "fi"
do
filtered_tgt_data="exp/$exp/data/text_common_sents_${tgt_lang}_len30_fidedupl.txt"
data_dir=( exp/$exp/splits/$prep_data_dir/comdiv${comdiv}_atomdiv0.0_seed${seed}_* )
data_dir=${data_dir[0]}
echo $data_dir
sbatch 07-prep-nmt-data.sh \
$exp \
exp/$exp/splits/$prep_data_dir \
$filtered_src_data \
$filtered_tgt_data \
$src_lang $tgt_lang \
$data_dir/{train,test}_set_iter${iter}_* \
$data_dir/iter${iter} \
false true
done
done
done
############################################
# Train tokenizer, build vocab, train NMT model
vocab_size_tgt="10000"
vocab_size_src="10000"
tok_method="bpe"
iter="220000_noaug_lowercase"
for seed in 11 22 33
do
for comdiv in 0.0 1.0
do
for tgt_lang in "de" "fr" "el" "fi"
do
datadir=exp/$exp/splits/$prep_data_dir/comdiv${comdiv}_seed${seed}_*/iter$iter
previous_job_id=$(sbatch \
08-train-spm.sh $exp $src_lang $tgt_lang $vocab_size_src $vocab_size_tgt \
$datadir $tok_method | awk '{print $NF}')
previous_job_id=$(sbatch --dependency=afterok:$previous_job_id \
09-build-vocab.sh $exp $src_lang $tgt_lang \
$vocab_size_src $vocab_size_tgt \
$datadir $tok_method | awk '{print $NF}')
sbatch --dependency=afterok:$previous_job_id \
--time=04:00:00 \
10-train-nmt-model.sh $exp $src_lang $tgt_lang \
$vocab_size_src $vocab_size_tgt $datadir $tok_method
done
done
done
############################################
# Translate
vocab_size_tgt="10000"
vocab_size_src="10000"
src_lang="en"
tokenizer="bpe"
for model_cp in 1000 2000 3000 4000 5000 6000
do
for seed in 11 22 33
do
for comdiv in 0.0 1.0
do
for tgt_lang in "de" "fr" "el" "fi"
do
datadir=( exp/$exp/splits/$prep_data_dir/comdiv${comdiv}_seed${seed}_*/iter$iter )
datadir=${datadir[0]}
11-translate.sh "$exp" "$datadir" $src_lang $tgt_lang \
"$vocab_size_src" "$vocab_size_tgt" "$tokenizer" "false" "$model_cp"
done
done
done
done
############################################
# Evaluate translation with SACREBLEU
for model_cp in 1000 2000 3000 4000 5000 6000
do
for seed in 11 22 33
do
for comdiv in 0.0 1.0
do
for tgt_lang in "de" "fr" "el" "fi"
do
datadir=( exp/$exp/splits/$prep_data_dir/comdiv${comdiv}_seed${seed}_*/iter$iter )
datadir=${datadir[0]}
nmt_dir="$datadir/nmt-fi-en_unigram_vocabs_${vocab_size_src}_${vocab_size_tgt}"
echo "Vocabs ${vocab_size_src}-${vocab_size_tgt}, comdiv ${comdiv}, seed ${seed}"
./utils/get-best-onmt-model-cp.sh "$nmt_dir" || continue
model_cp=$(cat $nmt_dir/best_model_cp.txt)
if [ -f "${nmt_dir}/test_pred_cp${model_cp}_full.bleu.chrf2.confidence" ] && \
[ -s "${nmt_dir}/test_pred_cp${model_cp}_full.bleu.chrf2.confidence" ]; then
echo "Skipping ${nmt_dir} because test_pred_cp${model_cp}_full.bleu.chrf2.confidence exists"
continue
else
echo "----------> Calculating BLEU, chrf for ${nmt_dir} with model cp ${model_cp}"
fi
sbatch 12-translation-eval.sh \
$exp $datadir \
$src_lang $tgt_lang \
$vocab_size_src $vocab_size_tgt \
$model_cp \
$tokenizer false
done
done
done
done
############################################
# Figures
# Figure 1 in the Genbench2023 paper
python figures/sns_plot.py \
--result_files \
exp/deprel-europarl-4way-len30/splits/filtered_en_len30_dedupl.300000_ranges200-1000000-100-100.10_comweight5/comdiv{0.0,1.0}_atomdiv0.0_seed{11,22,33}_subsample500every1iters_groupsize1_testsize0.1to0.2_leaveout0.0_moveasampleevery1iters_presplitno/iter2*0000_noaug_lowercase/nmt-en-{de,fi,fr,el}_bpe_vocabs_10000_10000/test_pred_cp*_full.bleu.chrf2 \
exp/deprel-europarl-4way-len30/splits/random_210000_30000_{1,2,3}/nmt-en-{de,fi,fr,el}_bpe_vocabs_10000_10000/test_pred_cp*_full.bleu.chrf2 \
--output figures/images/genbench_scatter.png --type 'genbench'